Evidence level: C
如果「哪种干预动哪几阶谐波」的结论只在健康年轻人里成立,它就不是脉象谐波的规律,只是那批受试者的画像。
正文
上一篇《四类干预的谐波「频谱指纹」到底长在哪几阶》用逐阶归因回答了「指纹长在哪」;本篇处理它的下一个致命问题:换一群人,指纹还长在原来的地方吗。这叫跨人群归因稳定性,是「干预指纹」这个概念能不能当规律用的生死线。
为什么必须问。因为现有的谐波证据几乎全部来自健康或单一人群队列:30 名健康受试者在五类日常活动下的桡动脉谐波状态信号(PMID 33082826),训练出的状态分类边界,换到患者或老年人身上未必成立。而且人群本身就在改写谐波底座——血管老化指数与年龄的相关系数达 0.56 到 0.80(PMID 28261531),老年人 fingerprints 的「纸」已经换了。
跨人群差异的实际证据已经在积累。把健康老年人与糖尿病患者放在一起对照,容量脉搏的 C0 到 C10 谐波系数出现系统性分离,C0 联合腰围对糖尿病风险有梯度关联(PMID 34945735,141 人)——同一套阶次,在两个队列里承担的判别角色不同。阿尔茨海默病患者与对照的 40 项谐波指数做多层感知机分类,准确率 82.86%、AUC 0.83(PMID 33903610),但该研究没有报告「哪些阶贡献最大」是否随人群改变。更直接的年龄分层证据来自血管老化判别研究:280 名 40 到 80 岁受试者用 40 项频域指数训练,40 到 50 岁段差异最突出,50 岁以上各段判别准确率降到 80% 上下、AUC 跌到 0.73 以上(PMID 34508787)——人群段一换,模型表现整个下台阶,归因位置更没人验过。患者侧的锚点是脉搏谐波预测冠脉病变严重度的 SYNTAX 评分研究(PMID 40396324),证明谐波在患者队列里确有判别力,但它的 Top-K 阶次与健康队列指纹是否重叠,同样未验。
裁决设计分三层。第一层是同协议双队列:健康队列与患者或老年队列用完全相同的干预协议、采集流程与特征管道,杜绝「差异来自仪器或方案」的低级混淆。第二层是逐阶归因对齐:两个队列各自训练同一架构的分类器,用 SHAP 与消融法做逐阶归因,取各队列的 Top-K 阶次,算 Jaccard 重叠率与 Kendall W 一致性系数——重叠高说明「哪阶对应哪种干预」近似普适;重叠低则说明指纹是样本特异的。第三层是归因差异的生理锚定:两队列归因向量之差,与血管状态指标(脉搏波传导速度 PWV、增强指数 AIx)做关联检验——如果归因漂移能被血管状态解释,那漂移就不是噪声,而是「底座变了」的可预测函数。
实施与验收口径需要写死四个细节,否则裁决会被统计口径偷换掉。第一,Top-K 的 K 必须在开跑前固定(建议 K 取 3 与 5 各跑一遍做敏感性),不许看完数据再挑 K——否则「重叠率」随时可以事后调高。第二,Jaccard 与 Kendall W 的随机基线要先算清楚:10 个候选阶次里随机抽 K 个的期望重叠率不是零,判「普适」必须显著高于该基线,判「样本特异」必须落在基线噪声带内。第三,两队列样本量不等时,归因向量的抽样方差不同,重叠率要做 bootstrap 置信区间,不许拿点估计直接下结论。第四,SHAP 与消融两种归因法要同时跑:两者给出的 Top-K 若在队列内部就互相打架,说明该模型的归因本身不稳定,跨队列比较先按下,回头解决模型可解释性问题——这也是「归因稳定性裁决」的前置门禁,归因在队列内都不稳,谈跨队列没有意义。
这套设计有一个潜在的反向收益:如果跨队列归因漂移确实能被 PWV 与 AIx 解释,那么「干预指纹」就从二值判断升级为连续刻度——同一干预在低血管年龄的人身上动第几阶、在高血管年龄的人身上动第几阶,可以做成一张「人群×阶次」的归因迁移表。那张表如果成立,对脉诊的现代化解释价值比单点指纹大得多:它等于把「同病异脉、异病同脉」的中医观察翻译成可计算的频域语言。但在跨队列对齐做完之前,这一切都只是设计意图,不是证据。
证据与局限
已验证: 谐波阶次在健康与患者队列间存在系统性数值差异(PMID 34945735、33903610);判别力随人群段衰减(PMID 34508787);血管状态随年龄系统改变谐波底座(PMID 28261531)。
部分验证: 患者队列中谐波对病变严重度有判别力(PMID 40396324);干预状态下健康人谐波携带状态信号(PMID 33082826)。但「干预归因的 Top-K 阶次」从未在两个队列间做过正式对齐。
待验证: 逐阶 SHAP 归因的跨队列 Jaccard 与 Kendall W 重叠率;归因漂移量与 PWV、AIx 的定量关联;「普适指纹」与「画像指纹」的分界阈值。这是本篇的三条证伪条件——若重叠率接近随机基线且归因漂移与血管状态无关,「普适规律」一说不成立。
L1 网络状态:该主题尚未进入 L1 对应网络——EDGE_TABLE_science-yixue.md 中脉象谐波相关边仅有 1d-model-pulse-tcm-classification-blind-comp(频率/心率/共振/谐波四行,代表文献 PMID 21622820)与 pulse-harmonic-organ-function-scoring-system(六腑,B 级)两条,均无「跨人群归因稳定性、SHAP、Top-K」维度;CORRESPONDENCE_NETWORK.md 中最近的条目为 E-060 的「人群×频段方向须分层预写」(KB-341/KB-351),与本主题相邻但不同边。故本文证据等级封顶 C。
相关阅读
证据等级: C
参考文献: PMID 40396324, PMID 34945735, PMID 33903610, PMID 34508787, PMID 33082826, PMID 28261531, PMID 32957479
证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索
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。
Cross-population stability verdicts on intervention fingerprint order attribution: Do Top-K orders found in healthy individuals still hold for patients and the elderly?
Evidence level: C
If the conclusion regarding “which intervention modulates which harmonic orders” holds true only for healthy young individuals, it does not constitute a regularity of pulse waveform harmonics but rather a profile specific to that cohort of subjects.
Body
The previous article, “Where do the harmonic ‘spectral fingerprints’ of four types of interventions actually lie?” (Part 1), answered “where they lie” via step-by-step attribution. This article addresses its next critical issue: Does the fingerprint remain in the same place when applied to a different group of people? This is cross-population attribution stability, the life-or-death threshold for whether the concept of “intervention fingerprint” can serve as a generalizable law.
Why is this question necessary? Because existing harmonic evidence almost entirely derives from healthy or single-population cohorts: the radial artery harmonic state signals of 30 healthy subjects under five daily activities (PMID 33082826); the trained state-classification boundaries may not hold for patients or the elderly. Furthermore, the population itself rewrites the harmonic foundation—the correlation coefficient between vascular aging index and age ranges from 0.56 to 0.80 (PMID 28261531); the “paper” of elderly fingerprints has changed.
Evidence of actual cross-population differences is accumulating. Comparing healthy elderly with diabetic patients reveals systematic separation in the C0 to C10 harmonic coefficients of pulse volume; C0 combined with waist circumference shows graded association with diabetes risk (PMID 34945735, 141 subjects)—the same set of orders assumes different discriminative roles in the two cohorts. A multi-layer perceptron classification of 40 harmonic indices between Alzheimer’s patients and controls achieved 82.86% accuracy and AUC 0.83 (PMID 33903610), yet the study did not report whether “which orders contribute most” shifts with population. More direct age-stratified evidence comes from vascular aging discrimination studies: 280 subjects aged 40–80 were trained on 40 frequency-domain indices; the 40–50 age band showed the most pronounced difference, while discrimination accuracy for ages >50 dropped to around 80%, with AUC falling below 0.73 (PMID 34508787)—once the population segment changes, model performance drops a step, and attribution locations remain unverified. On the patient side, the anchor is the SYNTAX score study predicting coronary lesion severity via pulse harmonic prediction (PMID 40396324), proving discriminative power in patient cohorts, but whether its Top-K orders overlap with the healthy cohort’s fingerprint remains untested.
The ruling design comprises three layers. The first layer is dual cohorts under the same protocol: healthy and patient/elderly cohorts undergo identical intervention protocols, acquisition procedures, and feature pipelines to eliminate low-level confounding where “differences stem from instruments or schemes.” The second layer is step-by-step attribution alignment: each cohort trains classifiers of the same architecture; SHAP and ablation methods perform step-by-step attribution, extracting Top-K orders for each cohort to compute Jaccard overlap rate and Kendall W consistency coefficient—high overlap suggests “which order corresponds to which intervention” is nearly universal; low overlap indicates the fingerprint is sample-specific. The third layer is physiological anchoring of attribution differences: the difference between attribution vectors of the two cohorts is tested against vascular status indicators (pulse wave velocity PWV, augmentation index AIx); if attribution drift can be explained by vascular status, the drift is not noise but a predictable function of “the foundation changing.”
Implementation and acceptance criteria must fix four details, otherwise the ruling risks statistical manipulation. First, K in Top-K must be fixed before running (suggest running K=3 and K=5 separately for sensitivity); selecting K post-hoc after seeing data is forbidden, as “overlap rate” can always be artificially inflated afterward. Second, the random baseline for Jaccard and Kendall W must be calculated precisely: the expected overlap rate of randomly selecting K out of 10 candidate orders is not zero; judging “universality” requires significant elevation above this baseline, while judging “sample specificity” requires falling within the baseline noise band. Third, when cohort sample sizes differ, attribution vector sampling variances differ; overlap rates must use bootstrap confidence intervals rather than direct point estimates. Fourth, both SHAP and ablation attribution methods must run simultaneously: if the Top-K they produce conflict within a cohort, the model’s attribution itself is unstable, so cross-cohort comparison is paused until model interpretability issues are resolved—this is the prerequisite gate for “attribution stability ruling”; if attribution is unstable within cohorts, cross-cohort comparison is meaningless.
This design has a potential reverse benefit: if cross-cohort attribution drift can indeed be explained by PWV and AIx, then “intervention fingerprint” upgrades from binary judgment to continuous scale—identifying which order an intervention affects in low-vascular-age individuals versus high-vascular-age individuals can form a “population × order” attribution transfer table. If this table holds, its value for modernizing pulse diagnosis far exceeds single-point fingerprints: it translates the traditional Chinese observation of “same disease, different pulse; different disease, same pulse” into computable frequency-domain language. However, until cross-cohort alignment is completed, all of this remains design intent, not evidence.
Evidence and Limitations
Verified: Systematic numerical differences in harmonic orders exist between health status and patient cohorts (PMID 34945735, PMID 33903610); discriminative power attenuates across population segments (PMID 34508787); vascular status systematically alters the harmonic baseline with age (PMID 28261531).
Partially Verified: Harmonic orders show discriminative power for lesion severity within patient cohorts (PMID 40396324); healthy subjects carry status signals in harmonics under intervention conditions (PMID 33082826). However, the “Top-K orders attributed to intervention” have never undergone formal alignment between the two cohorts.
Pending Verification: Cross-cohort Jaccard and Kendall W overlap rates for stepwise SHAP attribution; quantitative association between attribution drift and PWV, AIx; the threshold separating “universal fingerprints” from “portrait fingerprints.” These constitute the three falsification conditions for this paper—if overlap rates approach random baseline and attribution drift is uncorrelated with vascular status, the claim of “universal regularities” does not hold.
L1 Network Status: This topic has not yet entered the L1 corresponding network. In EDGE_TABLE_science-yixue.md, the only edges related to pulse-harmonic features are
1d-model-pulse-tcm-classification-blind-comp(covering frequency, heart rate, resonance, and harmonics; four rows representing cited literature PMID 21622820) andpulse-harmonic-organ-function-scoring-system(covering the six fu-organs, Level B). Neither includes dimensions for “cross-population attribution stability,” “SHAP,” or “Top-K.” The most recent entry in CORRESPONDENCE_NETWORK.md is E-060, titled “Population × Band direction must be pre-written stratified” (KB-341/KB-351), which is adjacent but represents a different edge. Therefore, the evidence level for this paper is capped at C.
Related Reading
- Harmonic order attribution of the four types of interventions: where do their spectral fingerprints lie?
- Division of labor between low-order identity and high-order states
- Amplitude vs. phase channel arbitration for intervention fingerprints
- Research pool: bio-frequency code
Evidence level: C
References: PMID 40396324, PMID 34945735, PMID 33903610, PMID 34508787, PMID 33082826, PMID 28261531, PMID 32957479
Evidence level: A = direct PubMed support; B = scientific analogy or indirect evidence; C = traditional consensus or preliminary exploration.
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.
Leave a Reply