状态分类前的底版减法:扣掉’你是谁’,才能看清’你现在怎样

Written by

in

Language

Evidence level: B

中医说脉”有常有变”。做状态分类时,如果”常”(个体底版)没有被扣掉,它就会伪装成”变”(状态信号)——把”这个人的脉”误判成”这个人此刻的脉”。

正文

上一篇文章提出”低阶谐波承载身份、高阶谐波承载状态”的谱段分工假说(PMID 38481553、PMID 42490820),并建议用”双任务裁决”验证。但真正动手做高阶谐波状态分类时,会遇到一个被普遍低估的坑:个体间底版差异(身份噪声)会污染状态信号。同样的应激下,张三的高阶谐波抬升 20%,李四可能只抬升 5%——不是张三应激更重,而是两人的”出厂设置”不同。如果不先扣掉底版,分类器学到的可能不是”应激 vs 静息”的差别,而是”张三 vs 李四”的差别。

个体差异真实存在,且大到可以认人。 脉搏波形本身就能当生物特征:46 人 PPG 数据集用心率与呼吸衍生特征个体识别准确率达 92.9%(PMID 33018130);更早的 pilot 研究用 PPG 导数做身份验证也证明了个体内一致性显著高于个体间(PMID 18003024)。这意味着同一段脉波数据里,”是谁”的信号强到足以压过”现在怎样”的信号——状态分类若不处理身份噪声,模型会优先利用最容易学的判别维度。

通用模型输给个体专属模型,是身份噪声污染的直接证据。 2020 年一项用 PPG 做压力状态分类的研究对比了三类模型:为每个人单独训练的个体专属模型测试准确率达 99.9%,而跨人群的通用模型明显劣化;仅用测试者 20% 的样本做校准,通用模型就显著回升(PMID 32957479)。同一批受试者、同样的压力任务,差别只在”模型知不知道这个人”——身份噪声不是理论担忧,而是实测的准确率缺口。

心率归一化的先例证明”底版减法”有效。 超短期压力测量领域有同样的困扰:PRV(脉搏率变异性)特征在同一压力水平下随心率个体差异变化。2025 年研究改用 NNI 计数分段后,压力指数预测 R² 从 0.42 提升到 0.94(PMID 39392540)——只做了一步”按个体心率分段”,个体差异造成的误差就被大幅压掉。这为谐波状态分类提供了现成的方法学模板。

状态信号确实存在于高阶谐波。 2026 年 348 例冠状动脉造影患者的谐波研究,用双手谐波参数差值判别 SYNTAX 评分≥22,AUC 达 0.89(PMID 40396324);系统综述确认情绪、认知负荷、自主神经活动都会系统性改写谐波谱(PMID 38481553);智能手机 PPG 在 127 人中证实脉搏波特征与负性心理状态显著关联(PMID 41629260)。状态信号是真实存在的——问题只在如何把它从身份噪声中分离出来。

底版减法的方法学选择。 多层模型文献对此有清晰建议:person-mean centering(个体均值中心化)把每个观测减掉该个体自己的均值,在统计上把”组间效应”(个体差异)与”组内效应”(状态变化)分离(PMID 25822206);grand-mean centering(总体均值中心化)则保留个体差异,两者拟合的是不同模型(PMID 17563168)。对状态分类而言,应该用 person-mean centering——把”这个人的常”先扣掉,剩下的才是”此刻的变”。

可检验的设计。 这是”状态分类前底版减法增益”的裁决实验:千人队列同步采集桡动脉高阶谐波全谱与 HRV,每人多次访问(静息/应激/运动后/进食后),两套特征工程对比——原始谐波特征 vs person-mean centering 后谐波特征,各自训练状态分类模型,检验 ΔAUC 与 NRI(净重分类改善)是否显著为正(PMID 23254468、PMID 25176621)。三种结局都是知识:底版减法显著提升→身份噪声确实污染状态分类,”扣常观变”获量化支撑;无提升→高阶状态信号本就个体无关,现行”谐波-脏腑”查表可放心跨人使用;部分提升→存在个体×状态交互,需按体质/血管类型分层建模。

为什么这对科学易学重要。 中医脉诊里”常脉”与”病脉”的区分历来依赖医家经验——同一个滑脉,在平和体质者身上是常态,在痰湿体质者身上就是病理。底版减法给了这条经验一个统计形式:先为每个人建立脉象底版,再看偏离。体质研究取身份分量做人群聚类、状态研究取状态分量做时辰对比,两者都能受益于这一步前置处理。这正是”常变分离”从心法变成算法的关键一步。

证据与局限

已验证: 脉搏波形个体识别可行(PMID 33018130、PMID 18003024);个体专属 PPG 压力模型显著优于通用模型(PMID 32957479);NNI 分段(个体心率归一化)将压力预测 R² 从 0.42 提至 0.94(PMID 39392540)。

部分验证: 高阶谐波携带状态信号(PMID 40396324、PMID 38481553、PMID 41629260);person-mean centering 分离组间/组内效应的统计性质明确(PMID 25822206、PMID 17563168)——但”底版减法对谐波状态分类的 ΔAUC 增益”尚无直接实测。

待验证: 千人队列上原始谱 vs 底版归一化谱的状态分类 ΔAUC/NRI 头对头;底版减法增益是否随个体间血管差异(年龄/体质/动脉硬化程度)放大;个体×状态交互的占比与分层建模收益。

相关阅读

证据等级: B(个体识别、个体专属模型优势、心率归一化增益均有 PubMed 直接支撑;”person-mean centering 对谐波状态分类的 ΔAUC 增益”直接研究尚不存在)
参考文献: PMID 33018130, PMID 18003024, PMID 32957479, PMID 39392540, PMID 40396324, PMID 38481553, PMID 41629260, PMID 25822206, PMID 17563168, PMID 23254468, PMID 25176621


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。

Subtraction Before State Classification: Subtracting ‘Who You Are’ to Reveal ‘How You Are Now

Evidence level: B

Traditional medicine describes the pulse as having both a constant baseline (“constant”) and variable states (“change”). When conducting state classification, if the “constant” (individual baseline) is not subtracted, it masquerades as “change” (state signal)—misjudging “this person’s pulse” as “this person’s pulse at this moment.”

Main Text

The previous article proposed a spectral division hypothesis: “low-order harmonics carry identity, high-order harmonics carry state” (PMID 38481553, PMID 42490820), suggesting verification via a “dual-task adjudication.” However, when actually performing high-order harmonic state classification, one encounters an underestimated pitfall: inter-individual baseline differences (identity noise) contaminate state signals. Under the same stressor, Zhang San’s high-order harmonic elevation may be 20%, while Li Si’s might only be 5%—not because Zhang San experiences more stress, but because their “factory settings” differ. Without first subtracting the baseline, the classifier may learn the difference between “Zhang San vs. Li Si” rather than “stress vs. rest.”

Individual differences are real and large enough to identify individuals. Pulse waveforms themselves function as biometric features: a 46-person PPG dataset achieved 92.9% individual identification accuracy using heart rate and respiration-derived features (PMID 33018130); an earlier pilot study using PPG derivatives also demonstrated that intra-individual consistency significantly exceeds inter-individual variability (PMID 18003024). This means that within a single pulse wave dataset, the signal of “who” is strong enough to overwhelm the signal of “how now”—if state classification does not handle identity noise, models will prioritize the easiest-to-learn discriminative dimensions.

General models underperform individual-specific models, serving as direct evidence of identity noise contamination. A 2020 study using PPG for stress state classification compared three model types: individual-specific models trained for each person achieved 99.9% test accuracy, whereas cross-population general models showed significant degradation; using only 20% of the test subject’s samples for calibration caused general models to recover significantly (PMID 32957479). With the same subjects and stress tasks, the difference lies solely in whether “the model knows this person”—identity noise is not just a theoretical concern but a measured accuracy gap.

Precedents of heart rate normalization demonstrate the efficacy of “baseline subtraction.” In the field of ultra-short-term stress measurement, similar challenges exist: PRV (pulse rate variability) features vary with individual heart rate differences at the same stress level. A 2025 study switched to NNI counting segments, raising the stress index prediction R² from 0.42 to 0.94 (PMID 39392540)—a single step of “segmenting by individual heart rate” drastically reduced error caused by individual differences. This provides a ready-made methodological template for harmonic state classification.

State signals do exist in high-order harmonics. A 2026 study of 348 coronary angiography patients used harmonic parameter differences between hands to discriminate SYNTAX scores ≥22, achieving an AUC of 0.89 (PMID 40396324); a systematic review confirmed that emotions, cognitive load, and autonomic activity systematically rewrite harmonic spectra (PMID 38481553); smartphone PPG confirmed significant associations between pulse wave features and negative psychological states in 127 individuals (PMID 41629260). State signals are real—the issue lies only in how to separate them from identity noise.

Methodological choices for baseline subtraction. Literature on multi-level models offers clear guidance: person-mean centering subtracts each observation by that individual’s own mean, statistically separating “between-group effects” (individual differences) from “within-group effects” (state changes) (PMID 25822206); grand-mean centering retains individual differences, fitting different models (PMID 17563168). For state classification, person-mean centering should be used—to first subtract “this person’s constant,” leaving only “this moment’s change.”

Testable design. This is the adjudication experiment for “baseline subtraction gain prior to state classification”: a thousand-person cohort synchronously collects full-spectrum high-order harmonics and HRV from the radial artery, with multiple visits per person (rest/stress/post-exercise/post-prandial). Two feature engineering approaches are compared: raw harmonic features vs. person-mean centered harmonic features, each training a state classification model to test if ΔAUC and NRI (Net Reclassification Improvement) are significantly positive (PMID 23254468, PMID 25176621). Three outcomes provide knowledge: significant gain → identity noise indeed contaminates state classification, “subtract constant, observe change” gains quantitative support; no gain → high-order state signals are inherently individual-irrelevant, allowing current “harmonic-organ” lookup tables to be safely cross-person used; partial gain → individual × state interaction exists, requiring stratified modeling by constitution/vessel type.

Why this matters for scientific Yijing Studies. The distinction between “constant pulse” and “pathological pulse” in Traditional Chinese Medicine pulse diagnosis has historically relied on physician experience—a slippery pulse is normal for a healthy constitution but pathological for a phlegm-damp constitution. Baseline subtraction gives this experience a statistical form: first establish an individual pulse baseline, then observe deviations. Constitution research uses identity components for population clustering, while state research uses state components for time-of-day comparison; both benefit from this pre-processing step. This is the critical transition from “method of mind” to algorithm for “constant-change separation.”

Evidence and Limitations

Verified: Pulse waveform individual identification is feasible (PMID 33018130, PMID 18003024); individual-specific PPG stress models significantly outperform general models (PMID 32957479); NNI segmentation (individual heart rate normalization) raises stress prediction R² from 0.42 to 0.94 (PMID 39392540).

Partially Verified: High-order harmonics carry state signals (PMID 40396324, PMID 38481553, PMID 41629260); the statistical property of person-mean centering separating between/within-group effects is clear (PMID 25822206, PMID 17563168)—but “baseline subtraction gain for ΔAUC in harmonic state classification” lacks direct empirical measurement.

To be Verified: Head-to-head comparison of ΔAUC/NRI for raw spectra vs. baseline-normalized spectra in a thousand-person cohort; whether baseline subtraction gain amplifies with inter-individual vascular differences (age/constitution/arteriosclerosis); the proportion of individual × state interaction and benefits of stratified modeling.

Related Reading

Evidence Level: B (Individual identification, individual-specific model advantages, and heart rate normalization gains have direct PubMed support; “person-mean centering gain for ΔAUC in harmonic state classification” lacks direct research)
References: PMID 33018130, PMID 18003024, PMID 32957479, PMID 39392540, PMID 40396324, PMID 38481553, PMID 41629260, PMID 25822206, PMID 17563168, PMID 23254468, PMID 25176621


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 *