底版减法增益的「个体间血管差异放大」裁决:person-mean centering 对谐波状态分类的 ΔAUC 增益是否随个体间血管差异而放大

Written by

in

Language

Evidence level: C

把每个人自己的谐波均值减掉(person-mean centering)再训练状态分类器,本该让模型专注学「状态」而不是学「认人」——但这个减法的增益有多大,很可能取决于队列里个体间血管差异有多大:差异越大的队列,减法越值钱。

正文

脉象谐波分析想把「寸关尺」的定性描述变成可量化的频谱指标,做法是把桡动脉压力波做傅里叶分解,得到 C0 到 C10 各阶谐波的能量分布。已有研究显示这套谱确实携带状态信息:一项 30 人可穿戴设备研究测了睡眠、运动、进食、排便、淋浴五种日常活动前后的桡动脉谐波,发现放松类活动(睡眠、排便)系统性抬升高阶谐波,负荷类活动(运动、进食)抬高低阶谐波,淋浴几乎无影响(PMID 33082826)。这说明「状态信号」真实存在于谐波谱的逐阶分布里。

但直接用原始谐波幅值做状态分类有一个结构性混淆:个体之间血管条件差异巨大。16867 人的欧洲多中心研究建立了颈股脉搏波传导速度(PWV)的参考值,年龄和血压是 PWV 的主导决定因素,且二者交互——年龄越大,血压对血管硬度的放大效应越强(PMID 20530030)。PPG 波形形态指标同样与年龄强相关,综合老化指数与年龄相关系数可达 0.56,经典的加速度脉搏图老化指数更高(PMID 28261531)。换句话说,原始谐波幅值 = 个体底版(由血管年龄、硬度、几何决定)+ 状态偏移(本次测量时的生理波动)。状态分类器吃原始特征,很容易走捷径学成「认人」而非「辨状态」。

person-mean centering 是这条捷径的减法解。方法论源头来自纵向数据的多层模型理论:时变协变量必须分解为个体内波动(within-person)与个体间差异(between-person)两个正交部分,二者含义不同、不可混用;减去个人均值只保留个体内部分,才能回答「这个人的状态变化与结果的关系」这类问题(PMID 19575624)。套到脉象谐波上:centered 特征只剩「相对自己底版偏了多少」,天然剔除「这个人是谁」的信息。

问题在于减法不是免费的。睡眠 PPG 研究提供了一个反向线索:反映动脉弹性的波形指标确实随睡眠分期变深而改善,但改善幅度随年龄显著衰减——血管越老,状态波动幅度本身越小(PMID 37379483)。这提示「个体间血管差异」不仅污染特征,还可能调制「个体内状态信号」的强度。如果分类器的判别信息部分编码在「哪类人的状态偏移模式不同」(个体×状态交互项),减法会连同这部分一起砍掉,ΔAUC 反而为负。

因此本文不预设 centering 一定有增益,而是把它做成一个可裁决的分层假设:增益是否随个体间血管差异单调放大。

实施与验收口径

  • 队列与数据:千人级队列,每人 ≥10 次标准化测量,覆盖至少 3 类已知状态标签(如餐前/餐后、运动前/后、睡眠分期);谐波全谱 C0-C10 提取统一管线(基线漂移校正、周期对齐、幅值归一化)。
  • 两套特征头对头:A=原始谐波幅值;B=person-mean centered(逐人逐阶减去个人跨次均值)。同一分类器(如梯度提升或逻辑回归)、同一交叉验证切分(按人分块,禁止同一人的样本跨训练/测试集),各跑一遍,记录 ΔAUC = AUC(B) − AUC(A)。
  • 分层裁决:按年龄三组(<40 / 40-60 / >60)、PWV 三分位、体质分型三层,各层内算 ΔAUC。检验「个体间血管差异放大」假设 = ΔAUC 随分层呈现单调递增(Jonckheere 趋势检验,α=0.05)。
  • 重测噪声带:用同一人短间隔重复测量估计 AUC 的重测标准差,构成噪声带。任何层的 ΔAUC 绝对值落在噪声带内判为「无差异」,不得声称增益。
  • 证伪条件:①若所有层 ΔAUC 均在噪声带内 → 裁决为减法无关紧要;②若 ΔAUC 与 PWV 分层无单调关系 → 「血管差异放大增益」假设被证伪,但 centering 仍可能因其他机制有效;③若 ΔAUC 显著为负 → 证明判别信息部分储存在个体×状态交互中,减法会伤及状态信号本身。

证据与局限

已验证: 谐波谱携带日常状态信息(PMID 33082826);个体间血管差异由年龄与血压主导且可量化(PMID 20530030);person-mean centering 的分层方法论在纵向统计中有坚实基础(PMID 19575624)。

部分验证: 状态波动幅度随血管年龄衰减(睡眠 PPG,PMID 37379483);波形形态指标与年龄强相关(PMID 28261531)——但均为形态学指标,尚未在谐波逐阶谱上系统复现。

待验证: 「centering 增益随个体间血管差异放大」是本文提出的可测假设,无直接文献;千人队列分层 ΔAUC 设计尚未实施;个体×状态交互被减法损伤的可能性同样待测。

相关阅读

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/谐波相关仅有 1d-model-pulse-tcm-classification-blind-comp 与 pulse-harmonic-organ-function-scoring-system 两条边,均不涉及 person-mean centering / 个体间血管差异放大;CORRESPONDENCE_NETWORK.md 中 centering/个体内中心化/底版 零命中。

证据等级: C
参考文献: PMID 19575624, PMID 20530030, PMID 33082826, PMID 37379483, PMID 28261531

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 ΔAUC gain of person-mean centering for harmonic state classification increase with inter-subject vascular differences

Evidence level: C

Subtracting each individual’s own harmonic mean (person-mean centering) before training the state classifier is intended to focus the model on learning “state” rather than “recognizing the person.” However, the magnitude of this subtraction’s gain likely depends on the degree of inter-individual vascular heterogeneity within the cohort: cohorts with greater vascular differences derive more benefit from this subtraction.

Body

Harmonic analysis of pulse waveforms aims to transform the qualitative description of “Cun-Guan-Chi” (寸关尺) into quantifiable spectral indices. This is achieved by applying Fourier decomposition to radial artery pressure waves, yielding the energy distribution across harmonics C0 through C10. Existing research indicates that this spectrum indeed carries status information: a study involving 30 participants using wearable devices measured radial artery harmonics before and after five daily activities—sleep, exercise, eating, defecation, and showering. The results showed that relaxing activities (sleep, defecation) systematically elevated higher-order harmonics, while load-bearing activities (exercise, eating) raised lower-order harmonics; showering had almost no effect (PMID 33082826). This demonstrates that “status signals” are genuinely embedded in the stepwise distribution of the harmonic spectrum.

However, directly using raw harmonic amplitudes for status classification introduces a structural confounding factor: individual vascular conditions vary significantly. A European multicenter study of 16,867 participants established reference values for carotid-femoral pulse wave velocity (PWV), identifying age and blood pressure as the dominant determinants of PWV, with an interactive effect—the older the individual, the stronger the amplifying effect of blood pressure on arterial stiffness (PMID 20530030). Similarly, PPG waveform morphology indices correlate strongly with age; the comprehensive aging index and its correlation coefficient with age can reach 0.56, while the classic acceleration pulse waveform aging index is even higher (PMID 28261531). In other words, raw harmonic amplitude = individual baseline (determined by vascular age, stiffness, and geometry) + status offset (physiological fluctuation at the time of measurement). Classification models relying on raw features are prone to taking shortcuts, learning to “recognize the person” rather than “discern the status.”

Person-mean centering is the subtractive solution to this shortcut. Its methodological origin lies in multilevel modeling theory for longitudinal data: time-varying covariates must be decomposed into within-person fluctuations and between-person differences, two orthogonal components with distinct meanings that cannot be conflated; subtracting individual means retains only the within-person component, enabling answers to questions such as “how does this person’s status change relate to the outcome” (PMID 19575624). Applied to pulse harmonics, centered features retain only “how much the current reading deviates from this individual’s baseline,” naturally eliminating information about “who this person is.”

The problem is that subtraction is not free. A reverse clue emerges from sleep PPG research: waveform indices reflecting arterial elasticity do indeed improve as sleep stages deepen, but the magnitude of improvement significantly attenuates with age—the older the vasculature, the smaller the inherent amplitude of status fluctuations (PMID 37379483). This suggests that “inter-individual vascular differences” not only contaminate features but may also modulate the intensity of “intra-individual status signals.” If part of the classifier’s discriminative information is encoded in “which type of person exhibits different status offset patterns” (individual × status interaction terms), subtraction will cut away this component as well, potentially resulting in a negative ΔAUC.

Therefore, this paper does not presuppose that centering necessarily yields gains but instead formulates it as a testable hierarchical hypothesis: whether gains increase monotonically with inter-individual vascular differences.

Implementation and Acceptance Criteria

  • Cohorts and Data: A cohort of approximately 1,000 individuals, each undergoing ≥10 standardized measurements, covering at least three categories of known state labels (e.g., pre/post meal, pre/post exercise, sleep stages). A unified pipeline extracts the full harmonic spectrum C0–C10 (baseline drift correction, cycle alignment, amplitude normalization).
  • Two Feature Heads Head-to-Head: A = raw harmonic amplitudes; B = person-mean centered (subtracting each individual’s cross-session mean from each order). Both are evaluated using the same classifier (e.g., gradient boosting or logistic regression) and the same cross-validation split (blocked by person, prohibiting samples from the same individual across training/test sets). Each is run separately, recording ΔAUC = AUC(B) − AUC(A).
  • Stratified Verdict: Three age strata (<40 / 40–60 / >60), three PWV tertiles, and three body constitution types. Within each stratum, ΔAUC is computed. The hypothesis of “inter-individual vascular differences amplification” is tested as a monotonic increase of ΔAUC across strata (Jonckheere trend test, α=0.05).
  • Retest Noise Band: The retest standard deviation of AUC is estimated by repeated short-interval measurements on the same individual, forming a noise band. Any ΔAUC absolute value falling within this noise band is judged as “no difference”; no gain may be claimed.
  • Falsification Conditions: ① If all strata ΔAUC lie within the noise band → verdict: subtraction is immaterial; ② If ΔAUC shows no monotonic relationship with PWV strata → the “vascular differences amplification gain” hypothesis is falsified, though centering may still be effective via other mechanisms; ③ If ΔAUC is significantly negative → this demonstrates that discriminative information resides partly in individual×state interactions, and subtraction would degrade the state signal itself.

Evidence and Limitations

Verified: Harmonic spectra carry information about daily states (PMID 33082826); inter-individual vascular differences are dominated by age and blood pressure and are quantifiable (PMID 20530030); the hierarchical methodology of person-mean centering has a solid foundation in longitudinal statistics (PMID 19579624).

Partially Verified: The amplitude of state fluctuations attenuates with vascular age (sleep PPG, PMID 37379483); waveform morphology indices correlate strongly with age (PMID 28261531)—but these are morphological indices and have not yet been systematically reproduced on the harmonic stepwise spectrum.

Pending Verification: The hypothesis that “centering gain is amplified by inter-individual vascular differences” is a measurable assumption proposed in this paper, with no direct literature; the thousand-person cohort stratified ΔAUC design has not yet been implemented; the possibility of subtractional damage to individual×state interactions also awaits testing.

Related Reading

L1 Network Status: This topic has not yet entered the L1 corresponding network. In EDGE_TABLE_science-yixue.md, the pulse/harmonic-related entries include only 1d-model-pulse-tcm-classification-blind-comp and pulse-harmonic-organ-function-scoring-system; neither involves person-mean centering or amplification of inter-individual vascular differences. In CORRESPONDENCE_NETWORK.md, centering/individual-mean centering/baseplate yielded zero hits.

Evidence Level: C
References: PMID 19575624, PMID 20530030, PMID 33082826, PMID 37379483, PMID 28261531

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.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *