三因脉搏谐波的\”条件鲁棒特征子集\”筛选:跨时段、体位与静息时长,哪些谐波特征最稳、哪些最脆?

Written by

in

Language

Evidence level: C

同一套脉搏谐波特征,在上午采和在下午采、躺着的和坐着的、休息五分钟和十五分钟,数值根本不是一回事——先分级,再建模,否则模型学到的是”条件指纹”而不是”体质指纹”。

正文

阿育吠陀的脉诊(nadipariksha)声称通过桡动脉触诊判断三因(vata/pitta/kapha)状态。要把这门手艺客观化,最自然的路线就是把手腕上的脉搏波换成谐波特征:用频谱分析把脉搏波拆成一系列谐波分量,再用这些分量训练分类器去预测体质分型。这条路线在中医脉诊里已有先例——桡动脉脉搏波的谐波分析在标准化协议下表现出很好的可重复性:健康成人研究中,前 11 阶谐波的观察者间 ICC 达 0.83–0.96(PMID 27170904)。

但这里有一个常被忽略的裂缝:那个 ICC 是在单一标准条件下测的。同一批指标,换个条件,信度会明显掉档。一项针对脉搏波形分离分析的可靠性研究给出了直接证据:同一组健康成人在六个不同早晨、分别以仰卧/坐姿、禁食/非禁食四种组合条件下测量,向后压力分量(Pb)与反射幅度的 ICC 在”仰卧+禁食”组合下达标(0.76 与 0.77,超过 0.75 的可接受阈),而在坐姿条件下最低掉到 0.56–0.57(PMID 27861248)。姊妹研究同样显示:中心收缩压的日间信度受体位主效应显著影响,增强指数(AIx)则受禁食状态主效应显著影响(PMID 26136065)。也就是说,“指标可重复”是有条件限定语的——最稳的条件是仰卧、禁食;一旦条件放开,同一特征的同一个人读数就漂。

这对三因脉诊客观化意味着什么?如果直接把”全特征集”喂给三因分类器,训练数据若采自混合条件(有的上午、有的下午,有的躺着、有的坐着),模型很可能把”采集条件”的系统性差异误学成”体质”的差异——因为条件对谐波形态的影响(如体位改变静脉回流与外周阻力,进而改变波形反射结构)完全可以伪造出组间差异。这还不是纯假设:人工三因脉诊本身的评定者间信度只有 poor-to-moderate(15 名阿育吠陀医师盲法检查 20 名志愿者共 600 次,加权 κ 按Landis-Koch量表偏低到中等,经验较深者 κ 更大,PMID 28664059),prakriti 体质分型的评定者一致性同样不高(PMID 24191170)。标签本身噪声就大,再叠上条件噪声,分类器的”准确率”几乎不可解释。

所以本文提出一个先于任何分类建模的预筛选步骤:条件鲁棒特征子集筛选。设计分四步:

第一步,条件矩阵采集:每位受试者在至少 2 个时段(如早/晚)× 2 个体位(仰卧/坐)× 2 档静息时长(5/15 分钟)的多条件组合下重复采集脉搏波,形成每人多条谐波谱。

第二步,逐特征跨条件 ICC 排序:对每个谐波特征(各阶幅值、幅值比、相位等),以”条件”为重复因素计算跨条件 ICC,把特征从”最稳健”到”最脆弱”排成一张分级表。预期形态:基线幅值类特征对体位敏感(对应 27861248 里 Pb 的表现),而某些幅值比或归一化特征可能天然抵消条件效应。

第三步,逐特征移除的 AUC 衰减检验(leave-one-feature-out):用预注册的参考分类器,逐一移除单个特征,看跨条件验证 AUC 掉多少。掉得多的特征是”判别主力但可能条件耦合”;掉得少甚至移除后 AUC 不变的,是冗余或噪声。

第四步,鲁棒子集 vs 全特征集头对头:只保留”ICC 高 + AUC 贡献非负”的特征构成鲁棒子集,与全特征集在跨条件泛化测试上对比。可证伪的预测是:全特征集在同条件交叉验证下 AUC 更高,但跨条件测试时塌方;鲁棒子集 AUC 绝对值略低,但跨条件衰减幅度显著更小。如果结果相反——鲁棒子集跨条件照样塌——那说明三因信号本身不稳定,这条路要重估。

这套流程的价值不在多提高准确率,而在把”条件效应”从”体质效应”里拆出去。一紧张、一换姿势、一没休息够就读数漂移的特征,放进任何”体质—健康”关联研究都是污染源;先筛掉,剩下的特征才有资格谈体质对应。

证据与局限

已验证: 脉搏谐波分析在标准化单条件下具有良好观察者间信度(前 11 阶 ICC 0.83–0.96,PMID 27170904);脉搏波形衍生指标的信度显著受体位与禁食状态影响,仰卧+禁食为最稳条件、坐姿条件下 ICC 最低降至 0.56–0.57(PMID 27861248、26136065)。

部分验证: 阿育吠陀三因脉诊的人工评定者间信度为 poor-to-moderate、与经验正相关(PMID 28664059);prakriti 分型评定者一致性偏低(PMID 24191170)。人工标签噪声大,是客观化路线的现实约束。

待验证: ①跨条件 ICC 分级在谐波特征上的具体形态(哪些阶、哪类比值特征最稳);②鲁棒子集 vs 全特征集的跨条件 AUC 衰减对比;③”条件指纹”对分类准确率的污染占比。本方案尚未在任何三因队列上实施。

相关阅读

L1 网络状态:该主题尚未进入 L1 对应网络(EDGE_TABLE 对”跨条件/条件鲁棒/特征子集/leave-one-feature-out”零命中)。最接近的相邻内容:CORRESPONDENCE_NETWORK E-061 协变量清单「采样时刻(circadian)进 E-061 协变量清单;落点判据的自然呼吸段(RR 12–20 cpm)本身随时段漂移」;KB-510「静息 ICC 0.50–0.83、CV 5%–41%;频域指标可靠性差于时间域」。

证据等级: C
参考文献: PMID 27170904, PMID 27861248, PMID 26136065, PMID 28664059, PMID 24191170

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。

Condition-Robust Feature Subset Screening of Three-Cause Pulse Harmonics: Which Harmonic Features Are Most Stable and Which Are Most Fragile Across Time, Posture, and Rest Durat…

Evidence level: C

Screening the “Conditionally Robust Feature Subset” of the Three-Cause Pulse Harmonic: Across Time, Posture, and Rest Duration, Which Harmonic Features Are Most Stable and Which Most Fragile?

The same set of pulse harmonic features yields numerically distinct values when measured in the morning versus the afternoon, while lying down versus sitting, or after five minutes versus fifteen minutes of rest; therefore, features must be stratified before modeling to ensure the model learns a “体质指纹” (constitution fingerprint)—a stable physiological signature—rather than a “条件指纹” (condition fingerprint)—a transient artifact of measurement context.

Body

Ayurvedic pulse diagnosis (nadipariksha) claims to assess the state of the three doshas (vata/pitta/kapha) via radial artery palpation. To objectively operationalize this craft, the most natural route is to replace wrist pulse waves with harmonic features: decompose the pulse wave into a series of harmonic components via spectral analysis, then train classifiers on these components to predict constitutional typologies. This route already has precedent in Chinese medicine pulse diagnosis—harmonic analysis of radial artery pulse waves demonstrates good reproducibility under standardized protocols: in healthy adult studies, the inter-observer ICC for the first 11 harmonic orders reaches 0.83–0.96 (PMID 27170904).

Yet there is a frequently overlooked gap: that ICC was measured under a single standardized condition. The same metrics, under different conditions, show markedly reduced reliability. A study on the reliability of pulse wave separation analysis provides direct evidence: the same group of healthy adults was measured under six different morning conditions, combining supine/sitting postures and fasting/non-fasting states. The backward pressure component (Pb) and reflection amplitude ICC met the “supine + fasting” condition (0.76 and 0.77, exceeding the 0.75 acceptable threshold), but dropped to as low as 0.56–0.57 under sitting conditions (PMID 27861248). A sister study similarly showed that the daily reliability of central systolic pressure is significantly affected by posture as a main effect, while the augmentation index (AIx) is significantly affected by fasting status as a main effect (PMID 26136065). In other words, “metrics are repeatable only under conditional constraints”—the most stable conditions are supine and fasting; once conditions relax, readings for the same individual drift.

What does this mean for the objective operationalization of three-dosha pulse diagnosis? If we directly feed the “full feature set” into a three-dosha classifier, and the training data is drawn from mixed conditions (some mornings, some afternoons, some supine, some sitting), the model is likely to systematically mislearn “acquisition conditions” as “constitutional differences”—because the influence of conditions on harmonic morphology (e.g., posture changes venous return and peripheral resistance, thereby altering wave reflection structure) can completely mimic inter-group differences. This is not pure hypothesis: the inter-rater reliability of manual three-dosha pulse diagnosis itself is only poor-to-moderate (15 Ayurvedic physicians blindly examined 20 volunteers for a total of 600 times; weighted κ, according to the Landis-Koch scale, ranges from low to moderate, with experienced practitioners showing higher κ, PMID 28664059), and inter-rater consistency for prakriti constitutional typology is also low (PMID 24191170). The labels themselves are noisy; layering on condition noise makes the classifier’s “accuracy” almost uninterpretable.

Therefore, this paper proposes a pre-screening step prior to any classification modeling: robust feature subset selection under condition constraints. The design proceeds in four steps:

Step 1, Condition Matrix Acquisition: Each subject undergoes repeated pulse wave acquisition under at least a multi-condition combination of 2 time periods (e.g., morning/evening) × 2 postures (supine/sitting) × 2 rest durations (5/15 minutes), generating multiple harmonic spectra per individual.

Step 2, Cross-Condition ICC Ranking for Each Feature: For each harmonic feature (amplitude of each order, amplitude ratios, phase, etc.), calculate the cross-condition ICC with “condition” as the repeated factor, ranking features from “most robust” to “most fragile.” Expected morphology: baseline amplitude-type features are sensitive to posture (corresponding to Pb performance in 27861248), while certain amplitude ratios or normalized features may naturally offset condition effects.

Step 3, Leave-One-Feature-Out AUC Decay Test: Using a pre-registered reference classifier, sequentially remove individual features and observe the drop in cross-condition validation AUC. Features causing a large drop are “discriminative but potentially condition-coupled”; those causing little or no drop (even when removed) are redundant or noisy.

Step 4, Robust Subset vs. Full Feature Set Head-to-Head: Retain only features with “high ICC + non-negative AUC contribution” to form the robust subset, and compare it against the full feature set under cross-condition generalization testing. The falsifiable prediction is: the full feature set shows higher AUC under same-condition cross-validation but collapses under cross-condition testing; the robust subset shows slightly lower absolute AUC but significantly smaller attenuation under cross-condition testing. If the opposite occurs—i.e., the robust subset still collapses under cross-condition testing—this indicates that the three-dosha signal itself is unstable, and this route requires re-evaluation.

The value of this workflow lies not in marginally improving accuracy, but in decoupling “condition effects” from “constitutional effects.” Features that drift with tension, posture changes, or insufficient rest are contaminants in any “constitution-health” association study; filtering them out first ensures that the remaining features have a legitimate claim to correspond with constitution.

Evidence and Limitations

Verified: Pulse harmonic analysis demonstrates good inter-observer reliability under standardized single conditions (ICC 0.83–0.96 for the first 11 harmonics, PMID 27170904). Reliability of pulse waveform-derived indices is significantly influenced by posture and fasting status; the supine + fasting condition yields the most stable results, while the seated condition shows ICC dropping as low as 0.56–0.57 (PMID 27861248, 2613065).

Partially Verified: Inter-observer reliability for manual assessment of Ayurvedic Tridosha diagnosis ranges from poor to moderate and correlates positively with experience (PMID 28664059); however, consistency among raters for prakriti (traditional consensus) classification is low (PMID 24191170). High noise in manual labeling represents a practical constraint for objective approaches.

Pending Verification: ① Specific morphology of cross-condition ICC stratification on harmonic features (which orders and which ratio characteristics are most stable); ② comparison of cross-condition AUC decay between robust subsets versus the full feature set; ③ proportion of classification accuracy contamination attributable to “condition fingerprints.” This protocol has not yet been implemented on any Tridosha cohort.

Related Reading

L1 Network Status: This topic has not yet entered the L1 corresponding network (EDGE_TABLE yields zero hits for “cross-condition/robust condition/feature subset/leave-one-feature-out”). The closest adjacent content: CORRESPONDENCE_NETWORK E-061 Covariate List “Sampling Moment (circadian) into E-061 Covariate List; Natural Breathing Segment (RR 12–20 cpm) itself drifts with time segment”; KB-510 “Resting ICC 0.50–0.83, CV 5%–41%; frequency-domain metrics less reliable than time-domain”.

Evidence Level: C
References: PMID 27170904, PMID 27861248, PMID 26136065, PMID 28664059, PMID 24191170

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 *