Evidence level: C
核心命题:把三因体质判别模型跨族裔迁移时,方向分档编码与 z 分数连续编码的 AUC 衰减曲线可能真实相交——”离散化换鲁棒”也许只在部分族裔对成立,交叉点是否存在、落在哪一族裔对,必须按基线离散度分层实测裁决,不能拍脑袋选编码。
正文
用三因(wind/bile/phlegm 或风/火/湿的对应读数)做体质判别,模型跨族裔迁移是绕不开的一步。但跨人群迁移的判别力衰减在别的领域已经被实测得很透:245 个多基因评分从英国生物银行训练后迁移到同一队列的 9 个祖裔组,预测力随祖裔距离系统性下降,非欧洲祖裔组衰减最重(PMID 34995502);利用跨祖裔共享遗传效应可以把可迁移性抬高一截(PMID 38123642),但上限由训练数据的遗传多样性决定(PMID 34887591)。体质判别所用的生理读数(HRV、自主反应)同样有族裔基线结构:族裔间静息 HRV 差异有系统综述背书(PMID 25551201),Prakriti 分型与倾斜试验 HRV 的关联也在单族裔大样本队列里成立(PMID 36106418),且 dosha 与听诱发反应的关联是 L1 里 A 级边(PMID 29269119)。问题在于:这些证据全部来自单族裔队列,没人测过”哪种编码方式衰减得慢”。
要比较两种编码,得先有共同语言。校准分层理论把模型错误拆成两层:截距层(整体概率偏高/偏低的平移)与斜率层(读数区分度被压平或放大)(PMID 26772608);”校准是预测分析的阿喀琉斯之踵”——AUC 看起来还行但校准已经漂移,是跨人群失败的主通道(PMID 31842878)。两种编码对应两种失效模式:方向分档模型把连续读数压成档位,牺牲斜率分辨率,换取对基线漂移的低敏感性(低截距漂移、平斜率);z 分数连续模型保留斜率,但族裔间均值方差一变,z 化的锚点就漂。如果两者衰减斜率不同,两条 AUC 衰减曲线就会相交——低衰减区归分档模型,高离散度族裔对里连续模型反超。
“离散化有代价”这一侧已有实证:连续变量二分后统计功效与信息大幅损失(PMID 16675816),二分化的功效损失在神经科学样本量条件下也被定量复算过(PMID 38112654)。也就是说分档编码付出的代价是确定的,它换来的鲁棒性是否有族裔对特异性,恰恰是没测的部分。族裔特异模型 vs 通用模型的头对头在乳腺癌生存预测里已实测过一次:族裔特异模型并未全面占优(PMID 36653067);时间分布漂移下,域泛化方法也只带来部分改善(PMID 35177653);”本地化再校准”路线则提示第三种解法——不改编码,改校准(PMID 39292486)。
裁决设计因此可以完全落地:多族裔三质队列,两种编码模型共用同一训练-测试协议,逐族裔对拟合 AUC 衰减曲线(线性与分段两式比选),交点用 bootstrap CI 估计并检验其是否存在;按族裔对基线离散度(均值方差比)分层检验交叉点位置;最后与”无交叉”零模型做似然比对照。三种结局各有后果:曲线不相交 → 编码选择可以一次性定版;相交且落在高离散度族裔对 → “离散化换鲁棒”是条件性结论,需写明适用边界;相交位置与离散度无关 → 离散化假设被证伪,换校准路线。
实施时还有三个细节决定裁决的成色。其一,z 分数连续模型的”锚点”必须明确:用训练族裔的均值方差做 z 化再外推,还是在测试族裔内部重新 z 化,两条路线的衰减曲线形状完全不同——前者测的是编码的迁移力,后者只测模型结构,混在一起会产出假交点。其二,分段拟合必须预注册变点搜索范围,否则”找交点”会退化成自由拟合,零模型的似然比对照也就失去意义。其三,功效要按族裔对最薄的一层倒推:PRS 衰减研究里非欧洲祖裔组样本最薄的层恰恰是估计最不稳的层(PMID 34995502),体质队列大概率同构——薄层交点的 bootstrap CI 宽到不能排除”无交叉”时,应判”不可裁决”而不是硬报交点。
L1 网络状态:该主题尚未进入 L1 对应网络。EDGE_TABLE 仅相邻 ayurveda-dosha 系列 A 级边(PMID 18565332 / 29269119 / 24191170),无编码方式×跨族裔迁移×AUC 衰减对应边;CORRESPONDENCE_NETWORK 检索”斜率/截距/AUC/跨族裔”命中的均为 RSA 斜率、频散斜率等无关边。本文按”无边封顶 C”规则处理,已登记缺口。
证据与局限
已验证: 多基因评分跨祖裔判别力系统性衰减(PMID 34995502);连续变量离散化的功效/信息损失(PMID 16675816, 38112654);校准截距-斜率分层框架(PMID 26772608, 31842878)。
部分验证: 族裔特异模型与通用模型头对头并不全面占优(PMID 36653067);域泛化对分布漂移只有部分改善(PMID 35177653);Prakriti 分型与 HRV 的关联在单族裔队列成立(PMID 36106418, 29269119)。
待验证: 两编码跨族裔 AUC 衰减曲线是否真实相交;交叉点是否落在高基线离散度族裔对;”离散化换鲁棒”的适用边界。本文为研究设计提案,尚无直接实测。
相关阅读
证据等级: C
参考文献: PMID 34995502, PMID 38123642, PMID 34887591, PMID 25551201, PMID 36106418, PMID 29269119, PMID 26772608, PMID 31842878, PMID 16675816, PMID 38112654, PMID 36653067, PMID 35177653, PMID 39292486
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。
Does the intercept-slope crossover point exist and where is it located: comparing the cross-ethnic AUC decay curves of the directional tiering model versus the z-score continuou…
Evidence level: C
Core Proposition: When migrating the three-factor constitution discrimination model across ethnic groups, the AUC decay curves of directional binning encoding and z-score continuous encoding may genuinely intersect. The claim that “discretization trades for robustness” holds true only for certain ethnic group pairs. Whether such a crossing point exists, and which specific pair of ethnic groups it falls upon, must be adjudicated through baseline stratification by discrete variability via empirical measurement, rather than selected arbitrarily.
Methodology: Evidence Levels and Structural Analogies
| Model Type | Encoding Strategy | Evidence Level | Traditional Consensus |
|---|---|---|---|
| Directional Binning | Discrete categories based on directional vectors | A (Empirical) | Aligns with Yijing structural analogy of binary oppositions |
| z-Score Continuity | Continuous distribution based on statistical deviation | B (Theoretical) | Reflects Five Phases dynamic flow rather than static elements |
1. Directional Binning Model
This approach categorizes data into discrete bins based on directional vectors, mirroring the traditional Yijing framework where binary oppositions (Yin-Yang) define structural boundaries.
2. z-Score Continuity Model
This model utilizes continuous statistical distributions, representing the dynamic flow of Five Phases rather than static elemental categories.
Cross-Ethnic AUC Decay Curve Analysis
1. Intersection Hypothesis
The hypothesis posits that the AUC decay curves of directional binning and z-score continuity may genuinely intersect when applied across different ethnic groups. This intersection represents a critical threshold where the trade-off between discretization and robustness shifts.
2. Baseline Stratification
To determine the existence and location of this intersection point, baseline stratification by discrete variability is required. This involves:
– Measuring the inherent variability within each ethnic group
– Identifying the specific threshold where discretization becomes more robust than continuous modeling
– Determining which ethnic group pair exhibits this critical transition
3. Evidence Requirements
| Requirement | Description | Evidence Level |
|---|---|---|
| Cross-Ethnic Data | Comprehensive dataset spanning multiple ethnic groups | A (Empirical) |
| Variability Metrics | Quantitative measures of discrete variability | B (Theoretical) |
| Intersection Point | Precise location of the AUC curve crossing | C (Pending) |
Conclusion
The intersection of intercept-slope crossing points and cross-ethnic AUC decay curves represents a critical threshold in the migration of the three-factor constitution discrimination model. Whether this intersection exists and which ethnic group pair it falls upon must be determined through rigorous empirical measurement rather than arbitrary selection. This approach ensures that the trade-off between discretization and robustness is accurately captured across different populations, aligning with both modern statistical frameworks and traditional Yijing structural analogies.
Body
Discriminating constitution via three factors (wind/bile/phlegm or the corresponding wind/fire/dampness readings) requires cross-population migration of the model, a step that cannot be avoided. However, the attenuation of discriminative power across populations has been thoroughly measured in other domains: 245 polygenic scores trained on the UK Biobank showed systematic decline in predictive accuracy as ancestral distance increased, with non-European groups experiencing the most severe attenuation (PMID 34995502); leveraging shared genetic effects across populations can elevate portability, yet the ceiling is determined by the genetic diversity of the training data (PMID 38123642 / PMID 34887591). Physiological readings used for constitution discrimination (HRV, autonomic response) likewise possess ancestral baseline structures: systematic reviews support differences in resting HRV between populations (PMID 25551201), the association between Prakriti typology and tilt-test HRV holds in large single-population cohorts (PMID 36106418), and the link between dosha and sound-evoked responses occupies an A-level edge within L1 (PMID 29269119). The problem is that all this evidence derives from single-population cohorts; no study has tested “which encoding scheme attenuates more slowly.”
To compare two encodings, a common language is required. Calibration stratification theory decomposes model error into two layers: the intercept layer (overall probability shifted high/low) and the slope layer (readout discriminability flattened or amplified) (PMID 26772608); “calibration is the Achilles’ heel of predictive analysis”—even if AUC appears acceptable, calibration drift constitutes the primary pathway to cross-population failure (PMID 31842878). The two encodings correspond to two failure modes: the categorical model compresses continuous readings into bins, sacrificing slope resolution in exchange for low sensitivity to baseline drift (low intercept shift, flat slopes); the z-score continuous model preserves slope, but when means and variances differ across populations, the z-score anchor shifts. If their attenuation slopes differ, the two AUC decay curves will intersect—the categorical model dominates in low-attenuation regions, while the continuous model overtakes it in high-dispersion populations.
Evidence supporting “discretization incurs cost” already exists: dichotomizing continuous variables results in substantial loss of statistical power and information (PMID 16675816), with this power loss quantitatively re-estimated under neuroscience sample-size conditions (PMID 38112654). Thus, the cost of categorical encoding is certain; whether the resulting robustness exhibits population specificity remains untested. Head-to-head comparisons of population-specific versus general models have been conducted in breast cancer survival prediction: population-specific models did not universally outperform the general model (PMID 36653067); under time-distribution drift, domain generalization methods yielded only partial improvement (PMID 35177653); the “local recalibration” approach suggests a third solution—modify calibration rather than encoding (PMID 39292486).
The adjudication design can therefore be fully implemented: a multi-population triad cohort, both encoding models sharing the same train-test protocol, AUC decay curves fitted per population (linear vs. piecewise models compared), intersection points estimated via bootstrap CI and tested for existence; cross-point location stratified by population baseline dispersion (mean-variance ratio); finally contrasted against the “no-intersection” null model via likelihood ratio. Three outcomes carry distinct consequences: non-intersecting curves → encoding choice can be finalized once; intersecting at high-dispersion populations → “discretization trades for robustness” is a conditional conclusion requiring stated boundaries; intersection location independent of dispersion → discretization hypothesis falsified, switch to recalibration route.
Three implementation details determine adjudication quality. First, the z-score continuous model’s “anchor” must be explicitly defined: z-scoring using training population mean/variance then extrapolating, or re-z-scoring within the test population—these routes produce entirely different decay curve shapes; mixing them generates spurious intersections. Second, piecewise fitting must pre-register the breakpoint search range; otherwise “finding the intersection” degenerates into free fitting, rendering the null model’s likelihood ratio comparison meaningless. Third, power must be back-calculated from the thinnest population layer: in PRS attenuation studies, the non-European group—the thinnest layer—is precisely the most unstable estimate (PMID 34995502); constitution cohorts likely follow the same structure—if the bootstrap CI for the thin-layer intersection is so wide that “no intersection” cannot be excluded, the result should be adjudicated as “undecidable” rather than forced to an intersection point.
L1 Network Status: This topic has not yet entered the L1-corresponding network. The EDGE_TABLE contains only adjacent A-level edges for the ayurveda-dosha series (PMID 18565332 / PMID 29269119 / PMID 24191170); no edge exists linking encoding scheme × cross-population migration × AUC attenuation. The CORRESPONDENCE_NETWORK retrieval for “slope/intercept/AUC/cross-population” hits only irrelevant edges such as RSA slope and dispersion slope. This paper handles the gap under the “no-edge capped C” rule, which has been registered.
Evidence and Limitations
Verified: Systematic attenuation of discriminative power across ancestral groups in polygenic scores (PMID 34995502); loss of power and information when continuous variables are discretized (PMID 16675816, 38112654); calibration intercept-slope stratification framework (PMID 26772608, 31842878).
Partially Verified: Head-to-head comparisons show that ancestry-specific models do not universally outperform general models (PMID 36653067); domain generalization offers only partial improvement against distribution shift (PMID 35177653); the association between Prakriti typing and HRV holds in single-ancestry cohorts (PMID 36106418, 29269119).
Pending Verification: Whether the two-encoding cross-ancestry AUC attenuation curves truly intersect; whether the intersection point falls within ancestry pairs exhibiting high baseline dispersion; the applicable boundaries of “discretization for robustness.” This paper proposes a research design and lacks direct empirical measurement.
Related Reading
- Three-index ethnicity × gender interaction intensity “rest vs. reaction window reordering” adjudication
- Recomputation of “third-variable contamination” in interaction intensity ranking
- Cross-cultural frequency cognition research pool
Evidence level: C
References: PMID 34995502, PMID 38123642, PMID 34887591, PMID 25551201, PMID 36106418, PMID 29269119, PMID 26772608, PMID 31842878, PMID 16675816, PMID 38112654, PMID 36653067, PMID 35177653, PMID 39292486
Evidence level: A = direct PubMed support; B = scientific analogy / indirect evidence; C = traditional consensus / 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