扣掉节律之后更像你自己了吗:用识别力 AUC 裁决校正谱到底买到了什么

Written by

in

Language

Evidence level: B

稳定性回答的是”同一个人两次测像不像”,识别力回答的是”两个人之间分不分得开”。扣掉节律能改善前者,但对后者可能一点用都没有,甚至帮倒忙。

正文

前面两篇一直在算一致性:扣掉昼夜与季节节律之后,脉搏谐波谱的重测 ICC 会涨多少、涨在哪几阶。但一致性高不等于这个人可辨认。这是两个数学上完全不同的问题,而脉象文献把两者混为一谈已经很久了。

先看差别在哪。ICC 是一个比值:个体间方差除以(个体间方差 + 个体内方差)。AUC 问的是另一件事:随便抽两次测量,来自同一个人的那对的相似度,超过来自不同人的那对的概率是多少。前者只看”波动占多大比例”,后者看”人和人之间的间隔,相对每个人的抖动有多大”。一个人可以极其稳定(ICC 0.95),但如果他和邻居本来就长得一样(间隔窄),AUC 依然逼近 0.5。反过来,抖动大但个体间隔更宽,AUC 照样能上 0.9。

这个区分不是文字游戏,它决定了节律校正到底该不该做。

校正有两种性质,ICC 分不出来,AUC 能。 如果把所有人按同一时刻的平均节律曲线扣掉——这是最常见也最容易实现的做法——本质上是把节律当成共同噪声。此时个体内方差下降、个体间方差不动,ICC 必然上涨。但 AUC 涨不涨,取决于节律里有没有个体成分。如果节律是全民共享的(所有人同相、同幅),扣它就是纯降噪,AUC 跟着涨;如果每个人有自己的相位和幅度——有人是早起型、有人是夜猫型,有人季节反应强、有人几乎不动——那么扣掉群体平均曲线,等于把每个人身上最独特的那部分一起削掉了,AUC 反而会跌。

所以 AUC 是唯一能裁决”节律是噪声还是身份”的指标,而 ICC 对此完全瞎。 前面那篇逐阶 ΔICC 的剖面,无论节律是哪种性质,都会画出上升曲线。

现有文献正好把这条路走了一半。脉搏波做生物识别已经相当成熟:40 人的小样本上用二维 Gram 矩阵加深度网络能到 99% 准确率(PMID 39796830);孪生网络做单周期匹配,AUC 0.988、准确率 0.972(PMID 37430548);跨 ECG/EEG/PPG 的统一框架里,多数数据集识别准确率超过 96%,且用 SHAP 拆出最有判别力的特征是小波分解的第 1 层与第 6 层系数以及收缩期形态(PMID 41812367)。这些数字看着漂亮,但有一篇复现研究把话说明白了:多数工作建立在受控环境、小样本、短期评估、且没有”未见用户”测试之上,一旦进入真实场景、面对信号变异与训练测试之间的时间间隔,性能会大幅下滑(PMID 41471580)。他们给出的整改清单第一条就是时间切分,而不是随机切分。

最接近”节律校正用于识别”的一项工作,做法是把心电与脉搏建模成变节律的循环随机过程,先用估计出的节律函数对齐心动周期再提特征,运动中 PPG 准确率 0.962、HTER 0.039(PMID 42385745)。但作者自己写得很清楚:该数据集每人每个活动只有一个记录会话,所以这是会话内可分性,不能当作长期身份恒定性的证据。换句话说,”先对齐节律再识别”这条路只在同一次坐下来的几十分钟里被验证过,跨时辰、跨季节一次都没测。

还有两个设计上的坑必须先填。第一是样本量本身会压低 AUC。用个性化心搏做 ECG 识别,参考库从 10 人扩到 230 人,12 导联准确率从 98.4% 掉到 87.4%(PMID 29382064)。所以”40 人 AUC 0.99″和”200 人 AUC 0.85″不是同一件事,跨研究比 AUC 毫无意义,这也是一致性文献里那句老话的翻版:相关系数类指标高度依赖测量值的分布范围,不能跨样本比较(PMID 9820922),选哪种 ICC 形式也必须显式交代(PMID 27330520)。第二是识别的可能不是人,是年龄。脉搏谱被反复证明能判血管老化(PMID 34508787)、能分代谢综合征(PMID 27067750,C1、C2 显著变小而 C4 至 C9 变大),这些恰好是”个体间差异”的主要来源。如果冒充者池里年龄、血压、体型都不匹配,模型很可能只是在猜年龄段。必须做年龄性别血压匹配的冒充者池。

可执行的裁决设计:同一批不少于 120 人,每人跨两个季节各测 6 次以上、覆盖不同时辰。三臂头对头——原始谱、群体节律校正谱、以及不做任何建模只用 24 小时均值(24 小时平均的增强指数与脉搏波速度 ICC 已证实大于 0.8,PMID 24063914,这是最笨但最不会过拟合的对照)。训练用季节 A、测试用季节 B,冒充者池按年龄性别血压匹配。主终点是开集识别 AUC 的差值及 bootstrap 置信区间,同时把 ΔAUC 分解成个体间方差变化与个体内方差变化两部分。

三种结局都有意义。校正后 AUC 显著高于原始谱——节律是共同噪声,扣掉确实更可辨人,中医”脉有常”这一侧拿到硬支撑。校正后 AUC 与原始谱无差异、但 ICC 明显上升——那说明前面几篇算到的稳定性增益只是把抖动压小了,身份信息一分没多,”更稳”和”更像你”被彻底拆开,这是对整套校正思路最不客气但最诚实的结果。校正后 AUC 反而下降——最有意思的一种:节律里带着个体成分,每个人有自己的时辰曲线,扣掉群体节律等于削掉了签名的一部分。若真如此,”脉有常”与”脉应时辰”不必再靠分配谐波阶次来调和,因为节律本身就是身份的一部分。

证据与局限

已验证: 脉搏波/PPG 具备生物识别能力,小样本受控条件下准确率可达 96% 至 99%(PMID 39796830、PMID 37430548、PMID 41812367);但复现研究显示进入真实场景、存在训练测试时间间隔后性能显著下降,现有文献普遍缺时间切分与未见用户测试(PMID 41471580);节律自适应对齐后会话内 PPG 识别准确率 0.962(PMID 42385745);识别准确率随参考库人数增加而单调下降,10 人 98.4%、230 人 87.4%(PMID 29382064);脉搏谱可判别血管老化与代谢综合征状态(PMID 34508787、PMID 27067750);24 小时平均的波反射与硬化指标 ICC 大于 0.8(PMID 24063914);节律性指标季节变异远大于血压,AVI 的 CV 为 12.6%、血压仅 5.1%(PMID 30409046);相关系数类信度指标高度依赖样本取值范围、不可跨样本比较(PMID 9820922),ICC 有十种形式须显式声明(PMID 27330520)。

部分验证: “节律校正提高 ICC”已有间接支持,但”提高 ICC 会同步提高 AUC”从未被检验——两者在数学上不必然同向,现有文献没有任何一篇在同一批人身上同时报告校正前后的一致性与判别力。SHAP 显示小波第 1、6 层与收缩期特征最具判别力(PMID 41812367),提示判别信息偏向高频细节与形态,而非低频谐波幅度,但该结论建立在单库、单次采集之上。

待验证: 群体节律扣除后个体识别 AUC 的净变化方向;节律的个体间相位与幅度差异占总节律方差的比例;跨季节外部验证下 AUC 的衰减率是否与 ICC 的衰减率同阶;以及”扣掉多少比例的群体方差”这一校正强度参数是否存在最优值(扣得越狠 ICC 越高,但 AUC 可能先升后降)。

相关阅读

证据等级: B
参考文献: PMID 41471580, PMID 39796830, PMID 37430548, PMID 41812367, PMID 42385745, PMID 29382064, PMID 27330520, PMID 9820922, PMID 34508787, PMID 27067750, PMID 30409046, PMID 24063914

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。

Without Rhythm, Do You Look More Like Yourself? What We Actually Bought When Using AUC to Arbitrate Corrected Spectra

Evidence level: B

Stability answers “Does this person look the same in two measurements?”, while identification power answers “Can we distinguish between two different people?”. Removing rhythm improves the former, but may be useless for the latter, or even counterproductive.

Main Text

The previous two articles have been calculating consistency: by removing circadian and seasonal rhythms, how much does the retest ICC of the pulse harmonic spectrum increase, and at which orders? But high consistency does not equal recognizability. These are two mathematically distinct problems, yet pulse literature has conflated them for a long time.

First, let’s clarify the difference. ICC is a ratio: between-subject variance divided by (between-subject variance + within-subject variance). AUC asks a different question: given two random measurements, what is the probability that a pair from the same person is more similar than a pair from different people? The former looks only at “what proportion of the variance is fluctuation,” while the latter looks at “how wide the gap is between people relative to each person’s jitter.” One can be extremely stable (ICC 0.95), but if they naturally resemble their neighbor (narrow gap), AUC still approaches 0.5. Conversely, if jitter is large but the inter-individual gap is wider, AUC can still reach 0.9.

This distinction is not semantic; it determines whether rhythm correction should be performed at all.

Correction has two natures that ICC cannot distinguish, but AUC can. If everyone’s average rhythm curve at the same time is subtracted—which is the most common and easiest implementation—rhythm is essentially treated as common noise. In this case, within-subject variance decreases while between-subject variance remains unchanged, so ICC must increase. However, whether AUC increases depends on whether the rhythm contains an individual component. If the rhythm is universally shared (same phase and amplitude for all), subtracting it is pure noise reduction, and AUC will rise. If everyone has their own phase and amplitude—some are early birds, some night owls, some have strong seasonal responses, others almost none—then subtracting the group average curve effectively chops off the most unique part of each individual, causing AUC to drop.

Therefore, AUC is the only metric capable of adjudicating whether rhythm is noise or identity, while ICC is completely blind to this. The previous profile of stepwise ΔICC will show an upward curve regardless of the nature of the rhythm.

Existing literature has walked half this path. Pulse wave biometric identification is already mature: a 40-person small sample achieved 99% accuracy using a 2D Gram matrix plus deep networks (PMID 39796830); twin networks for single-period matching achieved an AUC of 0.988 and accuracy of 0.972 (PMID 37430548); within a unified framework across ECG/EEG/PPG, most datasets exceed 96% identification accuracy, with SHAP analysis identifying the most discriminative features as coefficients from the 1st and 6th layers of wavelet decomposition and systolic morphology (PMID 41812367). These numbers look impressive, but one replication study clarified the situation: most works rely on controlled environments, small samples, short-term assessments, and lack “unseen user” testing; once moved to real-world scenarios facing signal variability and time intervals between training and testing, performance drops significantly (PMID 41471580). Their first recommendation for rectification is time splitting, not random splitting.

The work closest to “rhythm correction for identification” models ECG and pulse as cycle-random processes with varying rhythms, aligns cardiac cycles using estimated rhythm functions before feature extraction, achieving 0.962 accuracy and 0.039 HTER in motion PPG (PMID 42385745). However, the authors explicitly state: each participant has only one recording session per activity in this dataset; thus, this is within-session separability, not evidence of long-term identity constancy. In other words, the “align rhythm then identify” path has only been validated within a sitting session of tens of minutes; it has never been tested across hours or seasons.

Two design pitfalls must also be addressed first. First, sample size itself suppresses AUC. Using personalized heartbeats for ECG identification, expanding the reference library from 10 to 230 people caused 12-lead accuracy to drop from 98.4% to 87.4% (PMID 29382064). Therefore, “40-person AUC 0.99” and “200-person AUC 0.85” are not the same thing; cross-study comparison of AUC is meaningless, which is a reiteration of an old saying in consistency literature: reliability metrics like correlation coefficients heavily depend on the range of measurement values and cannot be compared across samples (PMID 9820922); which ICC form must also be explicitly stated (PMID 27330520). Second, identification may not be of the person, but of age. Pulse spectra have been repeatedly proven to judge vascular aging (PMID 34508787) and distinguish metabolic syndrome (PMID 27067750, where C1 and C2 significantly decrease while C4 to C9 increase)—these are exactly the main sources of “between-subject differences.” If the impostor pool does not match in age, blood pressure, or body type, the model might simply be guessing age groups. An impostor pool matched by age, gender, and blood pressure is required.

Executable adjudication design: A single batch of at least 120 people, each measured more than 6 times across two seasons covering different times of day. Three arms head-to-head: raw spectrum, rhythm-corrected spectrum, and no modeling at all using only the 24-hour mean (the enhancement index of the 24-hour average and pulse wave velocity ICC are already confirmed >0.8, PMID 24063914; this is the dumbest but least likely to overfit control). Training uses Season A, testing uses Season B; the impostor pool is matched by age, gender, and blood pressure. The primary endpoint is the difference in open-set identification AUC and its bootstrap confidence interval, simultaneously decomposing ΔAUC into changes in between-subject variance and within-subject variance.

All three outcomes are meaningful. If AUC is significantly higher after correction than the raw spectrum—rhythm is common noise, removing it makes people more distinguishable, providing hard support for the TCM concept “pulse has constants.” If AUC is unchanged but ICC rises significantly—then the stability gains calculated in previous articles merely suppressed jitter without adding any identity information; “more stable” and “more like you” are completely separated, which is the most unkind yet honest result for the entire correction framework. If AUC actually drops after correction—the most interesting case: rhythm contains individual components, everyone has their own time-specific curve, and subtracting the group rhythm chops off part of the signature. If true, there is no need to reconcile “pulse has constants” and “pulse responds to time” by assigning harmonic orders; because rhythm itself is part of identity.

Evidence and Limitations

Verified: Pulse wave/PPG possesses biometric identification capabilities; under small-sample controlled conditions, accuracy can reach 96% to 99% (PMID 39796830, PMID 37430548, PMID 41812367); however, replication studies show performance drops significantly upon entering real-world scenarios with time intervals between training and testing, and existing literature generally lacks time splitting and unseen user testing (PMID 41471580); after rhythm adaptive alignment, within-session PPG identification accuracy is 0.962 (PMID 42385745); identification accuracy monotonically decreases as the reference library size increases (98.4% for 10 people, 87.4% for 230 people) (PMID 29382064); pulse spectra can distinguish vascular aging and metabolic syndrome states (PMID 34508787, PMID 27067750); the ICC for wave reflection and stiffening indices of the 24-hour average is greater than 0.8 (PMID 24063914); rhythmic indices show seasonal variation far greater than blood pressure, with AVI CV at 12.6% vs. 5.1% for blood pressure (PMID 30409046); reliability metrics like correlation coefficients heavily depend on sample value ranges and cannot be compared across samples (PMID 9820922); there are ten forms of ICC requiring explicit declaration (PMID 27330520).

Partially Verified: “Rhythm correction improves ICC” has indirect support, but “improving ICC simultaneously improves AUC” has never been tested—mathematically they do not necessarily move in the same direction, and no existing literature reports both consistency and discriminative power before and after correction on the same batch of people. SHAP analysis indicates that wavelet layers 1 and 6 and systolic features are most discriminative (PMID 41812367), suggesting discriminative information leans toward high-frequency details and morphology rather than low-frequency harmonic amplitudes, but this conclusion is based on a single library and single acquisition.

To Be Verified: The net direction of change in individual identification AUC after subtracting group rhythm; the proportion of total rhythm variance accounted for by between-subject phase and amplitude differences; whether the attenuation rate of AUC under cross-season external validation is of the same order as the ICC attenuation rate; and whether there exists an optimal value for this correction intensity parameter (“how much group variance to subtract”—more subtraction yields higher ICC, but AUC might first rise then fall).

Related Reading

Evidence Level: B
References: PMID 41471580, PMID 39796830, PMID 37430548, PMID 41812367, PMID 42385745, PMID 29382064, PMID 27330520, PMID 9820922, PMID 34508787, PMID 27067750, PMID 30409046, PMID 24063914

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 *