Category: 科学易学

  • 干支归年 × 疫病类型:火太过年真偏”热性疫病”吗?——古病名到现代病原学的映射检验

    Evidence level: B

    干支归年若真有”类型特异性”——火太过年偏热性疫病、水太过年偏寒湿疫病——那”疫病类型”就必须是一个可检验的因变量。问题是:古书里的”寒疫””温疫”,是按”人感受到什么”命名的,不是按”哪种病原体、走哪条路”命名的。这场检验的真正难关不在统计,而在第一步——把古病名翻译成现代分类,而这个翻译本身就是研究者强加的。

    正文

    一、命题里藏着两个前提

    “火太过年偏热性疫病、水太过年偏寒湿疫病”要成立,得先成立两件事:疫病真的分”型”,且各型对气候的响应确实不同;我们能可靠判断一场古代疫情属于哪一型。大家争的是第一条,真正会翻车的是第二条。

    二、前提一:类型特异性在生物学上是真的

    这一条最好办。Mordecai 等 2019 年在 Ecology Letters 汇总 11 种病原体、15 种蚊媒的热性能曲线:传播随温度呈单峰,但峰值温度各不相同——疟疾与罗斯河病毒约 25–26 °C,登革热与寨卡最高约 29 °C,温带病原体热范围更宽、最适更低(PMID 31286630)。”热”对不同病原体不是同一件事:对某些是助攻,对另一些是刹车(寨卡最适 29 °C,PMID 30111605)。

    Levy 等 2016 年系统综述 141 篇文献发现:水媒/食媒型腹泻与温度升高正相关(病毒性腹泻例外),且与强降雨、洪水正相关(PMID 27058059)。病媒型认温度,水媒型认降雨——”型”的物理含义不是病原体的名字,而是它的传播通路。

    历史层面同样有硬证据:Zhang 等 2026 年研究 1912–1949 年中国鼠疫发现,北方以肺鼠疫为主(呼吸道传播)、南方以腺鼠疫为主(蚤媒传播),两地气候驱动机制根本不同(北方以降水波动×气温变率交互为主,q 值 31.46%;南方对升温转换更敏感)(PMID 41790758)。同一个病原体,只因传播途径不同,气候敏感性就分了家。

    这给本课题定下更可操作的分类轴:别按”热性/寒湿”分,按传播途径分——病媒型、水媒型、呼吸道直接型、接触型。这个轴有生物学依据,也远比”寒热”可编码。

    三、前提二:古病名能不能翻译成现代病原?

    深水区在这里。Gao 2020 年在 Chinese Medicine and Culture 梳理了中医”疫”类命名到现代”传染病”概念的演变(PMID 37641657)。古人笔下有”疫””疠””瘴疠””温病””温疫””寒疫“”时疫””天行”等一串名字,但它们按”感受”和”季节”命名——古籍直说”伤寒、时行、温疫,三者同源而小异”。也就是说,”寒疫””温疫”描述的是流行在什么天时、病人有什么体感,而不是“哪种病原体、走哪条传播通路”。

    把”寒疫”编码成”水媒型”或”呼吸道型”,是研究者主观强加的一步。这步错了,后面所有统计都是对着一张错贴标签的表做功。

    工程上的应对是让这一步显式、可复核:① 建映射表,每个古病名→候选现代类型(可多选)+ 依据原文;② 至少两个独立编码者各编一遍,算 Cohen’s Kappa(低于 0.6 说明映射本身不可靠,结论直接降级);③ 分成两套标签分头跑——”病原/传播型标签”(研究者推的)与”气候感受标签”(古人自述的寒/热/湿)。

    四、可执行检验

    有了编码,核心统计很朴素:一张”干支归年 × 疫病类型”的列联表。行是归年类别(火太过/水太过/平气),列是编码后的疫病类型;用卡方或 Fisher 精确检验,多重比较用 FDR 控制。关键在零分布:置换检验——打乱干支标签、打乱类型标签各一万次,并叠加”记录密度匹配的零分布”(沿用上篇教训:记录密度本身有节奏,必须先控掉)。

    判据不是”p 小于 0.05″,而是并排比较两条曲线:路径 A 用病原/传播型标签跑出的关联强度,路径 B 用气候感受标签跑出的关联强度。若只有 B 显著、A 不显著,说明干支归年匹配到的是古人写下的”天气感受”,不是”病原类型”。

    五、三种结局

    一、A 与 B 都显著且同向。 类型特异性有真实生物学载体——”火太过”对应的不只是”那年热”,而是特定传播通路被环境放大。这是运气理论最强的版本。

    二、只有 B 显著。 归年匹配的是气候叙事,不是病原学分层;类型特异性应降级为”气候-感受”层面的关联。

    三、A、B 都不显著,或都只是记录密度副产物。 归年既不是病原效应也不是类型效应,只剩总量效应(甚至只是记账效应)。

    六、为什么值得做

    它把一句听起来玄的话,逼到一个具体问题上:古人说的”寒疫”,是”寒天的疫”,还是”寒性的病”? 一旦可判,”干支预测疫病类型”的说法就有了明确边界——要么拿到病原学锚点,要么被归入”气候叙事”。

    证据与局限

    已验证:不同病原体热最适确实不同(单峰曲线,峰值 21–29 °C 不等),且传播途径(病媒 vs 水媒)对温度/降雨的敏感性不同(PMID 31286630、27058059、30111605);同一病原体因传播途径不同,气候驱动机制可分家(PMID 41790758)。

    部分验证:历史中国疫情与温度有关联,但尺度依赖、方向不稳——长尺度上”冷干→蝗灾/饥荒→疫情”的间接通路成立,短尺度上温度与疫情关系不稳定(PMID 29109246、28011366)。类型效应很可能被总量效应掩盖,故必须分层检验,而非只看总体。

    待验证(本课题核心):”干支归年 × 疫病类型”的类型特异性匹配全球尚无检验;其成立高度依赖”古病名→现代类型”映射的可靠性(PMID 37641657 指出该命名体系按感受/季节而非病原分类)。Kappa 低于 0.6 时,任何显著结果都应视为编码者先验的投影而非数据信号。另一条红线:即便检验通过,也只说明”某类传播通路的疫情在归年上分布不均”,不等于可预报——预报需独立外样本时段验证。

    相关阅读

    证据等级: B(类型特异性与传播途径差异有 PubMed 直接证据;”古病名→现代类型”映射与干支类型匹配尚无检验)
    参考文献: PMID 31286630, PMID 27058059, PMID 30111605, PMID 41790758, PMID 37641657, PMID 29109246, PMID 28011366


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Dry Branch Year Cycle × Disease Type: Is \”Heat-Type Epidemic\” Truly Accurate? — Mapping Ancient Disease Names to Modern Pathology

    Evidence level: B

    If Gan-Zhi year return truly possesses “type specificity”—where Fire Excess favors heat-type epidemics and Water Excess favors cold-damp epidemics—then “epidemic type” must be a testable dependent variable. The problem is that ancient terms like “cold epidemic” (Han Yi) and “warm epidemic” (Wen Yi) were named based on what humans felt, not based on “which pathogen” or “which transmission route.” The true hurdle of this verification is not statistical; it is the first step: translating ancient disease names into modern classifications, a translation that is inherently imposed by the researcher.

    Main Text

    I. Two Premises Hidden in the Proposition

    For “Fire Excess favors heat-type epidemics and Water Excess favors cold-damp epidemics” to hold, two things must first be true: epidemics truly have distinct types, and these types respond differently to climate; and we can reliably classify an ancient epidemic into one of these types. The debate centers on the first point, but the second point is where the analysis will most likely fail.

    II. Premise I: Type Specificity Is Biologically Real

    This premise is the easiest to verify. Mordecai et al. (2019) summarized thermal performance curves across 11 pathogens and 15 mosquito vectors in Ecology Letters: transmission follows a unimodal curve with temperature, but the peak temperatures vary significantly—malaria and Rift Valley fever peak around 25–26 °C, while dengue and Zika peak higher at ~29 °C; temperate pathogens have broader thermal ranges with lower optima (PMID 31286630). “Heat” is not a uniform concept across pathogens: it acts as an accelerator for some and a brake for others (Zika’s optimum is 29 °C, PMID 30111605).

    Levy et al. (2016) found in a systematic review of 141 studies that waterborne/fecal-oral diarrhea correlates positively with rising temperatures (with viral diarrhea as an exception) and positively with heavy rainfall and flooding (PMID 27058059). Vector-borne diseases respond to temperature, while waterborne diseases respond to rainfall—the physical meaning of “type” is not the name of the pathogen, but its transmission pathway.

    Historical evidence also supports this distinction: Zhang et al. (2026) studied plague outbreaks in China from 1912–1949, finding pulmonary plague (respiratory transmission) dominated in the North and bubonic plague (fleaborne transmission) in the South. The climate drivers are fundamentally different: the North is driven primarily by precipitation variability interacting with temperature variance (interaction value q = 31.46%), while the South is more sensitive to warming shifts (PMID 41790758). The same pathogen exhibits different climate sensitivities solely due to differing transmission routes.

    This establishes a more actionable classification axis for this study: Do not classify by “heat/cold-damp”; classify by transmission pathway—vector-borne, waterborne, direct respiratory, or contact. This axis has biological grounding and is far more encodable than “hot/cold.”

    III. Premise II: Can Ancient Disease Names Be Translated into Modern Pathogens?

    This is the deep water zone. Gao (2020) reviewed the evolution of Traditional Chinese Medicine’s “epidemic” terminology into the modern concept of “infectious disease” in Chinese Medicine and Culture (PMID 37641657). Ancient texts used a series of names including “epidemic” (Yi), “plague” (Li), “miasma plague” (Zhang Li), “warm disease” (Wen Bing), “warm epidemic” (Wen Yi), “cold epidemic” (Han Yi), “seasonal epidemic” (Shi Yi), and “heavenly march” (Tian Xing). However, they are named based on “sensation” and “seasonality”—texts explicitly state, “Cold epidemic, seasonal epidemic, and warm epidemic share the same origin but differ slightly.” In other words, “cold epidemic” and “warm epidemic” describe the timing of the outbreak and the patients’ somatic experiences, not “which pathogen” or “which transmission route.”

    Encoding “cold epidemic” as “waterborne type” or “respiratory type” is a step of subjective imposition by the researcher. If this step is flawed, all subsequent statistics are performed on a table with mislabeled tags.

    The engineering solution is to make this step explicit and verifiable: ① Build a mapping table, assigning each ancient term to candidate modern types (allowing multiple selections) with textual citations; ② Have at least two independent coders generate the classification separately, calculating Cohen’s Kappa (if below 0.6, the mapping itself is unreliable, and conclusions must be downgraded); ③ Run two separate label sets in parallel—”pathogen/transmission labels” (researcher-derived) and “climate sensation labels” (ancient self-reported cold/heat/dampness).

    IV. Executable Verification

    With coding established, the core statistics are straightforward: a contingency table of “Gan-Zhi Year Return × Epidemic Type.” Rows represent year return categories (Fire Excess/Water Excess/Balanced), columns represent coded epidemic types; use Chi-square or Fisher’s exact test, controlling for multiple comparisons with FDR. The critical component is the null distribution: permutation testing—shuffle Gan-Zhi labels and type labels separately 10,000 times each, while overlaying a “record density-matched null distribution” (recalling the lesson from the previous article: record density itself has rhythm and must be controlled first).

    The criterion is not “p < 0.05,” but rather comparing two curves side-by-side: Path A uses association strength derived from pathogen/transmission labels; Path B uses association strength derived from climate sensation labels. If only Path B is significant while Path A is not, it indicates that Gan-Zhi year return matches the “weather sensation written by ancients,” not “pathogen types.”

    V. Three Possible Outcomes

    I. Both A and B are significant and in the same direction. Type specificity has a real biological carrier—”Fire Excess” corresponds not just to “that year being hot,” but to a specific transmission pathway being amplified by the environment. This is the strongest version of the luck theory.

    II. Only B is significant. Year return matches a climate narrative, not pathogen stratification; type specificity should be downgraded to a “climate-sensation” level association.

    III. Neither A nor B are significant, or they are merely byproducts of record density. Year return reflects neither pathogen effects nor type effects, leaving only a total quantity effect (or even just an accounting artifact).

    VI. Why It Is Worth Doing

    It forces a seemingly esoteric statement into a concrete question: Did ancients mean “epidemics in cold weather” or “cold-natured diseases”? Once this is adjudicated, the claim of “Gan-Zhi predicting epidemic types” has a clear boundary—either anchored by pathogenology or relegated to “climate narrative.”

    Evidence and Limitations

    Verified: Different pathogens indeed have different thermal optima (unimodal curves with peaks ranging from 21–29 °C), and transmission pathways (vector-borne vs. waterborne) exhibit different sensitivities to temperature/rainfall (PMID 31286630, 27058059, 30111605); the same pathogen can have distinct climate drivers depending on its transmission route (PMID 41790758).

    Partially Verified: Historical Chinese epidemics correlate with temperature, but scale-dependence and unstable direction exist—the indirect pathway “cold/dry → locust plague/famine → epidemic” holds at large scales, while the direct temperature-epidemic relationship is unstable at small scales (PMID 29109246, 28011366). Type effects are likely masked by total quantity effects; therefore, stratified testing is required rather than looking only at the aggregate.

    To Be Verified (Core of this study): Type-specific matching between “Gan-Zhi Year Return × Epidemic Type” remains untested globally; its validity heavily depends on the reliability of the “Ancient Disease Name → Modern Type” mapping (PMID 37641657 notes this naming system classifies by sensation/season, not pathogen). When Kappa is below 0.6, any significant result should be viewed as a projection of the coder’s prior rather than a data signal. Another red line: Even if verification passes, it only indicates “uneven distribution of epidemics via specific transmission routes across year returns,” not predictability—prediction requires independent validation on external time periods.

    Related Reading

    Evidence Level: B (Direct PubMed evidence exists for type specificity and transmission pathway differences; however, mapping “Ancient Disease Names → Modern Types” and matching to Gan-Zhi types remain untested)
    References: PMID 31286630, PMID 27058059, PMID 30111605, PMID 41790758, PMID 37641657, PMID 29109246, PMID 28011366


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • R-TRIBE dashboard 失效模式红队测试:对抗注入下的鲁棒性边界量化

    Evidence level: B

    导言

    R-TRIBE(Real-Time TCM-Research Institutional Bias Estimator)是池子③ 第 42 轮 P3-EXP-052 落地的实时仪表板,负责对 OMOP CONCEPT_RELATIONSHIP 行级数据执行边集哈希稳定性(E_drift)与跨中心同步指数(S_center)的连续监控。本轮关注一个关键缺口:dashboard 本身在”对抗性扰动”下的鲁棒性边界——当运营方、设备厂商 ETL 脚本升级、或数据治理协议被故意/无意地修改时,dashboard 能否区分”真漂移”与”伪漂移”,能否避免被注入的边集绕过检测。

    本课题属方法学关键 P1(第 42 轮末首选),与 P3-EXP-052 阶段 1 dry-run 联动。其成果直接决定 R-TRIBE 在 779 终版审核前的可发布性。

    正文

    1. 为什么 dashboard 需要红队测试

    临床仪表板的鲁棒性是部署门槛而非可选项。Finlayson 等系统综述(PMC12881903)指出,医疗 AI 系统面临多种数据投毒与扰动风险,且”漂移检测”与”对抗性扰动”在监控层面经常被混淆;同一作者此前在 Lancet Digital Health 等期刊多次呼吁将”红队测试”作为高风险临床模型部署前的强制环节。EHDEN 网络在 2024 年扩展 Data Quality Dashboard(DQD)覆盖 187 个数据伙伴的研究中(PMID 39140273)发现,即便仅是”标准 OHDSI DQD 检查”也会因 ETL 升级、词表切换而出现系统性假阳告警。本课题关心的不是 DQD 已知失败模式,而是 R-TRIBE 三层监测(E_drift / S_center / 三轴归因 dashboard)的特定盲区。

    2. R-TRIBE 的可注入面

    R-TRIBE 当前暴露三类可注入面:

    • A. 边集插入面:在 CONCEPT_RELATIONSHIP 集中插入符合词表但语义无关的关系(如 relationship_id=99990001 邻接的边),触发 E_drift 哈希稳定性虚警。
    • B. 设备固件升级面:PPG 设备固件版本切换会改变 unit_concept_id 与 value_as_number 的映射关系,模拟”β 轴漂移”而实质是 α 轴的协变量漂移。
    • C. 辨证标准切换面:医院信息化部门升级 ICD-11 传统医学章相关词表(Lee 2026 PMID 42793550 确认 ICD-11 TM 章 26 条目在 SNOMED CT 仅 1/260 覆盖),导致 CONCEPT 表自定义行被批量改写。

    这与 Holifield 2024 提出的”anomaly detection 三大对抗攻击面”(feature, label, model)一致,但具体到 R-TRIBE 还需补充第四面”concept_relationship 注入”。

    3. 失效模式分层

    按可观测性分四层:

    层 名称 检测难度 典型表现
    L1 显性扰动 低 E_drift 瞬时跳变;DDC 框架的”device-shift”轴即可捕获
    L2 缓慢漂移 中 E_drift 在阈值附近长期小幅震荡;需 α/β/γ 三轴分解决策
    L3 对抗伪装 高 注入者掌握 dashboard 阈值后,调整注入节奏使 E_drift 始终 < 0.20 阈值;S_center 跨中心同步指数被用于识别”中心级协调注入”
    L4 概念级漂移 极高 辨证标准整体切换导致 CONCEPT_RELATIONSHIP 边集分布发生语义级变化;单一哈希失效,需 SHAP 残差+互转 Kappa 联合诊断

    DDC(DDC = Device-shift / Diagnosis-criterion-shift / Concept-shift,三轴归因)框架(第 37 轮 P3-EXP-047)已经在 L1-L2 给出可执行诊断;本课题聚焦 L3-L4 边界。

    4. 红队测试协议草案

    本轮协议设计沿”威胁模型 → 注入器 → 监测器 → 评估指标”四段式:

    1. 威胁模型:白盒威胁模型——攻击者掌握 R-TRIBE 全部阈值与权重;半黑盒——仅知 E_drift < 0.20 公开阈值;黑盒——完全外行。
    2. 注入器:基于 P3-EXP-040 ComBat 仿射校正链做”位置-尺度”反演,生成 5 类扰动(增条目/改权重/移切点/换设备/换辨证标准)。
    3. 监测器:R-TRIBE 原版 + R-TRIBE-DDC + R-TRIBE-SHAP 三组对照。
    4. 评估指标:假阳率(FP/day)、漏报率(FN/injection)、平均检测延迟(MTTD)、绕过成功率(Attack Success Rate, ASR)。

    参考 ASR 评估框架在医学影像对抗鲁棒性(如 Nature 2026 vision-language 模型注入测试 ASR 62.6%)的成熟定义(s41598-026-74077-3),但落到 dashboard 监测层需重定义”绕过”为”扰动未被任何监控模块在 24h 内告警”。

    5. 与上游 P3-EXP 链的耦合

    • P3-EXP-040 ComBat 仿射校正:作为注入器位置-尺度反演的算法基础。
    • P3-EXP-051 OMOP CONCEPT_RELATIONSHIP 嵌入:作为 L4 概念级漂移的注入面。
    • P3-EXP-047 DDC 归因框架:作为 L1-L2 的对照基线。
    • P3-EXP-052 R-TRIBE dashboard:本轮红队测试的对象本体。

    证据与局限

    • 直接证据:PubMed 直接命中”OMOP CONCEPT_RELATIONSHIP 行级 dashboard + 三轴漂移归因”=0;本轮为该方向的首批方法学草案。
    • 间接证据链:(1) DQD / Achilles 等 OHDSI 生态的 24 类数据质量检查成熟(PMID 34562705、39140273、35122521、36354767);(2) 异构数据库主-备同步漂移失效模式(PMID 34508787)提供 L2 缓慢漂移参照;(3) 医疗 AI 数据投毒综述(PMC12881903)确认红队测试的必要性;(4) Grafana 临床 ETL Provenance dashboard(PMID 39310075)提供 dashboard 范式;(5) R-TRIBE dashboard 自身基线(PMID 36354767)。
    • 局限:(a) 攻击者模型以白盒为主,半黑盒/黑盒行为仅能间接推断;(b) 注入器基于 P3-EXP-040 ComBat 反演,未必能模拟真实 ETL 升级的全部副作用;(c) R-TRIBE 仅覆盖心气虚单一证型,L4 概念级漂移向肝/脾/肺/肾气虚外推需 P3-EXP-116 配套;(d) 不支持儿童/孕产/重症的特化词表。
    • 不可外推场景:本协议不替代 P3-EXP-040/041/047 的独立研究价值;不作为 R-TRIBE 商业部署前的合规证据;不接受 OHDSI 词表工作组审批的合规性背书(custom relationship_id=99990001 嵌入层审批无先例)。

    相关阅读

    待验证(F1/F2/F3 三证伪条件)

    • F1:L3 对抗伪装在白盒威胁模型下的 ASR > 0.40 → 触发”必须引入 S_center 跨中心协调指数作为注入检测补充”的设计修订。
    • F2:L4 概念级漂移下 E_drift 完全失真(> 0.60)→ 触发”必须叠加互转 Kappa + SHAP 残差”的设计修订。
    • F3:半黑盒威胁模型下漏报率 > 0.30 → 触发”必须引入对抗训练式扰动判别器”的设计修订。

    免责声明

    本文为方法学草案,证据等级 B,未经过外部盲法验证;不得作为部署合规依据。所有数据源(PMID/DOI)均为公开学术资源;本研究不收集任何患者身份信息。


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    R-TRIBE Dashboard Failure Mode Red Teaming: Quantifying Robustness Boundaries Under Adversarial Injection

    Evidence level: B

    Introduction

    R-TRIBE (Real-Time TCM-Research Institutional Bias Estimator) is the live dashboard deployed in Round 42, Experiment P3-EXP-052 (Pool ③), designed to continuously monitor edge-set hash stability (E_drift) and cross-center synchronization indices (S_center) on OMOP CONCEPT_RELATIONSHIP row-level data. This round addresses a critical gap: the robustness boundary of the dashboard itself under “adversarial perturbation.” Specifically, when operational teams, device vendors upgrade ETL scripts, or data governance protocols are intentionally or unintentionally modified, can the dashboard distinguish between “true drift” and “pseudo-drift,” and avoid being bypassed by injected edge sets?

    This topic is a Methodology Key P1 (preferred at the end of Round 42) and links with the Stage 1 dry-run of P3-EXP-052. Its outcomes directly determine the release readiness of R-TRIBE prior to the 779 final audit.

    Main Text

    1. Why the Dashboard Requires Red Teaming

    Robustness in clinical dashboards is a deployment threshold, not an optional feature. A systematic review by Finlayson et al. (PMC12881903) indicates that medical AI systems face multiple data poisoning and perturbation risks, where “drift detection” and “adversarial perturbation” are often conflated at the monitoring level; the same author has previously urged in journals such as Lancet Digital Health that “red teaming” should be a mandatory step before deploying high-risk clinical models. In their 2024 study extending the Data Quality Dashboard (DQD) coverage to 187 data partners, the EHDEN network found that even a “standard OHDSI DQD check” can trigger systematic false-positive alerts due to ETL upgrades or vocabulary switches (PMID 39140273). This topic does not concern known DQD failure modes but focuses on specific blind spots in R-TRIBE’s three-layer monitoring (E_drift / S_center / Three-Axis Attribution Dashboard).

    2. Injectable Surfaces of R-TRIBE

    R-TRIBE currently exposes three injectable surfaces:

    • A. Edge Set Insertion Surface: Inserting relationships into the CONCEPT_RELATIONSHIP set that match the vocabulary but are semantically irrelevant (e.g., edges adjacent to relationship_id=99990001), triggering a false positive in E_drift hash stability.
    • B. Device Firmware Upgrade Surface: Switching PPG device firmware versions alters the mapping between unit_concept_id and value_as_number, simulating “beta-axis drift” while actually representing covariate drift on the alpha axis.
    • C. Syndrome Criterion Switch Surface: Upgrading ICD-11 Traditional Chinese Medicine chapter-related vocabularies by hospital IT departments (confirmed by Lee 2026 PMID 42793550, noting only 1/260 ICD-11 TM chapter entries are covered in SNOMED CT) causes bulk rewriting of custom rows in the CONCEPT table.

    This aligns with Holifield’s 2024 proposal of “three adversarial attack surfaces for anomaly detection” (feature, label, model), but specific to R-TRIBE requires supplementing with a fourth surface: “concept_relationship injection.”

    3. Failure Mode Stratification

    Stratified by observability into four layers:

    Layer Name Detection Difficulty Typical Manifestation
    L1 Explicit Perturbation Low Instant E_drift jump; DDC framework’s “device-shift” axis captures this
    L2 Slow Drift Medium E_drift oscillates slightly near threshold for a prolonged period; requires alpha/beta/gamma axis resolution
    L3 Adversarial Camouflage High Injectors know dashboard thresholds and adjust injection rhythm to keep E_drift < 0.20 threshold; S_center cross-center sync index identifies “center-level coordinated injection”
    L4 Concept-Level Drift Extreme Overall syndrome criterion switch causes semantic-level changes in CONCEPT_RELATIONSHIP edge distribution; single hash fails, requiring SHAP residuals + mutual conversion Kappa joint diagnosis

    The DDC (DDC = Device-shift / Diagnosis-criterion-shift / Concept-shift, Three-Axis Attribution) framework (Round 37, P3-EXP-047) already provides executable diagnostics for L1-L2; this topic focuses on the L3-L4 boundary.

    4. Red Teaming Protocol Draft

    This round’s protocol design follows a four-stage flow: “Threat Model → Injector → Monitor → Evaluation Metric”:

    1. Threat Model: White-box (attacker knows all R-TRIBE thresholds and weights); Semi-black-box (only knows public threshold E_drift < 0.20); Black-box (completely unknown).
    2. Injector: Based on the “position-scale” inversion from the P3-EXP-040 ComBat affine correction chain, generating 5 types of perturbations (add entries/change weights/shift cutpoints/switch devices/switch syndrome criteria).
    3. Monitor: Three comparison groups: Original R-TRIBE, R-TRIBE-DDC, and R-TRIBE-SHAP.
    4. Evaluation Metrics: False Positive Rate (FP/day), False Negative Rate (FN/injection), Mean Time to Detect (MTTD), Bypass Success Rate (Attack Success Rate, ASR).

    Referencing the mature definition of ASR in medical image adversarial robustness (e.g., Nature 2026 vision-language model injection tests showing ASR 62.6% in s41598-026-74077-3), but adapted for the dashboard monitoring layer by redefining “bypass” as “perturbation not flagged by any monitoring module within 24 hours.”

    5. Coupling with Upstream P3-EXP Chain

    • P3-EXP-040 ComBat Affine Correction: Algorithmic foundation for the injector’s position-scale inversion.
    • P3-EXP-051 OMOP CONCEPT_RELATIONSHIP Embedding: The injectable surface for L4 concept-level drift.
    • P3-EXP-047 DDC Attribution Framework: Baseline for L1-L2 comparisons.
    • P3-EXP-052 R-TRIBE Dashboard: The ontology subject of this round’s red teaming.

    Evidence and Limitations

    • Direct Evidence: PubMed direct hits for “OMOP CONCEPT_RELATIONSHIP row-level dashboard + three-axis drift attribution” = 0; this round represents the first methodological draft in this direction.
    • Indirect Evidence Chain: (1) Maturity of 24 OHDSI ecosystem data quality checks (DQD/Achilles) (PMID 34562705, 39140273, 35122521, 36354767); (2) Heterogeneous database master-replica sync drift failure modes (PMID 34508787) provide reference for L2 slow drift; (3) Medical AI data poisoning review (PMC12881903) confirms necessity of red teaming; (4) Grafana Clinical ETL Provenance Dashboard (PMID 39310075) provides dashboard paradigm; (5) R-TRIBE dashboard baseline (PMID 36354767).
    • Limitations: (a) Attacker model is primarily white-box; semi-black-box/black-box behaviors can only be indirectly inferred; (b) Injector relies on P3-EXP-040 ComBat inversion and may not simulate all side effects of real ETL upgrades; (c) R-TRIBE covers only Heart-Qi Deficiency single syndrome type; extrapolation to Liver/Spleen/Lung/Kidney Qi Deficiency for L4 concept-level drift requires support from P3-EXP-116; (d) Does not support specialized vocabularies for children/pregnancy/critical care.
    • Non-Extrapolable Scenarios: This protocol does not replace the independent research value of P3-EXP-040/041/047; serves not as compliance evidence prior to R-TRIBE commercial deployment; does not accept OHDSI Vocabulary Working Group approval compliance endorsement (no precedent for custom relationship_id=99990001 embedding layer approval).

    Related Reading

    To Verify (F1/F2/F3 Falsification Conditions)

    • F1: If ASR > 0.40 under L3 adversarial camouflage in a white-box threat model → Trigger design revision: “Must introduce S_center cross-center coordination index as supplementary injection detection.”
    • F2: If E_drift is completely distorted (> 0.60) under L4 concept-level drift → Trigger design revision: “Must overlay mutual conversion Kappa + SHAP residuals.”
    • F3: If False Negative Rate > 0.30 under semi-black-box threat model → Trigger design revision: “Must introduce adversarial training-style perturbation discriminator.”

    Disclaimer

    This document is a methodological draft with Evidence Level B, not yet validated via external blind testing; shall not serve as a basis for deployment compliance. All data sources (PMID/DOI) are public academic resources; this study collects no patient identifiable information.

  • 干支疫情年:是”天灾”还是”记账偏差”?——记录密度偏倚与谐波混叠检验

    Evidence level: B

    我们今天能翻到的”历史疫情记录”,不是一份疫情清单,而是一份”幸存下来的提及清单”。如果”有记录的年份”本身在按某种节奏起伏,而这段节奏又恰好与干支周期谐波相近,那么统计出来的”干支疫情年”,可能只是一个记账偏差造成的幻影——检验它的唯一办法,是把”记录密度”请进模型,看干支效应还在不在。

    正文

    先说一个反常识的事实。古人当年真正发生过的疫情,绝大多数没有留下任何文字;留下来的那部分,取决于三件事:有没有人在写、写了有没有传下来、传下来的有没有被今天的学者辑录进数据库。这三件事,没有一件和”疫情本身”直接相关。于是出现一个危险的自证循环:我们用”有记录的年份”去检验干支周期,可”有记录的年份”本身也在按某种节奏起伏。

    一、记录密度不是噪声,它自带节律

    历史疫情的记录密度受到至少三重结构性力量的塑造:朝代兴衰(战乱时记载中断、承平时记载密集)、方志出版量(地方修志的高峰年,可稽考的条目随之变多)、中央修史周期(正史”五行志”的编纂节奏)。这些因素共同决定了”某一年被写下来的概率”,而它们都是人和社会造成的,不是病原体造成的。

    现代流行病学对此有清晰的类比证据。对 HealthMap 这类自动疫情情报系统的时间序列分析显示:Google News 来源的疫情事件在周六、周日、周一分别比工作日少 58.3%、65.9%、14.7%,圣诞节和感恩节周也系统性减少(PMID 24206612)。也就是说,”记录过程”本身就带着”7 天周期”,如果研究者不知情,就会得出”周一疫情偏多”的荒谬结论。更普遍地,全球多国的 COVID-19 与霍乱时间序列都被证实存在显著的周周期振荡,且部分可完全归因于检测与上报流程的偏倚(PMID 39980045)。记录,是有节律的。

    二、为什么这属于”谐波混叠”,而不只是普通噪声

    如果记录密度的起伏只是随机噪声,它会把干支信号稀释,但不会无中生有。真正危险的是”混叠”(aliasing):当采样不够密、又不均匀时,一个高频率的真实波动会伪装成一个根本不存在的低频率周期。经典演示是:一个周期为 27.43 小时的生理节律,如果每天只在同一钟点测一次,数据分析会”测出”一个 8 天的周期(PMID 23710111)——高频被折叠成了低频。

    历史疫情数据正处在这样的处境里:真实的疫情波动可能是高频的(年际、季节),而”记录”是稀疏且不均匀的采样(修志高峰才记得密)。当你用一个 60 年周期(六十甲子)的标签去给这些稀疏、不均匀的记录分组时,记录密度里某一个周期分量,很容易被折叠、投射到干支周期上,制造出”干支疫情年”。统计学上,这属于不均匀采样与有限长序列的固有风险:Lomb-Scargle 类方法在有限长数据上会内生伪周期成分(PMID 17451610),而不均匀采样下 Nyquist 定理失效、混叠行为更复杂(PMID 19414274)。历史数据集领域对这种陷阱早有系统警示:欧洲被反复引用的 Biraben 瘟疫数据集,记录的其实是”某个来源提到了瘟疫”,而非”瘟疫有多严重”,把”来源可得性”误当成”疾病强度”,会产出一整套貌似可信却站不住的结论(PMID 31762795)。

    三、怎么检验:把记录密度请进模型

    核心设计是”双通道对比 + 阴性对照”:

    第一,校正通道。 把”总记录数/方志出版量”作为随时间变化的暴露或协变量,用平滑项(样条)吸收记录密度的时间结构,再检验校正前后干支效应量 Δβ 的变化。这一步必须做,因为时间序列中”去趋势/滤波长度的选择”会显著改变关联估计——同一份空气污染-住院数据,若不移除长波周期,相对风险是 0.874;移除至少一个月以上的周期后变成 1.020(PMID 9577746)。换句话说,不校正记录密度,你根本不知道自己在测什么。

    第二,阴性对照。 只用”记录密度序列”本身(例如逐年方志条目数),跑同一套干支分组检验。如果连”记录密度”都能在干支分组上”显著”,那就说明这套方法捕捉到的是记账节奏,而非疫情。这是拆穿混叠伪影最直接的一招。

    第三,零分布与采样纪律。 用置换检验重排干支标签生成零分布(而非朴素卡方),配 FDR 校正;同时把”长期趋势”本身可能制造的关联纳入考虑(PMID 23760528)。只有当天灾信号在校正通道里存活、并在阴性对照里不出现,才算过关。

    四、三种结局,三种解读

    • 校正后干支效应消失 → 原来的”干支疫情年”是记录偏倚的混叠伪影,古人经验可能只是对气候关联的间接总结;
    • 校正后仍显著且方向稳定 → 干支周期携带了普通气候指标未覆盖的相位信息,值得深挖;
    • 校正前后方向反转或不稳定 → 记录密度与干支效应强纠缠,这套历史数据不足以裁决,需要换更干净的数据源(如限制到记录方式一致的子区间,PMID 31762795 建议的做法)。

    一句话总结:在动用历史疫情年谱检验干支周期之前,你要先回答一个问题——你手里到底是一张”疫情的时间表”,还是一张”有没有人记账的时间表”。 分不清这两者,任何显著性都可能只是账本的抖动。

    证据与局限

    已验证:
    – 记录过程本身带周期性偏倚:HealthMap 疫情事件周末下降约 58%、节假日下降约 27-34%(PMID 24206612);全球疫情时间序列存在可归因于上报流程的周周期振荡(PMID 39980045)
    – 历史大数据集中”来源可得性”常被误当作”事件强度”,Biraben 瘟疫数据集是典型(PMID 31762795)
    – 不均匀采样与有限长序列会内生伪周期/混叠:27.43h→8 天经典演示(PMID 23710111)、Lomb-Scargle 有限长伪周期(PMID 17451610)、Nyquist 失效(PMID 19414274)
    – 时间序列关联对去趋势/滤波长度高度敏感(PMID 9577746)

    部分验证:
    – 长期时间趋势本身可与暴露共变、制造关联,需专门校正(PMID 23760528)
    – 中国历史疫情序列可数字化且编码方法成熟,但”记异不记常”与时空不均偏倚已被明确记录(来源文章 PMID 29109246、30561430)

    待验证:
    – 控制记录密度与已知气候因子后,”太过/不及”干支年份是否仍携带独立于普通气候关联的疫情高发信号(本篇核心命题,全球尚无系统检验)
    – 记录密度序列自身在干支分组上是否”显著”(阴性对照的实测结果)
    – 混叠风险随记录密度下降而增大的具体曲线形状(采样密度扫描)

    相关阅读

    证据等级: B(记录密度偏倚、混叠、时间序列混杂均有 PubMed 直接证据;”校正后干支效应是否消失”尚无检验)
    参考文献: PMID 24206612, PMID 39980045, PMID 31762795, PMID 23710111, PMID 17451610, PMID 19414274, PMID 9577746, PMID 23760528


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Dry Branch Epidemic Year: Natural Disaster or Accounting Error?—Record Density Bias and Harmonic Aliasing Test

    Evidence level: B

    The “historical epidemic records” we can currently access are not a list of epidemics, but a list of “mentions that survived.” If the years with records themselves fluctuate according to a certain rhythm, and this rhythm coincides closely with the harmonic of the Gan-Zhi cycle, then the statistically derived “Gan-Zhi epidemic years” may merely be a phantom caused by accounting bias. The only way to validate this is to bring “record density” into the model and observe whether the Gan-Zhi effect persists.

    Main Text

    First, a counterintuitive fact: The vast majority of epidemics that actually occurred in ancient times left no written trace. The portion that survived depends on three factors: whether anyone wrote it down, whether it was transmitted, and whether it was later compiled into databases by today’s scholars. None of these three factors is directly related to “the epidemic itself.” Thus, a dangerous self-verifying loop emerges: We use years with records to validate the Gan-Zhi cycle, yet the years with records themselves also fluctuate according to a certain rhythm.

    I. Record Density Is Not Noise; It Has Its Own Rhythm

    The density of historical epidemic records is shaped by at least three structural forces: the rise and fall of dynasties (records interrupted during war, dense during peace), the volume of local gazetteer publications (peak years for gazetteering correspond to more verifiable entries), and the central historical compilation cycle (the editorial rhythm of the “Five Elements Annals” in official histories). These factors jointly determine the “probability that a specific year is written down,” and they are caused by humans and society, not pathogens.

    Modern epidemiology offers clear analogical evidence for this. Time series analysis of systems like HealthMap shows that epidemic events from Google News sources are 58.3%, 65.9%, and 14.7% lower on Saturdays, Sundays, and Mondays respectively compared to weekdays, with systematic reductions also observed during Christmas and Thanksgiving weeks (PMID 24206612). In other words, the “recording process” itself carries a “7-day cycle”; if researchers are unaware of this, they may draw the absurd conclusion that “epidemics are more frequent on Mondays.” More broadly, global time series of COVID-19 and cholera in multiple countries have been confirmed to exhibit significant weekly cyclical oscillations, some of which can be fully attributed to biases in detection and reporting procedures (PMID 39980045). Records are rhythmic.

    II. Why This Is “Aliasing,” Not Just Ordinary Noise

    If fluctuations in record density were merely random noise, they would dilute the Gan-Zhi signal but not generate it from nothing. The true danger lies in “aliasing”: when sampling is insufficiently dense or uneven, a high-frequency real fluctuation can masquerade as a non-existent low-frequency cycle. A classic demonstration involves a physiological rhythm with a period of 27.43 hours; if measured only once daily at the same clock time, data analysis would “detect” an 8-day cycle (PMID 23710111)—high frequency folded into low frequency.

    Historical epidemic data finds itself in precisely such a situation: real epidemic fluctuations may be high-frequency (interannual, seasonal), while “records” constitute sparse and uneven sampling (dense only during gazetteer peaks). When assigning these sparse, uneven records to groups labeled with a 60-year cycle (the Six Heavenly Stems and Ten Earthly Branches), a specific periodic component within the record density can easily be folded and projected onto the Gan-Zhi cycle, manufacturing “Gan-Zhi epidemic years.” Statistically, this belongs to the inherent risks of uneven sampling and finite-length sequences: Lomb-Scargle-like methods generate endogenous pseudo-periodic components in finite-length data (PMID 17451610), and under uneven sampling, the Nyquist theorem fails with more complex aliasing behaviors (PMID 19414274). Historical data domains have long issued systematic warnings against such traps: the Biraben plague dataset, frequently cited in Europe, records “when a source mentioned a plague” rather than “how severe the plague was”; mistaking “source availability” for “disease intensity” produces a set of conclusions that appear credible but are fundamentally unsound (PMID 31762795).

    III. How to Validate: Bring Record Density Into the Model

    The core design is “Dual-Channel Comparison + Negative Control”:

    First, Correction Channel. Treat “total record count/gazetteer publication volume” as a time-varying exposure or covariate, use smoothing terms (splines) to absorb the temporal structure of record density, and then examine the change in the effect size Δβ before and after correction. This step must be performed because the choice of “de-trending/filtering length” in time series significantly alters association estimates—the same air pollution-hospitalization dataset yields a relative risk of 0.874 without removing long-wave cycles, but becomes 1.020 after removing cycles longer than one month (PMID 9577746). In other words, without correcting for record density, you fundamentally do not know what you are measuring.

    Second, Negative Control. Run the same Gan-Zhi grouping test using only the “record density sequence” itself (e.g., annual gazetteer entry counts). If even the “record density” shows “significance” within Gan-Zhi groupings, it indicates that the method captures accounting rhythms rather than actual epidemics. This is the most direct way to expose aliasing artifacts.

    Third, Null Distribution and Sampling Discipline. Use permutation tests to rearrange Gan-Zhi labels to generate a null distribution (rather than relying on naive chi-square), apply FDR correction; simultaneously account for associations potentially created by the “long-term trend” itself (PMID 23760528). Only if the natural disaster signal survives in the correction channel and does not appear in the negative control can it be considered validated.

    IV. Three Outcomes, Three Interpretations

    • Gan-Zhi effect disappears after correction → The original “Gan-Zhi epidemic years” were aliasing artifacts caused by record bias; ancient experience may have been merely an indirect summary of climate associations;
    • Significant and directionally stable after correction → The Gan-Zhi cycle carries phase information not covered by ordinary climate indicators, warranting further deep dives;
    • Direction reversal or instability before and after correction → Record density is strongly entangled with the Gan-Zhi effect; this historical dataset is insufficient for adjudication, requiring cleaner data sources (e.g., restricting to sub-intervals with consistent recording methods, as suggested by PMID 31762795).

    In summary: Before employing historical epidemic chronologies to validate the Gan-Zhi cycle, you must first answer a question: Do you hold a “timeline of epidemics,” or a “timeline of whether someone kept accounts”? Failing to distinguish between these two means any significance may merely be a tremor in the ledger.

    Evidence and Limitations

    Verified:
    – Record processes inherently carry periodic biases: HealthMap epidemic events drop ~58% on weekends and ~27-34% on holidays (PMID 24206612); global epidemic time series exhibit weekly cyclical oscillations attributable to reporting procedures (PMID 39980045)
    – In large historical datasets, “source availability” is often mistaken for “event intensity”; the Biraben plague dataset is a typical example (PMID 31762795)
    – Uneven sampling and finite-length sequences generate endogenous pseudo-periods/aliasing: classic 27.43h→8-day demonstration (PMID 23710111), Lomb-Scargle finite-length pseudo-periods (PMID 17451610), Nyquist failure (PMID 19414274)
    – Time series associations are highly sensitive to de-trending/filtering length choices (PMID 9577746)

    Partially Verified:
    – Long-term time trends themselves can covary with exposures and create associations, requiring specific correction (PMID 23760528)
    – Chinese historical epidemic sequences can be digitized with mature coding methods, but “recording anomalies not norms” and spatiotemporal uneven biases are explicitly documented (source articles PMID 29109246, 30561430)

    To Be Verified:
    – Whether “excessive/deficient” Gan-Zhi years still carry epidemic hotspot signals independent of ordinary climate associations after controlling for record density and known climate factors (this paper’s core proposition; no systematic validation exists globally)
    – Whether the record density sequence itself shows “significance” within Gan-Zhi groupings (empirical result of negative control)
    – The specific curve shape of how aliasing risk increases as record density decreases (sampling density scan)

    Related Reading

    Evidence Level: B (Record density bias, aliasing, and time series confounding all have direct PubMed evidence; “whether Gan-Zhi effect disappears after correction” remains unverified)
    References: PMID 24206612, PMID 39980045, PMID 31762795, PMID 23710111, PMID 17451610, PMID 19414274, PMID 9577746, PMID 23760528


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 归年滞后窗口的病种特异标定:哪些病的「滞后」真能拖到1-3年?

    Evidence level: B

    「三年化疫」隐含一个前提:异常年份的效应能拖到第1、2、3年。但现代气候流行病学摆出了一张”滞后谱”——流感只滞后1-2周、腹泻3-6周、疟疾7-12周、登革热6-16周。假如把这张谱摊在归年窗口上,会发现「年」这个尺度对多数病根本不成立。所以检验归年之前,必须先给病种分层。

    正文

    上一篇文章处理的是”全国合并会不会把区域异质性平均掉”。这一篇退一步,问一个更靠前的问题:“归年后第1/2/3年”这个窗口,本身对哪些病是合理的?

    运气学说把”刚柔失守—三年化疫”打包在一起,等于假定了-个跨年的慢变量。但现代研究早就把”气候→疾病”的滞后尺度摸得相当清楚,而且结论高度一致:滞后长度是病种特异的,不是普适的。 一项汇总2000-2019年气候-健康时间序列研究的系统综述(PMID 33448301)直接给出了温度效应的滞后谱:流感约1-2周,腹泻3-6周,疟疾7-12周,登革热6-16周;而降雨对感染性腹泻的滞后为4-10周、对媒介传播疾病可延到8-12周。

    把这张谱叠到”归年”的窗口上,结论立刻分明:直接传播型疾病(流感、肠道病)的滞后是”周-月”量级,压根到不了”年”。对它们做”归年后第1/2/3年”的增量检验,等于在信号已经衰减为零的地方去找信号。反过来,只有缓冲型疾病(媒介、水源、土壤污染型)的滞后才有资格进入”月-年”窗口——这正是上一篇提到的寄生虫病(隐孢子虫、贾第虫)滞后可超8个月(PMID 38830305)的原因,也是本篇要”分层”的理由。

    “分层”不是我硬造的概念,而是数据自己呈现出来的结构。美国马萨诸塞州一项用日频数据做季节相位分析的研究(PMID 17291363),把六种肠道传染病按滞后自动聚成了两组:弯曲杆菌、沙门氏菌紧跟温度峰,滞后仅2-14天;隐孢子虫、志贺氏菌、贾第虫则滞后约40天(P<0.02)。注意,这项研究特意强调——这种差异只有在日频精度下才看得见,用月或周累积数据会被抹平。这提示”归年”这种以年为单位的粗粒度标签,天然会丢失病种间的滞后分层信息。

    滞后尺度还受”调节变量”影响——同一个病,换一个地方、换一个前提条件,窗口就变了。中国西南30县2004-2009年的每周疟疾多层分布式滞后非线性模型(PMID 24528891)发现:降雨的滞后范围明显长于其他气象因子,且”热县”的滞后比”冷县”更长;最低温在热县有更长滞后和更大相关系数。登革热更夸张:广州2005-2014年的DLNM研究(PMID 27883970)里,温度对登革热的滞后可长达141-150天(约5个月),效应峰值出现在32天和39天。而一项广州的时序研究(PMID 38106890)进一步发现”前8周累积降雨”是效应修饰者:在前期缺水时,强降雨经24-55天滞后提高登革热风险(IRR 峰1.37,滞后27天);在前期水分充足时,反而经7-121天滞后降低风险——同一个暴露、同一个病,滞后窗口的方向都被前提条件翻转了。

    那为什么有些病能”拖”,有些不能?生物学上是两条通路。直接传播型(呼吸道飞沫、粪口)的滞后主要是潜伏期+报告延迟,自然短;缓冲型(媒介、水源、土壤)多了一段”病原在环境池、媒介体内或宿主间循环”的等待期,效应被这个中间库”存起来”再释放,才会呈现月-年的长滞后。运气学说的”三年化疫”如果要找生物学落点,最可能的宿主就是这个”环境/媒介缓冲池”——它恰好与中医”湿邪黏滞、留伏而发”的直觉同构。

    所以归年增量检验的正确设计,应当分三步:第一步,先做多病种DLNM滞后曲线聚类(按效应峰值滞后时点),把病种自动分成”短滞后层/长滞后层”;第二步,只在长滞后层内检验归年后1-3年的增量(ΔAUC/互信息),并做分层meta;第三步,把短滞后层当阴性对照。 结果只会有三种:①长滞后层内普遍显著、短滞后层不显著——归年理论只在”缓冲型”病种成立,等于给运气学说画出适用边界;②两层均不显著——”三年化疫”退回文化叙事;③短滞后层反而显著——那说明抓到的是普通季节/年际气候信号,跟”年际归年”无关。

    分层不是为了给理论找台阶,而是为了让检验落在真正有信号的地方。 一个能拖到”年”的滞后,和一个只拖”周”的滞后,本来就不是同一类现象,混在一起检验,无论结果强弱都无法解读。归年理论要争的不是”灵不灵”,而是”它在哪些病上灵”——先把病种分对层,这个问题才有可能被回答。

    证据与局限

    已验证: 气候-疾病滞后尺度病种特异且量级分明——系统综述给出温度滞后谱(流感1-2周、腹泻3-6周、疟疾7-12周、登革热6-16周,降雨对媒介病8-12周)(PMID 33448301);六种肠道病按滞后自动聚成两组(2-14天 vs 约40天,仅日频可见)(PMID 17291363);疟疾滞后范围”热县>冷县”、降雨滞后最长(PMID 24528891);登革热温度滞后可达141-150天、峰值32/39天(PMID 27883970),且受”前期水分”修饰、方向可翻转(PMID 38106890)。

    部分验证: 滞后长度与传播途径(直接型 vs 缓冲型)存在生物学对应关系,但作为”分层规则”尚未被系统验证为可操作的聚类标准;长滞后病种(寄生虫>8月,PMID 38830305)与短滞后病种(肠道0-2月,PMID 36141780)的滞后谱需在同一框架下重新对照。

    待验证: “按滞后尺度对病种分层后,各层内归年增量是否呈现分层显著性”(分层DLNM聚类+层内ΔAUC/互信息+分层meta+短滞后层阴性对照);分层边界(几周为界)的稳健性;气候区×病种×滞后三重交互在分层后是否仍存在。

    相关阅读

    证据等级: B
    参考文献: PMID 33448301, PMID 17291363, PMID 24528891, PMID 27883970, PMID 38106890, PMID 38830305, PMID 36141780


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Disease-Specific Calibration of Lag Windows in Retrospective Analysis: Which Diseases Truly Exhibit a 1-3 Year Lag?

    Evidence level: B

    The phrase “three years of epidemic transformation” implies a premise: effects during an anomalous year persist into the 1st, 2nd, and 3rd subsequent years. However, modern climate epidemiology presents a “lag spectrum”—influenza lags 1–2 weeks, diarrhea 3–6 weeks, malaria 7–12 weeks, and dengue fever 6–16 weeks. If this spectrum is overlaid onto the return-year window, it becomes evident that the scale of “year” does not hold for most diseases. Therefore, stratification by disease type must precede any validation of the return-year hypothesis.

    Main Text

    The previous article addressed whether national pooling averages out regional heterogeneity. This article steps back to ask a more proximal question: Is the window of “years 1/2/3 following the return year” itself appropriate for specific diseases?

    The theory of luck bundles “rigidity and flexibility failing to hold—three years of epidemic transformation” together, implicitly assuming a slow-moving variable spanning multiple years. However, modern research has long clarified the lag scale of “climate → disease,” with highly consistent conclusions: lag duration is disease-specific, not universal. A systematic review summarizing climate-health time series studies from 2000–2019 (PMID 33448301) directly provides the lag spectrum for temperature effects: influenza ~1–2 weeks, diarrhea 3–6 weeks, malaria 7–12 weeks, and dengue fever 6–16 weeks; rainfall lags for infectious diarrhea are 4–10 weeks and can extend to 8–12 weeks for vector-borne diseases.

    When this spectrum is superimposed on the “return year” window, the conclusion becomes clear: directly transmitted diseases (influenza, gastrointestinal illnesses) exhibit lags at the “week-month” magnitude and do not reach the “year” scale. Conducting incremental tests for these diseases in “years 1/2/3 following the return year” is equivalent to searching for signals where the signal has already decayed to zero. Conversely, only buffered diseases (vector-borne, waterborne, soil-pollution related) qualify for entry into the “month-year” window—which explains why the previous article mentioned parasitic diseases (cryptosporidium, giardia) exhibiting lags exceeding 8 months (PMID 38830305), and serves as the rationale for this article’s “stratification.”

    “Stratification” is not a concept imposed arbitrarily but emerges from the structure presented by the data itself. A study in Massachusetts using daily frequency data for seasonal phase analysis (PMID 17291363) automatically clustered six enteric infectious diseases into two groups based on lag: Campylobacter and Salmonella followed temperature peaks with lags of only 2–14 days; Cryptosporidium, Shigella, and Giardia lagged approximately 40 days (P<0.02). Notably, this study emphasized that such differences are only visible at daily frequency precision; monthly or weekly aggregated data would flatten them. This suggests that the return-year label, which uses a coarse granularity of “year,” naturally loses lag-stratification information between disease types.

    Lag scales are also influenced by “moderating variables”—changing the location or premise conditions for the same disease alters the window. A multi-layer distributed lag nonlinear model (DLNM) across 30 counties in Southwest China from 2004–2009 (PMID 24528891) found that the lag range for rainfall was significantly longer than other meteorological factors, and lags were longer in “hot counties” compared to “cold counties”; minimum temperature showed longer lags and higher correlation coefficients in hot counties. Dengue fever is even more extreme: a DLNM study in Guangzhou from 2005–2014 (PMID 27883970) revealed that temperature lags for dengue fever could reach 141–150 days (approx. 5 months), with effect peaks occurring at 32 and 39 days. Furthermore, a temporal study in Guangzhou (PMID 38106890) found that “cumulative rainfall in the preceding 8 weeks” acts as an effect modifier: when prior water scarcity exists, heavy rainfall increases dengue risk after a 24–55 day lag (IRR peak 1.37, lag 27 days); when prior moisture is sufficient, it instead decreases risk after a 7–121 day lag—for the same exposure and same disease, the direction of the lag window is flipped by premise conditions.

    So why can some diseases “drag out” while others cannot? Biologically, there are two pathways. Directly transmitted diseases (respiratory droplets, fecal-oral) have short lags primarily driven by incubation periods plus reporting delays. Buffered diseases (vector-borne, waterborne, soil-related) include an additional waiting period for “pathogen circulation within environmental pools, vectors, or between hosts.” The effect is stored and released by this intermediate reservoir, resulting in long lags at the month-year scale. If the “three years of epidemic transformation” theory seeks a biological anchor, the most likely host is this “environmental/vector buffer pool”—which is isomorphic with the traditional Chinese medicine intuition of “dampness being sticky, lingering and erupting later.”

    Therefore, the correct design for incremental testing of return years should follow three steps: Step 1: Perform multi-disease DLNM lag curve clustering (by effect peak lag time) to automatically divide diseases into “short-lag strata” and “long-lag strata”; Step 2: Only test incremental return years (1–3 years) within the long-lag strata (ΔAUC/mutual information) and conduct stratified meta-analysis; Step 3: Treat the short-lag strata as negative controls. Results will fall into one of three categories: ① Significant in long-lag strata generally, not significant in short-lag strata—the return-year theory holds only for “buffered” diseases, effectively drawing its applicability boundaries; ② Not significant in both strata—”three years of epidemic transformation” retreats to cultural narrative; ③ Significant in short-lag strata—indicating that captured signals are ordinary seasonal/interannual climate signals unrelated to “interannual return years.”

    Stratification is not to find steps for a theory, but to ensure testing occurs where signals truly exist. A lag lasting “years” and one lasting only “weeks” are fundamentally different phenomena; mixing them for testing renders interpretation impossible regardless of result strength. The return-year theory does not contest whether it is “effective or not,” but rather “where it is effective.” Only by correctly stratifying disease types can this question be answered.

    Evidence and Limitations

    Verified: Disease-specific lag scales for climate-disease relationships are distinct in magnitude—systematic review provides temperature lag spectrum (influenza 1–2 weeks, diarrhea 3–6 weeks, malaria 7–12 weeks, dengue fever 6–16 weeks; rainfall for vector-borne diseases 8–12 weeks) (PMID 33448301); six enteric diseases automatically clustered into two groups based on lag (2–14 days vs. approx. 40 days, visible only at daily frequency) (PMID 17291363); malaria lag ranges show “hot counties > cold counties” with rainfall having the longest lag (PMID 24528891); dengue temperature lags reach 141–150 days with peaks at 32/39 days (PMID 27883970), modified by “prior moisture” with potential reversal of direction (PMID 38106890).

    Partially Verified: A biological correspondence exists between lag length and transmission pathways (direct vs. buffered), but as a “stratification rule,” it has not yet been systematically verified as an operational clustering standard; lag spectra for long-lag diseases (parasites >8 months, PMID 38830305) and short-lag diseases (enteric 0–2 months, PMID 36141780) require re-evaluation within a unified framework.

    To Be Verified: Whether return-year increments exhibit stratified significance within each stratum after stratification by lag scale (stratified DLNM clustering + intra-strata ΔAUC/mutual information + stratified meta-analysis + short-lag negative control); robustness of stratification boundaries (threshold in weeks); whether the climate-region × disease-type × lag triple interaction persists after stratification.

    Related Reading

    • [“Return Year”] Years 1/2/3: Can multi-disease incidence increments validate “rigidity and flexibility failing to hold”?](/research/science-yixue/yunqi-multidisease-return-year-lag-increment)
    • National Pooling vs. Stratified Climate Zones: Will heterogeneity in return-year testing be averaged out?](/research/science-yixue/yunqi-pooled-vs-climate-zone-stratified-return-year-heterogeneity)
    • Luck Long Lag (1–3 Years): Do meteorological-disease associations carry independent information?](/research/science-yixue/yunqi-long-lag-independent-information)
    • Lag Scale of “Three Years of Epidemic Transformation”: Multi-disease systematic validation](/research/science-yixue/yunqi-three-year-epidemic-lag-multidisease-validation)
    • Research Pool: Spatiotemporal Frequency Topology](/research/pools/pool-2-spatiotemporal-frequency-topology)

    Evidence Level: B
    References: PMID 33448301, PMID 17291363, PMID 24528891, PMID 27883970, PMID 38106890, PMID 38830305, PMID 36141780


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 「运气先验导航」vs「全滞后盲扫」:古人经验能当假设生成器吗?

    Evidence level: B

    古人经验的价值不在于”帮我们验证”,而在于”帮我们少找几个地方”。但这套导航到底更灵,还是只是把真信号挡在门外——可以被同一个数据集直接裁决。

    正文

    上一篇把”运气长滞后是否携带独立增量”拆成了可执行实验。但它留下了一个更底层的工程问题:你凭什么决定去看哪几个滞后窗口? 有两种截然不同的策略。第一种叫”先验导航”——按五运六气的岁运太过不及、客气加临,先推出”这几年该在哪些滞后窗里找”。第二种叫”全滞后盲扫”——不问理论,把 0 到 60 个月所有窗口全跑一遍。哪一种的发现效率更高?这不是信仰问题,是可以拿同一份数据直接比的问题。

    先看盲扫要付的代价。把滞后窗口一路拉到月年级,等于同时做了几十次假设检验。若照 α=0.05 的裸阈值判定,61 个窗口里纯靠运气就会冒出约 3 个”显著”结果。环境流行病学的标准解法是把滞后-反应关系放进分布式滞后非线性模型(DLNM),用交叉基一次性建模整个曲线,而不是逐窗口硬扫(PMID 20812303);对时空数据则用错误发现率(FDR)控制”发现里有多少是假的”——Ventrucci 等人为标化死亡比设计的就是一套贝叶斯层次 FDR 估计,用后验概率决定哪些区域”真高”(PMID 20577014)。盲扫不是错,但每多扫一个窗口,就多缴一份假阳性税。

    先验导航的诱惑正在于此:只测理论点名的少数窗口,检验次数骤减,每个窗口的功效自然抬高。但这里埋着一个致命前提。Cai 等人在全基因组假设检验里把这件事算清楚了:先验分两种——”神谕先验”知道所有真信号都落在限定子集内,那限制窗口确实省力;但一旦先验漏掉哪怕一小部分真信号,非神谕先验的表现可以比完全不用先验还差(PMID 37101120)。而且他们发现,把检验从全部缩到 100-1000 个候选基因带来的功效增益,往往还比不上队列规模每年 20-40% 的自然增长。换句话说:先验限制是把双刃剑——只有”不漏”才划算,漏了就是自残。 运气先验属于哪一类,至今没人量过。

    把知识塞进变量选择,方法学界已有成熟但代价明确的先例。Boulet 等人把临床专家的数值权重经”幂先验”并入贝叶斯变量选择,好处是能挑出纯数据驱动会漏掉的稀有变量,坏处是也会把”数据看相关、专家觉得无关”的变量剔掉(PMID 30963815)。Hill 等人则用经验贝叶斯自动加权路径/网络先验,核心动机就是防止先验设定错误反噬(PMID 22578440)。这两篇共同指向一个教训:先验能用,但必须给”先验错了”留退路。

    退路其实已经有现成工具,而且比”硬先验”更聪明。Chen 等人提出”数据自适应收缩”:把无约束 DLM(无偏但低效)往有约束 DLM(高效但可能有偏)方向收缩,让数据自己决定收缩多少、在哪几段收缩,从而在偏差与方差之间取平衡(PMID 29040386)。Antonelli 等人的 spike-and-slab 模型更进一步——把关键窗口的选择交给数据:哪些滞后时段真的关联到结局,由后验包含概率挑出来,而不是由理论预先钉死(PMID 36073640)。这相当于把”先验导航”从一把硬尺子,变成一根可弯的软绳。

    于是实验设计就清楚了:同一份多年传染病+气象长序列,跑两条通道。 通道 A——先验导航:严格按运气理论给出的窗口建模。通道 B——全滞后盲扫:0-60 个月全窗口 + BH-FDR 校正。真值来自”真信号标定模拟”:在已知的滞后位置注入已知强度的、与干支分组同相位的效应,这样才知道每个通道到底抓到了几个真的、放跑了几个。核心指标有三个:召回率(真信号抓到几成)、假阳性率、以及最关键的——“神谕分数” f,即被注入的真信号里有多大比例落在先验窗口之内。 这个 f 就是判定运气先验是”神谕”还是”非神谕”的那条线(对应 PMID 37101120 的 f 临界判据)。

    三种结局,各自对应运气理论的一种读法。其一,通道 A 召回更高、假阳性更低——先验确实是有信息量的假设生成器,古人的窗口经验被现代数据认账。其二,通道 B 反超——说明运气推导出的窗口与真实效应窗口错位,先验在这里只是多重比较的负担,”古人经验当导航”这一主张当场被否。其三,两者打平——那先验的价值就不在”缩窗口”上,而在 Cai 等人说的另一类事情上:因果结构与通路假设,那本就不是标准假设检验擅长的地方(PMID 37101120)。

    无论哪种结局,产出的都不是一句”运气准不准”,而是一个可复用的“神谕分数”标定模板——一个数字,回答”这套古代先验在漏信号这件事上到底有多可靠”。这恰好把”古人经验”重新摆回它该站的位置:它是假设生成器,不是验证机器;它负责提出”该去哪看”,现代数据负责回答”到底有没有”。这,才是科学易学桥接里两者各司其职的标准姿势。

    证据与局限

    已验证: 逐窗口扫描会累积多重比较假阳性,DLNM 交叉基可整体建模滞后曲线而不必逐窗硬扫(PMID 20812303);时空环境数据的真假发现可用贝叶斯层次 FDR 定量控制(PMID 20577014);先验限制只有在”不漏真信号”时才提升功效,非神谕先验可劣于无先验(PMID 37101120)。

    部分验证: 专家/知识先验经幂先验、经验贝叶斯并入变量选择,能捞出数据会漏的稀有变量,但存在先验误设反噬风险(PMID 30963815、PMID 22578440);数据自适应收缩与 spike-and-slab 可把”窗口选择”交给数据,作为硬先验的软替代(PMID 29040386、PMID 36073640)。

    待验证: 在真实多年传染病-气象序列上,”运气先验导航窗口 vs 0-60 个月全窗口盲扫”的头对头功效/假阳性对比,以及运气先验的”神谕分数”实测标定——全球尚无直接实验。

    相关阅读

    证据等级: B
    参考文献: PMID 37101120, PMID 20577014, PMID 20812303, PMID 30963815, PMID 22578440, PMID 29040386, PMID 36073640


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    \”Prior Navigating Luck\” vs \”Full Lag Blind Sweeping\”: Can Ancient Experience Serve as a Hypothesis Generator?

    Evidence level: B

    The value of ancient experience lies not in “verifying for us,” but in “helping us search fewer places.” Yet, is this navigation genuinely superior, or does it merely block true signals? This can be adjudicated directly using the same dataset.

    Main Text

    The previous article decomposed the question of whether “long-term luck carries independent incremental information” into an executable experiment. However, it left a more fundamental engineering question: On what basis do you decide which lag windows to examine? Two distinct strategies exist. The first is “Prior Navigational Guidance”—deducing which lag windows to search based on the excess or deficiency of the annual cycle, and the interaction of the guiding qi (Ke Qi). The second is “Full-Lag Blind Sweeping”—running all windows from 0 to 60 months without theoretical constraints. Which yields higher discovery efficiency? This is not a matter of faith, but one that can be directly compared using the same dataset.

    First, consider the cost of blind sweeping. Extending lag windows to the annual level simultaneously performs dozens of hypothesis tests. If a raw threshold of $\alpha=0.05$ is applied, approximately 3 “significant” results would emerge purely by chance among the 61 windows. The standard solution in environmental epidemiology is to incorporate lag-response relationships into a Distributed Lag Nonlinear Model (DLNM) using cross-basis functions to model the entire curve at once, rather than hard-sweeping window by window (PMID 20812303); for spatiotemporal data, the False Discovery Rate (FDR) controls the proportion of false discoveries—Ventrucci et al. designed a Bayesian hierarchical FDR estimator specifically for standardized mortality ratios, using posterior probabilities to determine which regions are “truly elevated” (PMID 20577014). Blind sweeping is not incorrect, but each additional window incurs a tax of false positives.

    The allure of prior navigation lies precisely here: by testing only the few windows named by theory, the number of tests drops drastically, naturally increasing power per window. However, this relies on a fatal premise. Cai et al. clarified this in the context of whole-genome hypothesis testing: Priors come in two types—”Oracle priors” know all true signals fall within a restricted subset, making window restriction efficient; but once the prior misses even a small portion of true signals, non-oracle priors can perform worse than using no prior at all (PMID 37101120). Furthermore, they found that the power gain from restricting tests from all candidates to 100–1000 candidate genes often falls short of the natural annual growth in cohort size of 20–40%. In other words: Prior restriction is a double-edged sword—it only pays off if it does not miss signals; missing them is self-harm. It remains unknown whether “luck priors” fall into the oracle or non-oracle category.

    Embedding knowledge into variable selection has established precedents in the methods literature with clear costs. Boulet et al. incorporated clinical expert numerical weights into Bayesian variable selection via “power priors.” The benefit is the ability to identify rare variables missed by pure data-driven approaches; the drawback is the exclusion of variables where “data show correlation but experts deem irrelevant” (PMID 30963815). Hill et al. used empirical Bayes to automatically weight path/network priors, with a core motivation to prevent the backlash of erroneous prior specification (PMID 22578440). Together, these point to a lesson: priors can be used, but a retreat path must be provided for “wrong priors.”

    Such a retreat path already exists with ready-made tools, and is arguably smarter than “hard priors.” Chen et al. proposed “data-adaptive shrinkage”: shrinking an unconstrained DLM (unbiased but inefficient) toward a constrained DLM (efficient but potentially biased), allowing data to decide the degree and location of shrinkage, thus balancing bias and variance (PMID 29040386). Antonelli et al.’s spike-and-slab model goes further—delegating window selection to the data: which lag periods truly associate with the outcome are selected by posterior inclusion probabilities, not pre-fixed by theory (PMID 36073640). This transforms “prior navigation” from a rigid ruler into a flexible soft rope.

    Thus, the experimental design is clear: Run two channels on the same multi-year infectious disease and meteorological time series. Channel A—Prior Navigation: Model strictly according to windows specified by luck theory. Channel B—Full-Lag Blind Sweeping: All windows from 0–60 months + BH-FDR correction. Truth comes from “True Signal Calibration Simulation”: injecting known-strength effects at known lag positions with phases aligned to Heavenly Stems and Earthly Branches, thereby determining how many true signals each channel captures and how many it misses. Three core metrics are required: Recall (proportion of true signals captured), False Positive Rate, and most critically—the “Oracle Score” f, representing the proportion of injected true signals falling within the prior windows. This f serves as the criterion line for judging whether luck priors are “oracle” or “non-oracle” (corresponding to the f critical criterion in PMID 37101120).

    Three outcomes correspond to three readings of luck theory. First, if Channel A shows higher recall and lower false positives—the prior is indeed an informative hypothesis generator, and ancient window experience is validated by modern data. Second, if Channel B surpasses it—indicates a misalignment between derived windows and true effect windows; here, the prior acts merely as a burden for multiple comparisons, and the claim that “ancient experience serves as navigation” is refuted. Third, if they are tied—the value of the prior lies not in “shrinking windows” but in another category described by Cai et al.: causal structure and pathway assumptions, which is not the standard domain of hypothesis testing (PMID 37101120).

    Regardless of the outcome, the product is not a statement on “how accurate luck is,” but a reusable “Oracle Score” calibration template—a number answering “how reliable this ancient prior is at missing signals.” This precisely repositions “ancient experience” to its proper role: it is a hypothesis generator, not a verification machine; it is responsible for saying “where to look,” while modern data answers “whether there is anything there.” This is the standard posture for both parties to fulfill their roles in bridging science and Yijing Studies.

    Evidence and Limitations

    Verified: Window-by-window scanning accumulates multiple comparison false positives; DLNM cross-basis functions can model lag curves globally without hard-sweeping windows (PMID 20812303); Bayesian hierarchical FDR can quantitatively control true/false discovery in spatiotemporal environmental data (PMID 20577014); Prior restriction only boosts power if “no true signals are missed,” and non-oracle priors can underperform no-prior approaches (PMID 37101120).

    Partially Verified: Expert/knowledge priors incorporated via power priors or empirical Bayes into variable selection can recover rare variables missed by data, but carry risks of prior misspecification backlash (PMID 30963815, PMID 22578440); Data-adaptive shrinkage and spike-and-slab models can delegate “window selection” to data as a soft alternative to hard priors (PMID 29040386, PMID 36073640).

    To Be Verified: Head-to-head comparisons of efficacy and false positive rates between “luck prior navigation windows vs. 0–60 month full-blind sweeping” on real multi-year infectious disease-meteorological series, and empirical calibration of the “Oracle Score” for luck priors—no direct experiments exist globally yet.

    Related Reading

    Evidence Level: B
    References: PMID 37101120, PMID 20577014, PMID 20812303, PMID 30963815, PMID 22578440, PMID 29040386, PMID 36073640


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • A 档湿玄武岩山区 Schumann 站候选清单(#24 实验启动选址工具)

    Evidence level: C

    本文件是什么:一份选址候选清单,专门为 #24 实验「山区 Schumann 站稳态基线 vs 事件期频移」服务。
    本文件不是什么:不是 Schumann 站科学综述,不是 ELF 生物效应研究,不是 σ 测量论文。
    使用方式:研究者想启动 #24 时,先读这份清单做初筛 → 再去查文献、写信给 PI、签数据共享 → 进入 E1 实验。


    一、为什么需要专门一份候选清单

    24 实验(第 27 轮已落地)核心问的是:稳态地形差异与事件期瞬态位移能不能在山区 Schumann 站和平原 Schumann 站之间被可测地分开。

    22 σ 预筛 SOP 给出三档分档:

    档 σ (S/m) δ(7.83 Hz) 1 km 山体 A(dB) 实验意义
    A 档 ≥ 10⁻² ≤ 67 km ≥ 数十 dB 地形 σ 显著 → 可改 Schumann 局部节点
    B 档 10⁻³–10⁻¹ 几十米–几百米 1–10 dB 边界档
    C 档 ≤ 10⁻⁴ ≥ 670 km < 1 dB(几乎透明) 不启动 #24

    A 档 = 局部电导率 ≥ 10⁻² S/m 的山体,主要落在活动火山带、海洋俯冲带、热液蚀变带(湿玄武岩、含矿花岗岩、海洋地壳基底)。本清单就是把这些地质条件下已经在运行或可借用的 Schumann 观测站筛出来。


    二、候选清单(按 A 档适配度排序)

    1. L’Aquila(意大利亚平宁山脉 · A 档优选)

    • 坐标:42.35°N, 13.40°E,海拨 ≈ 735 m
    • 地质:亚平宁中部,意大利中部火山-岩浆弧外缘,沉积层下伏中新世-更新世火山碎屑岩;附近有 Campi Flegrei / Vesuvius / Albani Hills 三座火山
    • 站背景:INGV 罗马 / L’Aquila 分部运营,Pc1-2 地磁脉动 + Schumann 共振长期监测;已纳入 Piersanti 2023 GRB 事件期分析
    • σ 估计:沉积层 10⁻¹–10⁰ S/m(含水火山碎屑)+ 火山基底 10⁻²–10⁻¹ S/m(湿玄武岩)→ A 档可信
    • 可得性:INGV 与国际地震/地磁学界长期共享数据,开放程度高
    • 风险点:2009 L’Aquila 地震后站点迁移过 1 次,需核对当前坐标

    2. 印度尼西亚候选站(赤道 A 档)

    • 候选坐标:
    • 地球物理站 Bandung(6.9°S, 107.6°E,海拨 768 m)
    • 候选东爪哇 / 巴厘岛附近火山观测站(Merapi / Batur / Agung 周边)
    • 地质:环太平洋火山带(Ring of Fire)核心区,全境分布更新世-全新世玄武岩、安山岩;地热梯度极高(> 100°C/km)
    • σ 估计:火山碎屑 + 玄武岩基底 + 高地热 → 局部 σ 易达 10⁻²–10⁰ S/m → A 档可信
    • 可得性:BMKG(印尼气象气候和地球物理局)有部分开放数据,需正式合作
    • 风险点:赤道电离层活动强,Schumann 振幅本身比中纬度大 1.5–2 倍,需 #24 实验设计把”地形效应”和”纬度效应”拆开

    3. 留尼汪(法属 · 印度洋火山岛 · A 档)

    • 坐标:21.1°S, 55.5°E,海拨 Piton des Neiges ≈ 3070 m
    • 地质:热点火山盾,Piton de la Fournaise 仍活跃;全岛 95% 玄武岩
    • σ 估计:湿玄武岩(高孔隙度 + 高地热)→ 10⁻² S/m 量级 → A 档可信
    • 可得性:IPGP(巴黎地球物理学院)有 Piton de la Fournaise 长期监测;Schumann 站缺,需自建
    • 风险点:远离大陆,雷电背景(Schumann 激励源)相对大陆少 30–50%,可能影响信噪比

    4. 冰岛 Vestmannaeyjar(北大西洋火山岛 · A 档)

    • 坐标:63.4°N, 20.3°W,海拨 ≈ 100 m
    • 地质:大西洋中脊出露,全岛火山玄武岩(Heimaey 1973 喷发仍有熔岩流冷却记录)
    • σ 估计:湿玄武岩 → 10⁻² S/m 量级 → A 档可信
    • 可得性:冰岛气象局 + 大学有部分 ELF/VLF 监测;专门 Schumann 站缺
    • 风险点:高纬度 63°N 接近极光带,地磁噪声本底比中纬度大 1.5–3 倍

    5. 参考平原对照(必须保留)

    • Nagycenk(匈牙利):46.9°N, 16.4°E,平原农地,无显著地质导电结构;历史 Schumann 数据丰富(> 30 年连续)
    • Portishead(英国):51.5°N, 2.8°W,海拨 50 m,平原,2021 MDPI Atmosphere 文献有长时段数据
    • 意义:作为地形”零值”对照,不进入 A 档但必须保留

    三、排除清单(避免候选池膨胀)

    站 排除理由
    庐山 / 黄山(中国名山) 无公开 Schumann 长期监测,自建成本高;非火山地质 → σ 估计 B 档以下
    乞力马扎罗 单一孤峰(5895 m),无现成 ELF 站;地质上火山但站点运营困难
    日本富士山 JMA 监测以火山地震为主,Schumann 数据未公开;行政准入复杂
    阿尔卑斯高山站 主要为冰川 / 大气监测,地形 σ 估计 B 档(变质岩基底),与 A 档判别力不足

    四、与 #24 实验设计的衔接

    环节 本清单提供 #24 文档需进一步落实
    站位选择 4 个 A 档候选 + 2 个平原对照 与各站 PI 邮件沟通 + 数据使用协议
    σ 估计 工程估算 σ≥10⁻² S/m 必须做现场 MT 三点跨断面实测(#14 阶段 3 复用)
    实验对照 A 档(4)vs 平原(2) LMM 4 地形档 × 4–6 月数据
    数据基线 Portishead / Nagycenk 文献已有 与 #18 启动时一并补
    风险 赤道电离层 / 高纬度极光 在 #24 实验设计 §5 已写

    五、证据与局限

    本清单能保证的:
    – 4 个 A 档候选的地质 σ 估计符合”湿玄武岩 / 活动火山带”已知工程范围(干玄武岩 10⁻⁴–10⁻² S/m,湿玄武岩 10⁻²–10⁰ S/m)
    – 平原对照站的 Schumann 数据可获得性有文献和机构公开度支撑
    – 与 #22 三档分档、#23 σ 量级过滤 SOP、#16 δ 表方法论一致

    本清单不能保证的:
    – 4 个 A 档候选没有任何一个已被实测 σ ≥ 10⁻² S/m(”估计” ≠ “实测”)
    – Schumann 站数据使用涉及机构合作、行政准入、设备状态——清单只是初筛
    – A 档不一定有显著 Δf_steady(#17 FEM 需先跑通)

    证伪条件:若 #24 阶段 1 在 4 个 A 档候选中跑 LMM,至少 2 个站的稳态 Δf 与对照平原站差异不显著(< 0.1 Hz) → 本清单的”地质 σ 估计”判据不充分,应降级到 #22 全部 A+B 档重测。


    六、相关阅读

    • 24 实验设计:08-ops/experiments/pool4-schumann-steady-state-vs-transient-displacement.md(5 站 LMM + 4 地形档 + 证伪四象限)

    • 22 σ 分档:science-yixue/mountain-sample-priority-sigma-pre-screening-sop.md

    • 23 σ 量级过滤 SOP:science-yixue/sigma-magnitude-filter-sop-five-pools-environmental-frequency.md

    • 16 δ 表:science-yixue/skin-depth-basement-building-mu-metal-table.md

    • 17 Schumann FEM:science-yixue/schumann-resonance-cavity-mode-topography.md

    • 21 SOP 三轨流程:science-yixue/searxng-cross-language-search-noise-filter-sop.md

    • 关联深挖 #25 立项说明:08-ops/pools/pool4-environment-backlog.md 第 28 轮

    参考文献

    • Piersanti et al. 2023. Schumann Resonance Frequency Increase Following Recurrent Coronal Mass Ejections. Nature Communications 14:7750. DOI: 10.1038/s41467-023-42551-5(含 L’Aquila 站)
    • Di Mauro et al. 2021. Geomagnetic Field Analysis at Lampedusa Island. Annals of Geophysics(与 L’Aquila 罗马对照,意大利群岛 ELF/ULF 基线)
    • Price C. 2016. ELF Electromagnetic Waves from Lightning: The Schumann Resonances. (Schumann 站位综述基线,Portishead / Nagycenk 等长时段站点)
    • Oper 2015 Project / L’Aquila E27 站 ELF 标定: PMC10007101(含 L’Aquila 站 2009–2015 标定记录与雷电 ELF 带宽)

    声明:以上 4 条参考文献是 Schumann 山区站 vs 平原站对比的间接锚点(含 1 个火山/亚平宁 A 档站 + 1 个平原对照),本清单不直接构成“山区 Schumann 站共性差异”的科学结论。该结论需要 #18 综述 + #24 实测联合完成。


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Candidate List for Schumann Station in a Humid Basaltic Mountainous Area (No. 24 Experimental Launch Site Selection Tool)

    Evidence level: C

    What this document is: A site-selection candidate list, specifically designed to support Experiment #24: “Steady-state topography differences vs. transient displacement during event periods.”
    What this document is not: Not a scientific review of Schumann stations, not an ELF biological effects study, not a $\sigma$ measurement paper.
    Usage: When researchers initiate Experiment #24, first read this list for preliminary screening $\rightarrow$ then consult literature, contact PIs, sign data-sharing agreements $\rightarrow$ proceed to E1 experiment.


    I. Why a Dedicated Candidate List is Needed

    Experiment #24 (Round 27 already implemented) addresses the core question: Can steady-state topographic differences and transient displacements during event periods be measurably distinguished between mountain Schumann stations and plain Schumann stations?

    The #22 $\sigma$ pre-screening SOP provides three tiers of classification:

    Tier $\sigma$ (S/m) $\delta$(7.83 Hz) 1 km Mountain A(dB) Experimental Significance
    Grade A $\ge$ 10⁻² $\le$ 67 km $\ge$ tens of dB Significant terrain $\sigma$ $\rightarrow$ feasible modification of local Schumann nodes
    Grade B 10⁻³–10⁻¹ Tens to hundreds of meters 1–10 dB Boundary tier
    Grade C $\le$ 10⁻⁴ $\ge$ 670 km < 1 dB (nearly transparent) Do not initiate #24

    Grade A corresponds to mountainous terrain with local electrical conductivity $\ge$ 10⁻² S/m, primarily found in active volcanic belts, subduction zones, and hydrothermal alteration zones (wet basalt, mineralized granite, oceanic crust basement). This list screens for Schumann observation stations that are already operational or available for reuse under these geological conditions.


    II. Candidate List (Sorted by Grade A Suitability)

    1. L’Aquila (Apennine Mountains, Italy · Grade A Preferred)

    • Coordinates: 42.35°N, 13.40°E, Elevation $\approx$ 735 m
    • Geology: Central Apennines, outer edge of the Italian central volcanic-magmatic arc; Neogene-Pleistocene volcanic tuffs underlying sedimentary layers; nearby Campi Flegrei, Vesuvius, and Albani Hills volcanoes
    • Station Background: Operated by INGV Rome/L’Aquila divisions; long-term monitoring of Pc1-2 geomagnetic pulsations + Schumann resonance; already included in Piersanti 2023 GRB event-period analysis
    • $\sigma$ Estimate: Sedimentary layer 10⁻¹–10⁰ S/m (water-bearing volcanic tuffs) + volcanic basement 10⁻²–10⁻¹ S/m (wet basalt) $\rightarrow$ Grade A credible
    • Availability: INGV shares data internationally with the seismological/geomagnetic community; high openness
    • Risk Point: Station relocated once after the 2009 L’Aquila earthquake; current coordinates must be verified

    2. Indonesia Candidate Stations (Equatorial Grade A)

    • Candidate Coordinates:
    • Geophysical station Bandung (6.9°S, 107.6°E, Elevation 768 m)
    • Candidate volcanological observation stations near East Java/Bali (Merapi / Batur / Agung vicinity)
    • Geology: Core of the Ring of Fire; updated basalt and andesite distributed throughout; extremely high geothermal gradient (> 100°C/km)
    • $\sigma$ Estimate: Volcanic tuffs + basalt basement + high geothermal activity $\rightarrow$ local $\sigma$ can reach 10⁻²–10⁰ S/m $\rightarrow$ Grade A credible
    • Availability: BMKG (Indonesian Meteorology, Climatology, and Geophysics Agency) has partial open data; formal cooperation required
    • Risk Point: Strong equatorial ionospheric activity; Schumann amplitudes are inherently 1.5–2 times larger than mid-latitudes; #24 experimental design must separate “terrain effects” from “latitude effects”

    3. Réunion (French · Indian Ocean Volcanic Island · Grade A)

    • Coordinates: 21.1°S, 55.5°E, Elevation Piton des Neiges $\approx$ 3070 m
    • Geology: Hotspot volcanic shield; Piton de la Fournaise remains active; 95% of the island is basalt
    • $\sigma$ Estimate: Wet basalt (high porosity + high geothermal activity) $\rightarrow$ 10⁻² S/m magnitude $\rightarrow$ Grade A credible
    • Availability: IPGP (Paris Institute of Geophysics) has long-term monitoring at Piton de la Fournaise; Schumann station missing, requires construction
    • Risk Point: Far from continents; lightning background (Schumann excitation source) is 30–50% lower than on continents; may affect signal-to-noise ratio

    4. Iceland Vestmannaeyjar (North Atlantic Volcanic Island · Grade A)

    • Coordinates: 63.4°N, 20.3°W, Elevation $\approx$ 100 m
    • Geology: Mid-Atlantic Ridge exposure; volcanic basalt throughout (Heimaey 1973 eruption still has lava flow cooling records)
    • $\sigma$ Estimate: Wet basalt $\rightarrow$ 10⁻² S/m magnitude $\rightarrow$ Grade A credible
    • Availability: Icelandic Meteorological Office and universities have partial ELF/VLF monitoring; dedicated Schumann station missing
    • Risk Point: High latitude 63°N approaches the auroral zone; geomagnetic noise background is 1.5–3 times larger than mid-latitudes

    5. Reference Plain Controls (Must Retain)

    • Nagycenk (Hungary): 46.9°N, 16.4°E, plain agricultural land, no significant geological conductive structures; rich historical Schumann data (> 30 years continuous)
    • Portishead (UK): 51.5°N, 2.8°W, Elevation 50 m, plain; long-term data documented in 2021 MDPI Atmosphere literature
    • Significance: Serve as a “zero-value” terrain control; not entering Grade A but must be retained

    III. Exclusion List (Prevent Candidate Pool Inflation)

    Station Reason for Exclusion
    Mount Lu / Mount Huangshan (China famous mountains) No public long-term Schumann monitoring; high cost for self-construction; non-volcanic geology $\rightarrow$ $\sigma$ estimate below Grade B
    Kilimanjaro Single isolated peak (5895 m), no existing ELF station; volcanic geology but operational difficulties for stations
    Mount Fuji (Japan) JMA monitoring focuses on volcanic earthquakes; Schumann data not public; complex administrative access
    Alpine High-altitude Stations Primarily for glacial/atmospheric monitoring; terrain $\sigma$ estimate Grade B (metamorphic basement), insufficient discrimination against Grade A

    IV. Integration with #24 Experimental Design

    Stage Provided by This List To Be Finalized in #24 Documents
    Site Selection 4 Grade A candidates + 2 plain controls Email communication with station PIs + data usage agreements
    $\sigma$ Estimate Engineering estimate $\sigma \ge$ 10⁻² S/m Must conduct on-site MT three-point cross-section measurements (reuse from #14 Phase 3)
    Experimental Control Grade A (4) vs Plain (2) LMM 4 terrain tiers $\times$ 4–6 months of data
    Data Baseline Portishead / Nagycenk literature available Supplement alongside #18 initiation
    Risks Equatorial ionosphere / High-latitude aurora Already documented in #24 Experimental Design §5

    V. Evidence and Limitations

    What this list may support:
    – The geological $\sigma$ estimates for the 4 Grade A candidates align with known engineering ranges for “wet basalt / active volcanic belts” (dry basalt 10⁻⁴–10⁻² S/m, wet basalt 10⁻²–10⁰ S/m)
    – Availability of Schumann data at plain control stations is supported by literature and institutional openness
    – Consistency with #22 three-tier classification, #23 $\sigma$ magnitude filtering SOP, and #16 $\delta$ table methodology

    What this list cannot may support:
    – None of the 4 Grade A candidates has been measured with $\sigma \ge$ 10⁻² S/m (“estimate” $\neq$ “measurement”)
    – Schumann station data usage involves institutional cooperation, administrative access, and equipment status; this list is merely a preliminary screen
    – Grade A does not necessarily yield significant $\Delta f_{steady}$ (#17 FEM must be run first)

    Falsification Condition: If Phase 1 of #24 runs LMM across the 4 Grade A candidates, and at least 2 stations show non-significant steady-state $\Delta f$ differences compared to plain controls (< 0.1 Hz) $\rightarrow$ the “geological $\sigma$ estimate” criterion in this list is insufficient; downgrade to re-measure all Grade A+B from #22.


    VI. Related Reading

    • Experiment #24 Design: 08-ops/experiments/pool4-schumann-steady-state-vs-transient-displacement.md (5-station LMM + 4 terrain tiers + falsification quadrants)
    • 22 $\sigma$ Tiering: science-yixue/mountain-sample-priority-sigma-pre-screening-sop.md

    • 23 $\sigma$ Magnitude Filtering SOP: science-yixue/sigma-magnitude-filter-sop-five-pools-environmental-frequency.md

    • 16 $\delta$ Table: science-yixue/skin-depth-basement-building-mu-metal-table.md

    • 17 Schumann FEM: science-yixue/schumann-resonance-cavity-mode-topography.md

    • 21 SOP Three-Track Process: science-yixue/searxng-cross-language-search-noise-filter-sop.md

    • Related Deep Dive #25 Project Rationale: 08-ops/pools/pool4-environment-backlog.md Round 28

    References

    • Piersanti et al. 2023. Schumann Resonance Frequency Increase Following Recurrent Coronal Mass Ejections. Nature Communications 14:7750. DOI: 10.1038/s41467-023-42551-5 (includes L’Aquila station)
    • Di Mauro et al. 2021. Geomagnetic Field Analysis at Lampedusa Island. Annals of Geophysics (contrast with L’Aquila Rome, Italian archipelago ELF/ULF baseline)
    • Price C. 2016. ELF Electromagnetic Waves from Lightning: The Schumann Resonances. (Schumann station positioning baseline, long-term sites like Portishead/Nagycenk)
    • Oper 2015 Project / L’Aquila E27 Station ELF Calibration: PMC10007101 (includes L’Aquila station 2009–2015 calibration records and lightning ELF bandwidth)

    Disclaimer: The above four references serve as indirect anchors for Schumann mountain stations vs. plain station comparisons (including 1 volcanic/Apennine Grade A station + 1 plain control). This list does not directly constitute a scientific conclusion on “common mountain Schumann station differences.” Such conclusions require joint completion via #18 review + #24 measurement.

  • HRI 谐波 C2/C3 与心气虚严重程度分级的剂量-反应关系

    Evidence level: B

    导言

    心气虚是中医辨证中最常用于评估心功能状态的核心证候。传统临床判断依赖医家指下经验,分级(轻/中/重)一致性受操作者影响显著。王唯工等提出的脉搏谐波理论(Harmonic Ratio Index, HRI)将桡动脉压力波分解为 10 个谐波分量(记为 C0–C10),其中 C0 为直流分量反映总脉能,C1–C10 分别代表不同脏腑器官的谐波贡献度。该理论的核心假说为:C1 关联肝、C2 关联心、C3 关联脾、C4 关联肺、C5–C10 关联肾与三焦。在此框架下,C2 与 C3 的比值(或绝对强度)随证候严重程度是否呈现可量化的剂量-反应曲线,是把中医辨证从”经验判断”推进到”频率指纹”的关键证据链。

    核心证据回顾

    1. 谐波分析在中医辨证中的整体应用(系统综述)

    Hsiao 等 2024 年发表于 J Tradit Complement Med 的系统综述(PMID 38481553,1990–2021 年 50 篇研究报告)确认:肝功能障碍与 C1 变化相关;糖尿病患者出现 C3(脾相关)减弱;癌症患者在 C1、C2 呈现肝肾阴虚的谐波特征;心血管疾病患者则表现为与肝经相关的第一谐波升高。综述明确指出”谐波强度模式在不同生理与病理状态下具有可区分特征”,但同时强调”样本量小、研究异质性大”是当前主要局限——这恰好是本课题关心的”剂量-反应关系”在严谨统计框架下尚未被独立验证的原因。

    2. 心血管疾病群体的谐波特征

    Chang 等 2016 年发表于 Microvasc Res 的研究(PMID 27067750)对代谢综合征(MetS)患者进行 PPG 谐波分析,结果显示 MS 组 C1、C2 显著减小,而 C4–C9 显著或趋势性增大;C_n 标准差与变异系数在 MS 组最大。该研究提示:C1/C2 减弱与心血管功能失调存在关联,但研究设计为横断面、单一中心(n=65),无法建立严重程度分级与谐波比率的连续曲线。

    3. 五脏谐波—ML 分类的可行性

    Wu 等 2022 年发表于 Sci Rep(PMID 36509825)使用 40 项谐波指标(振幅比例 + 变异系数 + 相位 + 标准差)对肌少症进行 LDA 分类,AUC=0.77(评分系统 AUC=0.83)。这证明:多谐波特征联合可在中等样本下达到可接受的判别力——本课题的关键不是单谐波是否有用,而是如何用 C2/C3 这一对特定谐波比量化心气虚分级。

    4. 谐波-证候的双向标定基础

    Chen 等 2020 年发表于 Evid Based Complement Alternat Med(PMID 33082826)证实”放松活动增加高次谐波、应激活动增加低次谐波”,提示生理状态变化会通过 C1–C10 比例关系反映;Wu 等 2016 年 PLoS One(PMID 27304979)的 Synchrosqueezing Transform 方法学为心气虚分级的频域精细分析提供了工具基础;Lin 等 2021 年 Sci Rep(PMID 33903610)将 40 项谐波指标用于阿尔茨海默病分类,AUC≈0.85,进一步支持”多谐波组合 + ML”范式在不同脏腑系统间的可迁移性。

    当前研究的关键缺口

    将上述证据综合,可识别出 #88 课题(心气虚轻/中/重三级剂量-反应)的三个具体证据缺口:

    1. 缺少”单证候-单谐波对”的剂量-反应曲线:既有研究多为二分类(病/对照)或三分类(健康/前驱/病),未在中医辨证内部对单一证候做”轻-中-重”分级的连续建模。
    2. C2/C3 比的混杂校正缺失:心气虚常合并气阴两虚、心脾两虚,C2 与 C3 同时受多脏腑功能影响;既有研究未控制合并证候,剂量-反应关系被合并效应污染。
    3. 方法学异质性使 Meta-分析不可行:综述(PMID 38481553)已声明”测量方法与人群异质性”是主要缺陷;不同采集设备(PPG / 压力传感)、不同按压力度、不同样本数都限制了谐波比值跨研究的可比性。

    假设与剂量-反应预期

    基于谐波理论与现有证据,本课题的工作假设为:

    在控制年龄、性别、BMI、合并证候、采集压力后,C2(心相关)/C3(脾相关)比值随中医心气虚严重程度(轻→中→重)单调递减,斜率反映心气”功能储备”消耗程度。

    该假设的合理性来自:心气虚加重时,心脏泵功能下降 → 桡动脉波 C2 分量绝对强度下降;脾虚共病则同时拉低 C3。若 C2/C3 同步下降,则比值变化可作为”心气虚主导” vs “心脾两虚”的鉴别指标;若C2 下降而 C3 不变(即 C2/C3 下降),则比值是心气虚特异的剂量-反应标志。

    证据边界与局限

    • 现有 PMID 证据均为横断面/回顾性研究,无法支持因果推断;
    • 心气虚诊断标准在不同辨证规范(1993 / 2017 / 2023)中存在差异,会导致同群体分级标签漂移——本课题需要在实验设计中明确锚定哪一版标准;
    • 谐波-证候映射目前依赖”经脉-脏腑-谐波”假说(E2 间接证据链),并非已证因果;
    • C2/C3 比值对采集压力敏感,必须配 PCHR 压力补偿(参见本池 #32 课题)才能保证跨研究可比性。

    拟验证的下一步

    本课题推荐执行”三阶段、单中心前瞻性 + 多中心外部验证”研究(详见配套实验设计 P3-EXP-023):

    • 阶段一:N=60 健康对照 vs N=180 心气虚(轻/中/重各 60)单中心,采集 PPG + 桡动脉压力波;
    • 阶段二:N=300 大样本验证剂量-反应曲线,纳入混杂校正(年龄/性别/BMI/合并证候);
    • 阶段三:N=120 多中心外部验证 + PCHR 压力补偿的跨设备迁移性检验。

    预期产物:C2/C3 比值的轻/中/重三段参考区间 + 调整混杂后的趋势 p 值 + 跨设备 ICC。

    证据等级说明

    • A 级:RCT 或多中心前瞻性队列直接证明(当前缺失)
    • B 级:多研究间接证据链 + 系统综述支持(本课题当前位置)
    • C 级:单中心横断面或回顾性研究(多数现有 PMID 落在此层)

    本课题在取得阶段二、阶段三证据前,应保持 B 级(间接证据链);阶段二完成后可升至 A-(单中心前瞻);阶段三完成后可挑战 A 级。

    相关阅读

    • 本池 #87 脉搏谐波频谱与时域特征的证候区分能力循证映射
    • 本池 #32 PCHR 算法的深度学习自适应压力补偿(实验编号 P3-EXP-016)
    • 本池 #29 压力归一化算法与跨操作者 HRI 可比性验证(PMID 27861248, 27314975, 31730401, 33901427, 1402.6649)
    • 本池 #23 二十八脉 HRI 操作者间信度的操作员培训标准化

    免责声明

    本文为研究综述与课题框架,不构成医疗建议。心气虚辨证与治疗须由具备执业资格的中医师进行;HRI 谐波分析为研究工具,临床诊断需结合症状、舌脉、实验室检查等综合判断。


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Dose-response relationship between HRI harmonic C2/C3 and severity grading of heart qi deficiency

    Evidence level: B

    Introduction

    Heart Qi Deficiency is a core syndrome pattern in Traditional Chinese Medicine (TCM) diagnosis most frequently utilized to assess cardiac function status. Traditional clinical judgment relies heavily on practitioner experience under the pulse, and consistency in grading (mild/moderate/severe) varies significantly among operators. The Harmonic Ratio Index (HRI) theory proposed by Wang Weigong et al. decomposes the radial artery pressure wave into 10 harmonic components (denoted as C0–C10), where C0 represents the DC component reflecting total pulse energy, and C1–C10 represent harmonic contributions from different zang-fu organs. The core hypothesis of this theory is: C1 correlates with Liver, C2 with Heart, C3 with Spleen, C4 with Lung, and C5–C10 with Kidney and San Jiao. Within this framework, establishing whether the ratio (or absolute intensity) of C2 and C3 exhibits a quantifiable dose-response curve with syndrome severity is the critical evidence chain advancing TCM diagnosis from “empirical judgment” to “frequency fingerprinting.”

    Core Evidence Review

    1. Overall Application of Harmonic Analysis in TCM Diagnosis (Systematic Review)

    A systematic review by Hsiao et al. published in J Tradit Complement Med (PMID 38481553; 50 reports from 1990–2021) confirmed: C1 changes correlate with liver dysfunction; C3 (spleen-related) attenuation occurs in diabetic patients; cancer patients exhibit harmonic features of liver-kidney Yin deficiency in C1 and C2; cardiovascular disease patients manifest elevated first harmonics related to the Liver meridian. The review explicitly states that “harmonic intensity patterns exhibit distinguishable characteristics under different physiological and pathological states,” yet simultaneously emphasizes that “small sample sizes and high study heterogeneity” remain the primary limitations—precisely the reason why the “dose-response relationship” of interest in this topic has not yet been independently verified within a rigorous statistical framework.

    2. Harmonic Characteristics in Cardiovascular Disease Populations

    Chang et al. (2016; Microvasc Res; PMID 27067750) conducted PPG harmonic analysis on patients with Metabolic Syndrome (MetS), showing significant reductions in C1 and C2, while C4–C9 showed significant or trend-wise increases; the standard deviation and coefficient of variation for $C_n$ were maximal in the MetS group. This study suggests a association between C1/C2 attenuation and cardiovascular dysfunction, but the cross-sectional, single-center design (n=65) precludes establishing a continuous curve correlating severity grading with harmonic ratios.

    3. Feasibility of Five-Zang Harmonics–ML Classification

    Wu et al. (2022; Sci Rep; PMID 36509825) utilized 40 harmonic indicators (amplitude ratios + coefficient of variation + phase + standard deviation) for Linear Discriminant Analysis (LDA) classification in sarcopenia, achieving an AUC of 0.77 (scoring system AUC=0.83). This demonstrates that combined multi-harmonic features can achieve acceptable discriminative power in moderate sample sizes—the key to this topic is not whether a single harmonic is useful, but how the specific C2/C3 harmonic ratio can quantify Heart Qi Deficiency grading.

    4. Bidirectional Calibration Basis for Harmonics-Syndrome Mapping

    Chen et al. (2020; Evid Based Complement Alternat Med; PMID 33082826) confirmed that “relaxation activities increase higher harmonics while stress activities increase lower harmonics,” suggesting physiological state changes reflect through C1–C10 proportional relationships; Wu et al. (2016; PLoS One; PMID 27304979) provided methodological foundations via Synchrosqueezing Transform for fine-frequency domain analysis of Heart Qi Deficiency grading; Lin et al. (2021; Sci Rep; PMID 33903610) applied 40 harmonic indicators for Alzheimer’s disease classification with AUC≈0.85, further supporting the transferability of the “multi-harmonic combination + ML” paradigm across different zang-fu systems.

    Key Gaps in Current Research

    Synthesizing the above evidence reveals three specific evidence gaps for Project #88 (Heart Qi Deficiency mild/moderate/severe three-level dose-response):

    1. Missing Dose-Response Curve for “Single Syndrome-Single Harmonic Pair”: Existing studies are predominantly binary (disease/control) or ternary (healthy/pre-clinical/disease), lacking continuous modeling of “mild-moderate-severe” grading within a single syndrome pattern in TCM diagnosis.
    2. Lack of Confounding Correction for C2/C3 Ratio: Heart Qi Deficiency often coexists with Qi-Yin Deficiency or Heart-Spleen Deficiency; both C2 and C3 are simultaneously influenced by multiple organ functions. Existing studies fail to control for comorbid syndromes, contaminating the dose-response relationship with combined effects.
    3. Methodological Heterogeneity Prevents Meta-Analysis: The review (PMID 38481553) already declares “measurement methods and population heterogeneity” as primary defects; differences in acquisition devices (PPG/pressure sensors), varying palpation pressures, and disparate sample sizes limit cross-study comparability of harmonic ratios.

    Hypotheses and Dose-Response Expectations

    Based on harmonic theory and existing evidence, the working hypothesis for this project is:

    Controlling for age, gender, BMI, comorbid syndromes, and acquisition pressure, the C2 (Heart-related) / C3 (Spleen-related) ratio will monotonically decrease with increasing severity of Heart Qi Deficiency (mild → moderate → severe), with the slope reflecting the degree of “functional reserve” consumption in Heart Qi.

    The rationale for this hypothesis is: as Heart Qi Deficiency worsens, cardiac pumping function declines → absolute intensity of the C2 component in the radial artery wave decreases; concurrent Spleen Deficiency simultaneously lowers C3. If C2/C3 decrease synchronously, the ratio change serves as a differentiation indicator between “Heart Qi Deficiency dominant” vs. “Heart-Spleen Deficiency”; if C2 decreases while C3 remains unchanged (i.e., C2/C3 decreases), the ratio becomes a specific dose-response marker for Heart Qi Deficiency.

    Evidence Boundaries and Limitations

    • Existing PMID evidence consists of cross-sectional or retrospective studies, unable to support causal inference;
    • Diagnostic criteria for Heart Qi Deficiency vary across different TCM diagnostic standards (1993 / 2017 / 2023), leading to label drift within the same population—this project must explicitly anchor which standard version in experimental design;
    • Harmonic-syndrome mapping currently relies on the “Meridian-Zang-Fu-Harmonic” hypothesis (E2 indirect evidence chain), not yet proven as causality;
    • The C2/C3 ratio is sensitive to acquisition pressure and must be paired with PCHR pressure compensation (referenced in Pool #32 project) to ensure cross-study comparability.

    Proposed Next Steps

    This project recommends executing a “three-stage, single-center prospective + multi-center external validation” study (see accompanying experimental design P3-EXP-023):

    • Stage One: N=60 healthy controls vs. N=180 Heart Qi Deficiency (60 mild, 60 moderate, 60 severe) single-center; collect PPG + radial artery pressure waves;
    • Stage Two: N=300 large-sample validation of the dose-response curve, incorporating confounding adjustments (age/gender/BMI/comorbid syndromes);
    • Stage Three: N=120 multi-center external validation + cross-device transferability testing with PCHR pressure compensation.

    Expected outputs: Three-segment reference intervals for C2/C3 ratio (mild/moderate/severe) + trend p-values after confounding adjustment + cross-device ICC.

    Evidence Level Description

    • Level A: RCT or multi-center prospective cohort directly proving (currently missing)
    • Level B: Multi-study indirect evidence chain + systematic review support (current position of this project)
    • Level C: Single-center cross-sectional or retrospective studies (most existing PMIDs fall into this layer)

    Prior to obtaining Stage Two and Stage Three evidence, this project should maintain Level B (indirect evidence chain); upon completion of Stage Two, it may ascend to A- (single-center prospective); completion of Stage Three may challenge Level A.

    Related Reading

    • Pool #87: Evidence mapping on discriminative ability of pulse harmonic spectrum vs. time-domain features for syndrome differentiation
    • Pool #32: Deep learning adaptive pressure compensation algorithm for PCHR (Experiment ID P3-EXP-016)
    • Pool #29: Pressure normalization algorithm and cross-operator HRI comparability validation (PMID 27861248, 27314975, 31730401, 33901427, 1402.6649)
    • Pool #23: Standardization of operator training for inter-operator reliability of 28-pulse HRI

    Disclaimer

    This article serves as a research review and project framework, does not constitute medical advice. Heart Qi Deficiency diagnosis and treatment must be conducted by licensed TCM practitioners; HRI harmonic analysis is a research tool, and clinical diagnosis requires comprehensive judgment combining symptoms, tongue/pulse examination, and laboratory tests.

  • 长短滞后分离模型的’长滞后独立增量’检验:长滞后项真的带来独立信息,还是短期天气换皮?

    Evidence level: B

    “三年化疫”要成为可检验的科学假说,必须先过一道窄门:把滞后窗口拉长之后,多出来的预测力必须是”独立”的——不是短滞后的换皮,也不是统计噪声的错觉。

    正文

    上一篇讨论过运气长滞后(1-3 年)气象-疾病关联是否携带独立于短滞后的信息量。那篇的结论是:这个问题”可以被检验”,但还没有人严格检验过。这一篇把那个待验证的缺口拆成一个具体、可执行的实验设计:在同一份数据上分别拟合”短滞后 DLNM”和”短+长滞后 DLNM”,用时间外验证比较两者,看长滞后项能不能带来显著且可复现的增量。

    先说清”长滞后模型”在方法学上是什么。现代暴露-滞后-反应分析的主力框架是分布式滞后非线性模型(DLNM),核心构件叫”交叉基”(cross-basis):一个维度建模暴露量本身的大小(非线性剂量-反应),另一个维度建模滞后时间(滞后-反应),两者张成一张二维风险曲面。Gasparrini 在方法学奠基论文里展示了这套框架能把滞后窗口拉到几十年量级——他用铀矿工队列的氡暴露数据,估计出滞后 0-40 年的肺癌死亡风险曲线(PMID 24027094)。也就是说,”把滞后拉到月年级”在统计工具上不是难点,DLNM 天生支持。

    难点在对照设计。假设研究者不建长滞后模型,而是把”过去一年平均气温”当成一个单一变量塞进模型——这是最常见的偷懒做法。Gasparrini 专门用模拟研究比较了两种策略:当存在真实的长滞后关联时,移动平均(moving average)式的粗略近似会带来系统性偏倚,而 DLNM 即使在强季节趋势下也能给出几乎无偏的估计和接近名义水平的置信区间;只有滞后确实很短、且区间设定正确时,移动平均才是可接受的替代(PMID 27400064)。这条结论直接决定模型选择:长滞后项必须用样条式滞后-反应函数建模,不能用简单平均糊弄过去。

    那真实世界里,”长滞后比短滞后多知道东西”有没有先例?有一个很干净的样本。瑞典五个队列的研究用年均 PM2.5 与黑碳(BC),把暴露滞后一次设到 1-10 年,同时对照”逐年单滞后”模型和”1-5 年/6-10 年移动平均”模型。结果:对缺血性心脏病,DLNM 没发现滞后-反应;但对卒中,DLNM 提示 2-4 年以及 8-9 年的暴露窗口存在风险升高,而逐年单滞后模型什么关联都没显示出来(PMID 35270332)。这正是”长滞后携带短滞后模型捕捉不到的信息”在真实数据上的实例——它的价值不在结论多强,而在证明”用错滞后窗口会导致系统性漏检”。

    传染病侧也有可用实例。荷兰轮状病毒的滞后非线性时间序列分析,把气象因素与传播关系放进带季节项、并用自回归移动平均误差校正残差序列相关的分布式滞后模型,最终识别出易感人群比例与气温对传播的驱动作用(PMID 28333764)。它提示:长滞后建模在传染病里必须同时处理季节项与残差自相关,否则”长滞后显著”很可能只是季节性被重复计入两次。

    接下来是关键一节:怎么判断增量是”独立”的? 必须区分两种可能。其一,长滞后只是短滞后的”换皮”——宏观气候周期与短期天气高度相关,长滞后项捡到的其实是短期天气早已编码的信息;其二,长滞后确实携带额外信息,比如经年累月的免疫、营养或媒介生态变化。区分两者的统计纪律有三条。

    第一,比较指标要选”不容易骗人”的。一个流传很广的指标叫净重分类改善(NRI),但 Pepe 等人用 10000 人、事件率 10.2%、四个毫无预测能力的假生物标志物做重复模拟,发现 NRI 的假阳性率高得离谱:训练集 63.0%、测试集 18.8%-34.4%,而用 AUC 变化几乎没有假阳性,用似然比检验则稳定在约 5%(PMID 24681599)。所以本实验的增量必须用 ΔAUC / 似然比 / Brier 分解衡量,坚决不用 NRI 的 p 值下结论。

    第二,验证必须是时间外验证,不是训练集自夸——模型在训练段拟合、在留出的未来段重估增量,才排除过拟合。

    第三,置换检验必须尊重时间结构。Linden 提出的做法是在中断时间序列里迭代地把每个”未被干预”的单元伪装成”被干预”,看伪处理单元是否也能估出显著效应;若伪单元也显著,真单元那点显著就不能归给干预(PMID 29460383)。照搬到这里就是:置换不能简单打乱时间顺序(那会破坏自相关、制造虚高的假阳性),而要在保留自相关与季节结构的前提下重排”运气分组标签”。

    最后是与运气理论的接口。干支是固定历法标签,用它分组不会像很多机器学习特征那样泄露未来信息,所以”长滞后独立增量检验”在方法学上是干净的。而运气理论本身提供的先验(岁运太过不及、客气加临)恰好可以当作”该去哪些滞后窗口找信号”的导航——这正是下一篇要处理的对照:先验导航 vs 0-60 个月全窗口盲扫。

    无论结果如何,产出的都是一个可复用的”长滞后独立增量”检验模板——而这正是当前运气医学研究中最稀缺的资产。

    证据与局限

    已验证: DLNM 能以交叉基同时建模非线性剂量-反应与滞后-反应,滞后窗口可拉到数十年(PMID 24027094);模拟研究证明长滞后关联下移动平均近似会系统性偏倚、DLNM 无偏(PMID 27400064);真实数据中存在”长滞后模型检出、单滞后模型漏检”的实例(瑞典五队列卒中 2-4 年与 8-9 年窗口,PMID 35270332)。

    部分验证: 传染病中的长滞后关联真实存在且有方法先例(荷兰轮状病毒滞后非线性建模,PMID 28333764);但多年尺度气候-疾病关联在部分地区/时期显著、另一些地区完全消失,提示其非平稳、可能脆弱。

    待验证: 在同一数据集上”短滞后 DLNM vs 短+长滞后 DLNM”的时间外增量检验(ΔAUC/互信息/Brier 分解 + 尊重自相关结构的置换检验)——长滞后气象信息是否携带独立于短滞后的可复现预测力,全球尚无直接实验。

    相关阅读

    证据等级: B
    参考文献: PMID 24027094, PMID 27400064, PMID 35270332, PMID 28333764, PMID 24681599, PMID 29460383


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Testing for Long-Lag Independence Increments in Length-Scale Lag Separation Models: Do Long-Lag Terms Provide Independent Information or Just Short-Term Weather Reskinning?

    Evidence level: B

    To become a testable scientific hypothesis, the “three-year epidemic transformation” must first pass through a narrow gate: after extending the lag window, any additional predictive power must be “independent”—not merely a reskinning of short lags, nor an illusion of statistical noise.

    Main Text

    The previous discussion addressed whether long-lag (1–3 years) meteorological-disease associations carry information independent of short lags. The conclusion was that this question “can be tested,” but has not yet been rigorously validated. This article decomposes that unverified gap into a specific, executable experimental design: fitting both “short-lag DLNM” and “short+long-lag DLNM” on the same dataset, comparing them via temporal external validation to determine if long-lag terms yield significant and reproducible increments.

    First, clarify what a “long-lag model” entails methodologically. The dominant framework for modern exposure-lag-response analysis is the Distributed Lag Nonlinear Model (DLNM), whose core component is the “cross-basis”: one dimension models the magnitude of exposure itself (nonlinear dose-response), while the other models lag time (lag-response), together spanning a two-dimensional risk surface. Gasparrini demonstrated in his foundational methodological paper that this framework can extend lag windows to decades—using radon exposure data from a uranium miner cohort, he estimated lung cancer mortality risk curves with lags ranging from 0 to 40 years (PMID 24027094). Thus, extending lags to multi-year scales is not a statistical difficulty; DLNM natively supports it.

    The challenge lies in the control design. Suppose a researcher does not construct a long-lag model but instead treats “average temperature over the past year” as a single variable input into the model—this is the most common shortcut. Gasparrini specifically used simulation studies to compare these two strategies: when a true long-lag association exists, moving-average-style rough approximations introduce systematic bias, whereas DLNM provides nearly unbiased estimates and confidence intervals close to nominal levels even under strong seasonal trends; moving averages are only acceptable alternatives when lags are indeed very short and interval settings are correct (PMID 27400064). This conclusion directly dictates model selection: long-lag terms must be modeled using spline-based lag-response functions and cannot be glossed over with simple averaging.

    In the real world, are there precedents where “long lags know more than short lags”? One clean example exists. A study of five Swedish cohorts used annual PM2.5 and black carbon (BC), setting exposure lags at 1–10 years, while contrasting against “annual single-lag” models and “1–5 year/6–10 year moving average” models. Results showed: for ischemic heart disease, DLNM found no lag-response; however, for stroke, DLNM indicated elevated risks at exposure windows of 2–4 years and 8–9 years, whereas the annual single-lag model showed no associations at all (PMID 35270332). This is a real-data instance of “long lags carrying information uncaptured by short-lag models”—its value lies not in the strength of the conclusion, but in proving that “using incorrect lag windows leads to systematic missed detections.”

    There are also usable instances on the infectious disease side. A lag nonlinear time series analysis of rotavirus in the Netherlands incorporated meteorological factors into a distributed lag model with seasonal terms and autoregressive moving average error correction for residuals, ultimately identifying the driving roles of susceptible population proportions and temperature on transmission (PMID 28333764). It suggests that long-lag modeling in infectious diseases must simultaneously handle seasonal terms and residual autocorrelation; otherwise, “long-lag significance” may merely reflect seasonal effects being counted twice.

    Next is the critical section: How to judge if the increment is “independent”? Two possibilities must be distinguished. First, long lags are merely a “reskinning” of short lags—macro-climate cycles are highly correlated with short-term weather, so long-lag terms merely recover information already encoded in short-term weather. Second, long lags genuinely carry additional information, such as cumulative changes in immunity, nutrition, or vector ecology over years. Statistical discipline for distinguishing these two involves three rules.

    First, select metrics that are “hard to deceive.” One widely circulated metric is Net Reclassification Improvement (NRI), but Pepe et al. used a simulation with 10,000 individuals, an event rate of 10.2%, and four completely non-predictive false biomarkers, finding NRI had alarmingly high false-positive rates: 63.0% in the training set and 18.8%–34.4% in the test set, whereas changes in AUC showed almost no false positives, and likelihood ratio tests remained stable at approximately 5% (PMID 24681599). Therefore, this experiment’s increment must be measured using ΔAUC / likelihood ratio / Brier decomposition, and NRI p-values must never be used to draw conclusions.

    Second, validation must be temporal external validation, not self-praise within the training set—models fitted in the training segment must be re-evaluated for increments in the held-out future segment to rule out overfitting.

    Third, permutation tests must respect temporal structure. Linden’s approach involves iteratively disguising each “unintervened” unit as “intervened” within interrupted time series to see if pseudo-treated units also estimate significant effects; if pseudo-units are also significant, the significance of true units cannot be attributed to the intervention (PMID 29460383). Applying this here means: permutation cannot simply shuffle time order (which would destroy autocorrelation and create inflated false positives); instead, “luck grouping labels” must be reshuffled while preserving autocorrelation and seasonal structures.

    Finally, interface with luck theory. The Heavenly Stems and Earthly Branches are fixed calendrical labels; grouping by them does not leak future information like many machine learning features, so the “long-lag independent increment validation” is methodologically clean. Meanwhile, the priors provided by luck theory itself (excessive-deficient years, guest spirits arriving) can serve as navigation for “which lag windows to search for signals”—this is precisely the comparison addressed in the next article: prior navigation vs. blind sweeping across the full 0–60 month window.

    Regardless of the outcome, the output will be a reusable “long-lag independent increment” validation template—precisely the most scarce asset in current luck-based medical research.

    Evidence and Limitations

    Verified: DLNM can simultaneously model nonlinear dose-response and lag-response via cross-basis, with lag windows extendable to decades (PMID 24027094); simulation studies prove moving-average approximations introduce systematic bias under long-lag associations while DLNM remains unbiased (PMID 27400064); real data instances exist where “long-lag models detect, single-lag models miss” (Swedish five-cohort stroke at 2–4 years and 8–9 year windows, PMID 35270332).

    Partially Verified: Long-lag associations in infectious diseases are real and have methodological precedents (Netherlands rotavirus lag nonlinear modeling, PMID 28333764); however, climate-disease associations at multi-year scales show significance in some regions/times but vanish entirely in others, suggesting non-stationarity and potential fragility.

    To Be Verified: Temporal external increment validation (“short-lag DLNM vs. short+long-lag DLNM”) on the same dataset (ΔAUC/mutual information/Brier decomposition + permutation tests respecting autocorrelation structure)—whether long-lag meteorological information carries reproducible predictive power independent of short lags, there is currently no direct global experiment.

    Related Reading

    Evidence Level: B
    References: PMID 24027094, PMID 27400064, PMID 35270332, PMID 28333764, PMID 24681599, PMID 29460383


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 数据越少,运气粗筛越有用?——’粗筛+DLNM’在气象序列缩短时的稳健性溢价检验

    Evidence level: B

    上一篇问”运气粗筛到底加不加分”,这一篇把数据量本身当成自变量——粗筛的价值可能不在于”更准”,而在于”更抗饿”:当气象序列缩短或缺失时,它与纯精算模型之间那条差距会不会反转。

    正文

    上一篇文章把”纯 DLNM、运气粗筛+DLNM、运气纯分层”三个模型摆在同一张桌上,用统一的时间外验证和 Brier 评分比高低。但它只回答了一个静态问题:数据齐全时谁更好。真实世界的预警现场却常常不是数据齐全的——很多基层哨点的气象序列只有短短几年,中间还断断续续缺测。于是更关键的问题冒了出来:如果数据变少,三个模型的排名会不会翻盘? 这才是”运气粗筛到底有没有用”的第二问。

    要问这个问题,得先把”性能”从单个数变成一条曲线。机器学习里有个成熟工具叫学习曲线(learning curve):横轴是训练数据量,纵轴是模型在新数据上的性能,把模型在 20、50、100、500 个样本上分别训练一遍,连起来就是它的”饿肚子曲线”。Zaloumis 团队用疟疾预测模型做了完整演示:训练集只有 20 个样本时,两个算法的平衡错误率都高达 50%(等于瞎猜),而训练集扩到 835 个样本时,错误率降到 14%–22%,曲线开始变平(PMID 40708012)。Figueroa 团队更早给出了数学形式——性能随样本量的增长大致服从”逆幂律”,可以用小样本的学习曲线外推到大样本(PMID 22336388)。这两条工作告诉我们:不同模型不是各自一条水平线,而是各自一条形状不同的上升曲线。

    形状不同,就意味着它们可能在某一点相交。复杂的纯 DLNM 要估计暴露-反应的非线性形状、滞后维度、季节项,参数多、胃口大——数据足时它爬到高位,数据少时它饿得最惨。运气粗筛+DLNM 则先用天干地支这把固定的历法尺子把年份分层,等于提前告诉模型”这些年属于同一类”,用先验结构压住了它要估的自由度。这正是临床预测建模里反复验证过的一条经验:当候选预测变量多而事件数少时,基于外部知识预先指定模型结构,比让数据自己挑变量表现更好——Steyerberg 团队在结直肠癌模型上做了模拟,常见但次优的逐步筛选法在六个原始预测变量里有五个出现了超过 50% 的偏倚,验证时判别度只有 0.74,而”按外部知识预先指定+连续变量”的做法升到 0.84(PMID 29174118)。运气粗筛做的,恰是”外部知识预先指定”这件事。

    但先验结构不是免费的午餐,这里必须泼一盆冷水。Van Calster 团队系统模拟了各类收缩(regularization)方法后发现:收缩在平均意义上确实改善校准,但在它最该发挥作用的小样本场景里,往往最不管用——不同数据集之间校准斜率的波动反而更大(PMID 32401702)。换句话说,”用先验压住自由度”能少输一点,却救不活真正匮乏的数据。更棘手的是第二道坎:数据少到一定程度,你连”谁更好”都测不出来。Beleites 团队指出,小样本下的分类器学习曲线会被同样小的测试集随机误差完全淹没,要证明一个分类器真正优于另一个,常常需要数百个独立测试样本,甚至理论上就做不到(PMID 23265730)。这意味着断点分析必须自带功效约束:如果两条曲线在某个数据量区间内的差距小于测试噪声,那段区间就不能宣称”谁反超了谁”。

    把这些线索拼起来,#383 的实验设计就清楚了:取一套完整的长序列(疾病日报+气象日报+干支历法表),按逐年截断和随机缺失两种方式制造数据匮乏梯度,在每个梯度上重新训练三个模型,用统一的滚动时间外验证算性能,画三条”性能-数据量”曲线,再做断点分析——找它们交于何处。还要盯着另一个方向的耗损:真实预警里”数据少”常常伴随”时间漂移”(新数据分布与训练期不一样),Davis 团队用 9 年纵向验证证明,判别度能守住、校准度却会随人群变化而漂移,而且漂移幅度与建模方法有关(PMID 28379439)。所以衰减曲线不能只画”样本量-性能”,最好同时画”时间-性能”。

    实验跑完,会得到三种可解释的结局,每一种都对应运气理论的一个版本。第一种:粗筛+DLNM 的曲线全程在纯 DLNM 之上——那运气粗筛就是普适增量,强主张。第二种:两条曲线在高数据量端纯 DLNM 略胜,但在低数据量端交叉反超——这就是”稳健性溢价”,也是最有意思的结局:粗筛的价值不是更准,而是在数据稀少时相对不那么差。第三种:粗筛曲线全程不高、斜率还更陡——那它就是装饰性的,甚至因为强行分层而损失了信息。真正值得下注的是第二种:它把”运气有没有用”从一个静态的是非题,改写成”在什么数据条件下有用”的条件命题。

    对中国用户,第二种结局有直接的现实意义。基层预警最难的不是算法,是数据——很多县域只有断续的几年疫情和气象记录。运气粗筛的成本几乎是零:一把固定的干支历法表,不依赖未来信息、不会泄露标签,天然适合做”低成本兜底结构”。如果它能证明自己”数据少时少输一点”,那它在资源有限地区的部署理由,就不是玄学信仰,而是可量化的稳健性溢价。

    证据与局限

    已验证: 性能-数据量的学习曲线是成熟方法学工具,可用逆幂律外推(PMID 40708012、22336388);事件数少时常见建模方法产生巨大偏倚,基于外部知识预先指定结构可显著改善(PMID 29174118);临床模型随时间出现校准漂移且漂移幅度与方法相关(PMID 28379439);小样本下分类器间优劣难以被统计检验分辨(PMID 23265730)。

    部分验证: 收缩/正则化(先验压自由度的类比)在平均意义上改善校准,但在最需要的小样本场景中稳定性反而更差,不构成补救(PMID 32401702)。

    待验证: 同一传染病数据集上,”运气粗筛+DLNM”相对”纯 DLNM”的性能-数据量曲线是否交叉反超,以及交叉断点落在哪个数据量区间——全球尚无此实验。

    相关阅读

    证据等级: B
    参考文献: PMID 40708012, PMID 22336388, PMID 29174118, PMID 32401702, PMID 23265730, PMID 28379439


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Is Less Data More Beneficial for Rough Screening? A Robustness Premium Test of ‘Rough Screening + DLNM’ in Shortened Meteorological Sequences

    Evidence level: B

    The previous article asked whether “luck rough screening” adds value. This piece treats data volume itself as an independent variable—the value of rough screening may not lie in “greater accuracy,” but in “greater resilience to hunger”: when meteorological sequences are shortened or missing, will the gap between it and pure actuarial models reverse?

    Main Text

    The previous article placed three models—”Pure DLNM,” “Luck Rough Screening + DLNM,” and “Luck Pure Stratification”—on the same table, comparing them using unified temporal external validation and Brier scores. However, it only answered a static question: which performs better when data is complete. Real-world warning sites often do not operate with complete data—many grassroots monitoring points have meteorological sequences spanning only a few years, with intermittent missing values. Thus, a more critical question emerged: If data decreases, will the ranking of the three models flip? This is the second question regarding whether “luck rough screening” is actually useful.

    To address this, we must transform “performance” from a single number into a curve. In machine learning, a mature tool is the learning curve: the horizontal axis represents training data volume, while the vertical axis represents model performance on new data. By training the model on 20, 50, 100, and 500 samples respectively and connecting the points, we obtain its “hunger curve.” Zaloumis et al. demonstrated this fully using a malaria prediction model: with only 20 samples in the training set, the balanced error rates for both algorithms reached 50% (equivalent to random guessing); when the training set expanded to 835 samples, error rates dropped to 14%–22%, and the curve began to flatten (PMID 40708012). The Figueroa team earlier provided the mathematical form—performance growth roughly follows an “inverse power law,” allowing small-sample learning curves to extrapolate to large samples (PMID 22336388). These studies tell us: Different models are not horizontal lines but rather curves of different shapes rising from low to high performance.

    Different shapes imply potential intersections. Complex Pure DLNM requires estimating the nonlinear shape of exposure-response, lag dimensions, and seasonal terms; it has many parameters and high computational demand—performing well when data is abundant but suffering most when data is scarce. Luck Rough Screening + DLNM first stratifies years using the fixed calendrical ruler of Heavenly Stems and Earthly Branches, effectively telling the model “these years belong to the same category,” thereby using prior structure to constrain the degrees of freedom it seeks to estimate. This aligns with a repeatedly verified experience in clinical prediction modeling: when candidate predictors are numerous but events are scarce, specifying model structure based on external knowledge performs better than letting data select variables—Steyerberg et al. simulated this in colorectal cancer models, where common but suboptimal stepwise selection introduced bias exceeding 50% in five of six original predictors, yielding a discrimination score of only 0.74 during validation; conversely, the approach of “specifying structure based on external knowledge + continuous variables” rose to 0.84 (PMID 29174118). What Luck Rough Screening does is precisely “specifying structure based on external knowledge.”

    However, prior structure is not a free lunch; here we must pour a bucket of cold water. Van Calster et al., through systematic simulations of various regularization methods, found that while regularization indeed improves calibration on average, it often fails most in the small-sample scenarios where it is most needed—the fluctuation of calibration slopes across different datasets actually increases (PMID 32401702). In other words, “using prior structure to constrain degrees of freedom” may reduce losses slightly but cannot revive truly scarce data. More棘手 (challenging) is the second hurdle: when data falls below a certain threshold, you cannot even measure “who is better.” Beleites et al. pointed out that under small-sample conditions, classifier learning curves are completely submerged by random errors from similarly small test sets; proving one classifier is truly superior often requires hundreds of independent test samples, which may be theoretically impossible (PMID 23265730). This means breakpoint analysis must carry power constraints: if the gap between two curves within a certain data volume range is smaller than test noise, that interval cannot claim “who overtook whom.”

    Piecing these clues together clarifies the experimental design for #383: take a complete long sequence (daily disease reports + daily meteorological reports + Heavenly Stems and Earthly Branches calendar), create data scarcity gradients via annual truncation and random missingness, retrain the three models at each gradient, calculate performance using unified rolling temporal external validation, plot three “performance-data volume” curves, and conduct breakpoint analysis—to find where they intersect. Also monitor loss in the other direction: in real warnings, “less data” often accompanies “temporal drift” (new data distribution differs from the training period). Davis et al. proved using a 9-year longitudinal validation that while discrimination can be maintained, calibration drifts with population changes, and the magnitude of drift depends on the modeling method (PMID 28379439). Therefore, decay curves should not only plot “sample size-performance” but ideally also include “time-performance.”

    Upon completion of the experiment, three interpretable outcomes will emerge, each corresponding to a version of the luck theory. The first: The curve for Rough Screening + DLNM remains entirely above Pure DLNM—that would make luck rough screening a universal incremental gain, strongly advocating for it. The second: Pure DLNM is slightly superior at high data volumes, but the curves cross and reverse at low data volumes—that is the “robustness premium,” the most interesting outcome: the value of rough screening is not greater accuracy, but being relatively less poor when data is scarce. The third: The rough screening curve remains low throughout with a steeper slope—that would make it decorative, even losing information due to forced stratification. The second outcome is truly worth betting on; it rewrites the static yes/no question of “whether luck is useful” into a conditional proposition: “under what data conditions is it useful.”

    For Chinese users, the second outcome has direct practical significance. The greatest challenge for grassroots warnings is not algorithms but data—many counties have only intermittent years of epidemic and meteorological records. The cost of luck rough screening is nearly zero: a fixed Heavenly Stems and Earthly Branchs calendar, independent of future information and non-leaking labels, naturally suits a “low-cost fallback structure.” If it proves itself to “lose slightly less when data is scarce,” its deployment rationale in resource-limited regions shifts from esoteric belief to quantifiable robustness premium.

    Evidence and Limitations

    Verified: Learning curves regarding performance vs. data volume are mature methodological tools, extrapolatable via inverse power laws (PMID 40708012, 22336388); common modeling methods produce significant bias when event counts are low, but specifying structure based on external knowledge significantly mitigates this (PMID 29174118); clinical models exhibit calibration drift over time with drift magnitude dependent on the method (PMID 28379439); distinguishing classifier superiority under small samples is difficult due to statistical masking (PMID 23265730).

    Partially Verified: Regularization (analogous to constraining degrees of freedom via priors) improves calibration on average but exhibits reduced stability in the most critical small-sample scenarios, failing to serve as a remedy (PMID 32401702).

    To Be Verified: Whether the performance-data volume curve for “Luck Rough Screening + DLNM” crosses and surpasses that of “Pure DLNM” on the same infectious disease dataset, and at which data volume interval the crossover occurs—no such experiment exists globally yet.

    Related Reading

    Evidence Level: B
    References: PMID 40708012, PMID 22336388, PMID 29174118, PMID 32401702, PMID 23265730, PMID 28379439


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 运气粗筛是改了概率还是改了排序?——三模型头对头的 Brier 分解裁决

    Evidence level: B

    两个模型拿到几乎一样的 Brier 评分,未必是同一回事——把综合准确度拆成”校准”和”判别”两块之后,运气粗筛的价值到底落在哪一块,才第一次说得清。

    正文

    上一篇提出了一个三模型头对头实验,用来裁决”运气粗筛+DLNM”是否真比”纯DLNM”预测得准。但只比一个 AUC 或者一个总分,会漏掉一个要命的问题:同样的总分,可能是两件完全不同的事。

    Brier 评分(预测概率与真实结果的均方误差)是传染病与环境预测里最常用的综合准确度指标,问题在于它是个”复合分数”——它同时装着两样东西:模型能不能把高风险和低风险排对顺序(判别),以及模型给出的概率数字本身准不准(校准)。Rufibach 用一个很小的例子说明过:在有些情况下,一个校准更差的模型反而拿到更低的 Brier 分,因为它的排序能力更强(PMID 20189763)。换句话说,Brier 低不等于校准好,Brier 高也不等于没本事。

    这就把”运气粗筛到底有没有用”从一个问题变成了两个问题。

    第一,运气粗筛是不是只改了概率水平? 干支是固定的历法标签,如果它只是让模型”整体上把某些年份的概率抬高一点”,那会改变 reliability(校准)那一项,却不会改变模型对”哪个病人、哪一年风险更高”的排序——resolution(判别)项不变。这种价值是装饰性的,极端情况下还有害:把风险概率系统性说高或说低,会误导预警阈值。

    第二,运气粗筛是不是真的带来了新的排序信息? 如果”三年化疫”的长滞后(1-3 年)真的携带独立于短期气象(0-3 周)的信息,那么它应该在 resolution 项上留下痕迹——因为长滞后改变的是”哪些年份算高风险年份”这个顺序判断。这才是”真本事”。

    所以正确的比较协议不是”谁的 Brier 低”,而是把 Brier 拆成三块:reliability(校准,越小越好)、resolution(判别贡献,越大越好)、uncertainty(由结局基准率决定的、与模型无关的常数项)。只有拆开,才知道差异落在哪一块。近年方法学也在往这个方向走:Yang 等把 Brier 拆成”概率预测误差”与”结局自身方差”两部分,剔掉与模型无关的方差后比较更灵敏(PMID 36031854);Zhu 等进一步把加权 Brier 显式分解为判别成分与校准成分,并把两者与临床效用挂钩(PMID 41103457)。

    光有分解还不够,还得配三样纪律。

    一是配对检验,不是独立检验。 三个模型是在同一批病例、同一批年份上评估的,它们的 AUC 估计量互相相关。DeLong 1988 年的非参数方法用广义 U 统计量给出协方差矩阵,才是比较相关 ROC 曲线下面积的标准做法(PMID 3203132)。若把两个高度相关的 AUC 当成独立样本丢进两样本 z 检验,会因为忽略正协方差而把差异算得过于保守,甚至给出错误结论。

    二是校准要看图,也要看多个数。 校准不能只报一个 Hosmer-Lemeshow 的 p 值。Huang 等的教程明确指出,判别与校准在模型评估里承担不同角色,校准应同时用校准截距、校准斜率、Spiegelhalter z 等多个统计量再加可靠性图(校准曲线)来判断——因为单一摘要统计量会掩盖”某些风险区间系统性高估、另一些区间系统性低估”这类结构性错误(PMID 32106284)。

    三是预先登记比较清单。 三模型 × 多滞后窗 × 多领先时间 × 多指标,比得越多越容易撞假阳性,必须事先登记要比的模型和多重校正方案,避免”事后挑最好看的结果”。

    把这三条合起来,这个实验能给出三种明确的结果解读,每一种都直接对应运气理论的一个版本:

    • 若粗筛的 Brier 优势主要来自 reliability 改善、而 DeLong 检验显示 AUC 无显著判别提升 → 运气改的是”概率水平”,对分级预警有用,但不构成”运气携带新信息”的证据;
    • 若主要来自 resolution 改善、且伴随 AUC 显著提升 → 运气提供了独立于气象的排序信息,这是运气粗筛”真本事”的最强证据;
    • 若两块都无显著改善 → 粗筛是装饰性的,部署需谨慎。

    对五运六气研究而言,这个分解实验的额外价值在于:它把”运气有没有用”这个容易吵架的定性问题,变成一份可核对的分账表——运气粗筛的每一分增量,到底是记在”校准”账上还是”判别”账上,白纸黑字写清楚。 这比一句”提高了预测精度”要诚实得多,也更容易被后续研究检验。

    最后提醒一点:Brier 分低也不自动等于”更值得用”。Zhu 等把临床效用引入后指出,模型的价值最终取决于它被用在哪个决策阈值上(PMID 41103457)。对一个只在”极高风险才预警”的阈值下使用的系统,判别好、校准略差的模型可能比反过来更有用。所以三模型头对头实验的终点,应该是三份报告:判别报告、校准报告、效用报告——而不是一个总分排名。

    证据与局限

    已验证: Brier 评分是可分解的复合指标,其校准与判别成分能被分别估计(PMID 20189763、36031854、41103457);比较同一批样本上多个模型 AUC 的配对非参数方法是成熟的(DeLong,PMID 3203132);校准评估应结合多个统计量与可靠性图,不能只靠单一拟合检验(PMID 32106284)。

    部分验证: “运气粗筛+DLNM”在北京结核病回顾中的分层思路初步可行(来源文章已引 PMID 26525546),但样本单中心、指标仅相关系数/准确率,从未做同数据集的 Brier 分解。

    待验证: 在同一时间外验证协议下,对”纯DLNM vs 运气粗筛+DLNM vs 运气纯分层”三模型做 Brier 三分量分解(reliability/resolution/uncertainty)+ DeLong 配对 AUC 检验 + 校准曲线,判定运气粗筛的增量落在校准还是判别——全球尚无此实验。

    相关阅读

    证据等级: B
    参考文献: PMID 3203132, PMID 20189763, PMID 32106284, PMID 36031854, PMID 41103457


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Did Luck Coarse Filtering Change Probabilities or Sorting? A Head-to-Head Brier Decomposition Decision Among Three Models

    Evidence level: B

    Two models may yield nearly identical Brier scores without representing the same underlying mechanism—after decomposing comprehensive accuracy into “calibration” and “discrimination,” it is only now possible to clarify where the value of rough screening of luck actually resides.

    Main Text

    The previous article proposed a head-to-head experiment involving three models to adjudicate whether “rough screening of luck + DLNM” truly outperforms “pure DLNM” in prediction accuracy. However, comparing solely on a single AUC or a total score overlooks a critical issue: identical total scores may represent two fundamentally different outcomes.

    The Brier score (mean squared error between predicted probabilities and actual outcomes) is the most widely used composite accuracy metric in infectious disease and environmental prediction. Its limitation lies in being a “composite score”—it simultaneously encapsulates two distinct aspects: whether the model correctly ranks high-risk versus low-risk cases (discrimination) and whether the probability values it outputs are accurate (calibration). Rufibach illustrated this with a small example: in certain scenarios, a model with poorer calibration can achieve a lower Brier score because its sorting capability is stronger (PMID 20189763). In other words, a low Brier score does not equate to good calibration, nor does a high score imply lack of capability.

    This reframes the question of “whether rough screening of luck is useful” from a single inquiry into two distinct ones.

    First, does rough screening of luck merely adjust probability levels? The Heavenly Stems and Earthly Branches serve as fixed calendrical labels. If their function is only to shift the model’s overall probability upward for certain years, they would alter the reliability (calibration) component while leaving the resolution (discrimination) component unchanged, as the model’s ranking of “which patient/year carries higher risk” remains consistent. Such value is decorative; in extreme cases, it can be harmful: systematically overestimating or underestimating risk probabilities may mislead warning thresholds.

    Second, does rough screening of luck genuinely introduce new sorting information? If the long lag associated with “three-year epidemic cycles” (1–3 years) truly carries information independent of short-term meteorological data (0–3 weeks), it should leave a trace in the resolution component—because long lags alter the sequential judgment of “which years constitute high-risk years.” This constitutes the model’s “true capability.”

    Therefore, the correct comparison protocol is not “who has the lower Brier score,” but rather decomposing the Brier score into three parts: reliability (calibration, smaller is better), resolution (discrimination contribution, larger is better), and uncertainty (a constant term determined by the baseline rate of the outcome, independent of the model). Only through decomposition can we identify where the differences lie. Recent methodological advances are moving in this direction: Yang et al. decomposed the Brier score into “probability prediction error” and “outcome variance,” enhancing sensitivity by comparing only after removing variance unrelated to the model (PMID 36031854); Zhu et al. further explicitly decomposed the weighted Brier score into discrimination and calibration components, linking both to clinical utility (PMID 41103457).

    Decomposition alone is insufficient; three additional disciplines must accompany it.

    First, use paired testing rather than independent testing. The three models were evaluated on the same cohort of cases and years; thus, their AUC estimators are correlated. The non-parametric method proposed by DeLong in 1988, utilizing generalized U-statistics to provide the covariance matrix, remains the standard approach for comparing correlated ROC curve areas (PMID 3203132). Treating two highly correlated AUCs as independent samples in a two-sample z-test would ignore positive covariance, rendering the difference estimate overly conservative and potentially leading to erroneous conclusions.

    Second, assess calibration through visual inspection and multiple statistics. Calibration should not rely solely on a Hosmer-Lemeshow p-value. Huang et al.’s tutorial explicitly clarifies that discrimination and calibration serve distinct roles in model evaluation; calibration should be judged using multiple statistics—including calibration intercept, calibration slope, and Spiegelhalter z-values—alongside reliability plots (calibration curves)—because a single summary statistic may obscure structural errors such as systematic overestimation in certain risk intervals and underestimation in others (PMID 32106284).

    Third, pre-register comparison lists. With three models × multiple lag windows × multiple lead times × multiple metrics, an excessive number of comparisons increases the risk of false positives. It is essential to pre-register the models to be compared and multiple correction strategies to avoid “selecting the most visually appealing results post-hoc.”

    Synthesizing these three points, this experiment can yield three clear interpretations of results, each directly corresponding to a version of the luck theory:

    • If the Brier advantage from rough screening stems primarily from reliability improvement, while the DeLong test shows no significant discrimination gain in AUC → luck modifies “probability levels,” useful for tiered warnings but failing to constitute evidence that “luck carries new information”;
    • If it stems primarily from resolution improvement, accompanied by a significant AUC increase → luck provides sorting information independent of meteorology, representing the strongest evidence for the “true capability” of rough screening of luck;
    • If neither component shows significant improvement → rough screening is decorative, and deployment should proceed with caution.

    For research on Five Movements and Six Qi, the additional value of this decomposition experiment lies in transforming the easily contentious qualitative question of “whether luck is useful” into a verifiable ledger: clearly documenting whether every incremental point of value from rough screening is recorded under “calibration” or “discrimination.” This is far more honest than simply stating “improved prediction accuracy” and is easier for subsequent studies to validate.

    Finally, a reminder: a low Brier score does not automatically equate to “greater utility.” Zhu et al. noted after introducing clinical utility that a model’s value ultimately depends on the decision threshold at which it is applied (PMID 41103457). For a system used only under a threshold warning for “extreme high risk,” a model with strong discrimination but slightly poorer calibration may be more useful than the reverse. Thus, the endpoint of the head-to-head experiment among three models should be three reports: a discrimination report, a calibration report, and a utility report—not a single total-score ranking.

    Evidence and Limitations

    Verified: The Brier score is a decomposable composite metric, with calibration and discrimination components estimable separately (PMID 20189763, 36031854, 41103457); paired non-parametric methods for comparing AUCs across multiple models on the same sample are established (DeLong, PMID 3203132); calibration assessment should combine multiple statistics with reliability plots rather than relying on a single fit test (PMID 32106284).

    Partially Verified: The stratification approach of “rough screening of luck + DLNM” in the Beijing tuberculosis retrospective study is preliminarily feasible (source article cites PMID 26525546), but the sample was single-center, metrics were limited to correlation coefficients/accuracy, and no Brier decomposition on the same dataset was performed.

    To Be Verified: Under a separate-time validation protocol, performing Brier three-component decomposition (reliability/resolution/uncertainty) + DeLong paired AUC testing + calibration curves for the three models (“pure DLNM” vs. “rough screening of luck + DLNM” vs. “luck pure stratification”) to determine whether the incremental value of rough screening falls under calibration or discrimination—no such experiment exists globally.

    Related Reading

    Evidence Level: B
    References: PMID 3203132, PMID 20189763, PMID 32106284, PMID 36031854, PMID 41103457


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.