谐波状态分类的「个体×状态交互」占比裁决:底版减法后仍无法解释的残差里,藏着多少只属于你自己的规律

Written by

in

Language

Evidence level: C

同一个状态(紧张、运动、睡眠)对每个人的脉象谐波改写量并不一样——这个「不同人不一样」的部分占了多少方差,直接决定谐波状态分类该做通用模型还是个性化模型。

正文

前两篇我们拆了两层:一层是「底版减法」本身能带来多大增益(个体间血管差异会不会被放大),另一层是常脉的长期可识别性。这一篇轮到第三层,也是最关键的一层:底版减法做完之后,仍然解释不了的残差里,「个体×状态交互」到底占多少。

先把三个成分摆清楚。把每个人在不同状态下测到的谐波特征(各阶相对能量、H2/H1、H3/H1 等)放进方差分解框架,总方差可以拆成三块:个体主效应(你我的底子不同,血管粗细、弹性、年龄带来的静态差异)、状态主效应(紧张让所有人低阶谐波抬升这类「普遍反应」)、以及个体×状态交互(同样是紧张,A 的 H1 飙 30%,B 只动 5%,甚至方向相反)。前两块好理解,第三块最容易被忽略,却恰恰是「状态读数能不能跨人通用」的命门。

交互不是空想,文献里已经有三条独立线索。第一条来自压力分类:PPG 应激检测里,通用模型在训练集上看着很好,换到没见过的人身上测试准确率从约 97% 掉到约 61%;给每个新人补 20% 校准样本就能拉回 88%——这说明应激对波形的改写方式高度因人而异,通用模型默认「人人同款反应」是错的(PMID 32957479)。第二条来自情绪生理学的方差排序:32 人两天多情境采集的 PPG 特征,方差从大到小依次是个体间差异、日间差异,最后才是情绪本身——状态信号不但排在最后,而且区分愤怒的特征在不同人、不同日子上稳定性不一致,这正是「交互存在」的直白表述(PMID 37582733)。第三条来自睡眠分期的个性化实验:通用 PPG 睡眠模型考虑到个体自主神经平衡差异后做个性化微调,82.5% 的受试者性能提升,年轻人和失眠人群收益最大——个性化收益本身就是交互占比的宏观代理指标(PMID 40882681)。

对脉象谐波来说,状态信号本身是实的:30 人在五类日常活动下的桡动脉谐波显示状态确实在谐波结构上留痕(PMID 33082826)。但「状态留痕」和「状态留痕的方式人人一样」是两回事。方法学上,把个体内与个体间效应分开处理、并在混合效应模型里显式放交互项,是纵向数据分解的标准做法(PMID 19575624)。

这条线的易学对接很直接。中医讲「因人制宜」,脉书里的「常脉」本来就承认各人有各人的常——用现代话说,就是承认个体×状态交互不可约简。如果交互方差占比真的很大,那么任何「一套阈值判所有人的紧张脉」的做法都在系统性地误读,分层建模(按体质、按血管年龄)甚至完全个性化才是正路;反过来,如果交互占比很小,通用读数就值得坚持,个性化只是锦上添花。

裁决设计:同一批受试者×多状态×多日重复测量,谐波特征做三方差分解——混合效应模型放个体随机截距与随机斜率,交互方差占比为首要读数;预注册阈值(例如交互占比超过状态主效应的一半)触发分层建模分支;然后头对头比较全局模型与「按交互占比画像分层」模型的分类增益,用留一人交叉验证避免个性化收益被泄漏夸大。三条证伪条件:① 交互方差占比低于预设阈值且置信区间不跨阈——全局模型胜出;② 交互占比高但画像分层带不来显著增益——说明交互虽在、但画像变量抓不住它;③ 分层增益不随交互占比增大而增大——因果链断裂。

证据与局限

已验证: 谐波特征随状态改变(PMID 33082826);PPG 类信号里个体差异主导方差排序(PMID 37582733);通用模型跨人衰减、个性化可挽回(PMID 32957479、40882681)。

部分验证: 交互效应在 PPG 心理生理任务里被间接证实,但全部是间接证据——没有一项在「底版归一化后的脉象谐波残差」上直接量化交互占比。

待验证: 脉象谐波残差中交互占比的具体数值;体质/血管画像能否解释交互部分;分层建模相对全局模型的增益幅度。

相关阅读

证据等级: C
参考文献: PMID 32957479, PMID 37582733, PMID 40882681, PMID 33082826, PMID 19575624

L1 网络状态:该主题尚未进入 L1 对应网络——EDGE_TABLE 中脉象/谐波相关仅 1d-model-pulse-tcm-classification-blind-comp(频率/心率/共振/谐波/脉象条目)与 pulse-harmonic-organ-function-scoring-system(六腑)两条边,均无「个体×状态交互 / 方差分解 / 个性化建模」维度;CORRESPONDENCE_NETWORK.md 无对应边。本文按仅文献证据处理,证据等级封顶 C。


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。

Individual-State Interaction Proportion in Harmonic State Classification: How Many of Your Own Patterns Are Hidden in the Residuals That Remain Unexplained After Subtracting the…

Evidence level: C

In any given state (tension, exercise, sleep), the magnitude of harmonic waveform modification induced varies across individuals; the proportion of variance attributable to this “individual difference” directly determines whether harmonic state classification should rely on a general model or an individualized model.

Body

In the previous two articles, we decomposed two layers: first, how much gain the “baseline subtraction” itself can provide (whether inter-individual vascular differences will be amplified), and second, the long-term identifiability of normal pulses. This article addresses the third layer, which is also the most critical: within the residual variance that remains unexplained after baseline subtraction, what proportion is attributable to the “individual × state interaction.”

Let us clarify three components. We place the harmonic features measured in each individual under different states—specifically, relative energies of various orders, H2/H1, H3/H1, etc.—into a variance decomposition framework. The total variance can be decomposed into three parts: individual main effects (static differences arising from individual baselines, such as variations in vessel caliber, elasticity, and age), state main effects (universal responses, such as stress causing an elevation in low-order harmonics across all individuals), and individual × state interactions (where, under the same stress condition, A’s H1 surges by 30%, while B shifts only 5%, or even in opposite directions). The first two components are straightforward; the third is most easily overlooked yet precisely the crux of whether “state readings can be universally applied across individuals.”

Interaction is not merely speculative; literature provides three independent lines of evidence. First, from stress classification: in PPG stress detection, general models appear robust on the training set but drop to approximately 61% accuracy when tested on unseen individuals, compared to about 97% on the training set; adding just 20% calibration samples per new individual restores accuracy to 88%. This indicates that stress rewrites waveforms in highly individualized ways, rendering the general model’s default assumption of “everyone reacts alike” incorrect (PMID 32957479). Second, from emotion physiology variance ranking: in a study of 32 individuals across multiple contexts over two days, PPG feature variance ranked from largest to smallest as inter-individual differences, diurnal differences, and finally emotion itself—meaning state signals not only rank last but also exhibit inconsistent stability across individuals and days for distinguishing anger, which directly expresses the existence of interaction (PMID 37582733). Third, from personalized sleep staging experiments: after applying individualized fine-tuning to a general PPG sleep model to account for differences in autonomic balance, performance improved by 82.5% across subjects, with the greatest gains observed in young individuals and those with insomnia; this personalized gain itself serves as a macro proxy for the proportion of interaction variance (PMID 40882681).

Regarding pulse harmonic features, state signals are indeed real: harmonic analysis of the radial artery in 30 individuals across five daily activities shows that state leaves traces on the harmonic structure (PMID 33082826). However, “state leaves traces” and “everyone leaves traces in the same way” are distinct matters. Methodologically, separating within-individual and between-individual effects and explicitly including interaction terms in mixed-effects models constitutes the standard approach for decomposing longitudinal data (PMID 19575624).

The alignment with Yijing Studies is direct. Traditional Chinese medicine emphasizes “individualized treatment” (因人制宜), and the classical pulse texts already acknowledge that each person has their own “normal pulse”—in modern terms, this acknowledges the irreducibility of individual × state interaction. If the proportion of interaction variance is indeed large, any approach relying on “a single threshold to judge stress pulses for everyone” systematically misreads the data; instead, layered modeling (by constitution, by vascular age) or fully personalized approaches become necessary. Conversely, if the interaction proportion is small, universal readings remain justified, with personalization serving merely as refinement.

Decision Design: Conduct repeated measurements on the same cohort across multiple states and days, then perform three-way variance decomposition of harmonic features using mixed-effects models with random intercepts and random slopes for individuals, taking the interaction variance proportion as the primary readout. Pre-register thresholds (e.g., triggering layered modeling if interaction variance exceeds half of the state main effect) and directly compare classification gains between the global model and the “interaction-profiled layered” model, using leave-one-out cross-validation to avoid inflating personalized gains through data leakage. Three falsification conditions: ① interaction variance proportion falls below the preset threshold and confidence intervals do not cross it—global model wins; ② high interaction proportion but no significant gain from profile-based layering—indicating that while interaction exists, the profiling variables fail to capture it; ③ layered gain does not increase with rising interaction proportion—causal chain broken.

Evidence and Limitations

Verified: Harmonic features shift with state (PMID 33082826); individual differences dominate variance ordering in PPG-like signals (PMID 37582733); general models degrade across individuals, but personalization recovers performance (PMID 32957479, 40882681).

Partially Verified: Interaction effects were indirectly confirmed in PPG psychophysiological tasks, yet all evidence remains indirect—no study has directly quantified the proportion of interaction within the residual harmonics after baseline normalization.

Pending Verification: The specific magnitude of the interaction component within harmonic residuals; whether constitution/vascular profiles can explain the interactive portion; and the gain of hierarchical modeling relative to global models.

Related Reading

Evidence level: C
References: PMID 32957479, PMID 37582733, PMID 40882681, PMID 33082826, PMID 19575624

L1 network status: This topic has not yet entered the L1 corresponding network. In EDGE_TABLE, the only edges related to pulse/harmonic are 1d-model-pulse-tcm-classification-blind-comp (frequency/heart rate/resonance/harmonic/pulse) and pulse-harmonic-organ-function-scoring-system (six fu organs); neither includes the “individual × state interaction / variance decomposition / personalized modeling” dimension. CORRESPONDENCE_NETWORK.md lacks corresponding edges. This article is processed based on literature evidence only, with the evidence level capped at C.


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 *