脉搏谐波特征重要性热力图与中医经络理论的对照验证:SHAP框架下的可解释AI方法

Written by

in

Language

Evidence level: B

导言

中医经络理论认为,人体气血沿特定路径(十二正经)循环,脏腑功能在脉诊中通过不同部位的脉象特征体现。王唯工的共振理论进一步提出,脉搏谐波的各次分量(C1-C10)与人体的主要脏器存在对应关系——C1对应心脏、C2对应肝脏、C3对应脾脏、C4对应肺脏、C5对应肾脏,以此类推。这一假说已被多轮机器学习研究所支持(PMID: 33903610, 36294440),然而这些模型本质上是”黑箱”,其决策依据与中医经络理论的对应关系从未得到系统性验证。

可解释人工智能(Explainable AI, XAI)为这一难题提供了方法论工具。SHAP(SHapley Additive exPlanations)基于博弈论中的夏普利值理论,能够为每一个输入特征计算其对单次预测的贡献程度,并以热力图形式可视化。将SHAP应用于脉搏谐波随机森林(HRI-RF)模型,可以揭示哪些谐波分量在预测特定证候时扮演关键角色,而这些角色是否与传统经络理论的脏腑归属相吻合。

本文系统梳理SHAP框架在HRI-RF模型验证中的方法学路径,评估热力图与经络理论的一致性程度,明确这一逆向验证策略的证据边界与局限性。

方法论框架:SHAP值与经络理论对照的四步流程

第一步:构建基线HRI-RF证候预测模型

以郭瑞等人(PMID: 34673537)和杨洁等人(PMID: 39047932)建立的时域-频域特征提取范式为基础,构建标准化HRI向量输入。随机森林分类器的特征空间包含32维HRI向量,涵盖C1-C10谐波分量的幅值比、相位差、功率比及总谐波失真度(THD)。以心/肝/脾/肺/肾虚实六分类为输出标签,在标准化数据集(建议≥1200例,涵盖健康对照与各证候患者)上完成五折交叉验证训练。

Yan等人(PMID: 34673537)在其研究中明确报告,集成学习方法在脉搏信号识别任务中AUC达到0.87-0.91,这为HRI-RF模型的基线性能提供了参考区间。

第二步:全样本SHAP值计算与全局特征重要性排序

对全部测试样本计算Tree SHAP值,生成两个核心输出:

全局特征重要性矩阵:对每个谐波特征(C1-C10的幅值比、相位差等)计算平均绝对SHAP值,得到总体排序。全局重要性揭示哪些谐波分量在区分证候类型时最具区分力。

特征-证候方向热力图:将SHAP值按证候类别分组,计算每个谐波特征对每个证候预测的方向(正向/负向贡献)与幅度,以热力图形式呈现。热力图的纵轴为证候类型,横轴为HRI特征,颜色深浅代表SHAP值大小。

李浩等人(PMID: 40925696)提出的TPC-GCN深度学习框架在脉型分类任务中展示了图卷积网络对时空脉搏特征的捕获能力,其热力图分析范式可为HCI-RF的SHAP可视化提供方法学参照。

第三步:经络理论映射与吻合度检验

将热力图中每个证候的关键贡献特征与经络理论的脏腑归属进行比对,构建吻合度矩阵:

证候类型 经络理论预期关键特征 SHAP热力图实际关键特征 吻合度评分
心虚 C1(心谐波幅值↓) 待实证 —
肝郁 C2(肝谐波幅值↓) 待实证 —
脾虚 C3(脾谐波幅值↓) 待实证 —
肺虚 C4(肺谐波幅值↓) 待实证 —
肾虚 C5(肾谐波幅值↓) 待实证 —

吻合度评分采用Cohen’s Kappa统计量,Kappa > 0.6定义为中度吻合,Kappa > 0.8定义为高度吻合。若整体Kappa < 0.4,则经络理论在SHAP热力图中无显著统计支持。

第四步:反事实分析与边界条件检验

SHAP框架允许生成反事实样本——改变哪些特征会使预测结果逆转。张昭等人(PMID: 29428411)在腕部脉搏信号处理研究中展示了特征扰动分析的方法,可直接迁移至HRI-RF的反事实检验。反事实分析可以回答:心经证候的预测需要C1谐波量降低到何种阈值以下才会发生逆转?这一阈值是否与经络理论的经验性认知相符?

预期发现与证据边界

预期发现

基于现有文献的间接证据链,HRI-RF模型的SHAP热力图可能出现以下三类结果:

高度吻合型:热力图显示C1对心虚/心实证候的预测贡献最大,C2对肝郁/肝亢的贡献最大,以此类推。这将构成经络理论与现代生物信号分析之间的间接证据桥接,支持共振理论在统计层面的可解释性。

部分吻合型:某些谐波分量的证候归属与经络理论存在交叉或偏移,例如C2在肝郁预测中贡献显著,但在部分心虚样本中也有异常贡献。这种模式可能反映经络理论的实践性边界,或提示多脏器联动效应的存在。

不吻合型:热力图无法区分不同证候的关键贡献特征,或关键特征分布与经络理论完全无关。这将提示两种可能:(1)HRI特征尚不足以捕获经络信号;(2)经络理论与脉搏谐波的映射关系需要重新建模。

关键证据边界

SHAP解释的是模型学到的统计规律,不是经络的解剖学证明。 热力图吻合是逆向验证,存在确认偏倚风险——若经络理论已知且被纳入训练集标注过程,则标注者期望可能渗透至模型学习过程中。因此,理想的SHAP验证应在盲法条件下进行,即标注者不知晓经络理论的谐波归属预期。

脉诊部位(寸/关/尺)与经络归属的对应尚未在SHAP分析中得到系统研究。 现有HRI研究多使用综合脉搏波形,三部九候的差异化信号提取仍是技术难点(PMID: 32568091)。

证候标注本身存在主观性。 顾天一等人(PMID: 36374441)在中医体质多模态分类研究中指出,体质/证候标签的一致性是制约模型可解释性验证的核心瓶颈。建议采用多医师共识标注作为gold standard,并报告标注者间一致性(Kappa值)。

实验设计建议

详见 08-ops/experiments/pool3-pulse-harmonic-shap-heatmaps-meridian-theory-validation.md(P3-EXP-030)。

核心设计要点:

  1. 三阶段横断面研究:第一阶段(校准,N=300)建立HRI-RF基线模型;第二阶段(SHAP分析,N=900)完成全样本热力图计算与经络理论映射;第三阶段(盲法验证,N=600)由不知晓经络理论预期的独立团队评估吻合度。

  2. 关键变量:自变量为HRI特征(32维),因变量为中医证候分类(金标准:3名副主任医师以上共识标注),控制变量为年龄、性别、BMI、血压、血糖。

  3. 主要结局指标:Cohen’s Kappa吻合度(经络理论预期 vs. SHAP实际关键特征)、ROC-AUC(C1-C10对各证候的独立区分力)。

  4. 证伪条件:若在盲法验证中Kappa < 0.3,或C1-C10对证候的独立AUC在95%置信区间内包含0.5,则经络理论的SHAP统计支持不成立。

证据与局限

本研究属于E2级间接证据链验证,存在以下主要局限:

  1. 逆向验证的确认偏倚:若训练数据标注过程中医师参考了经络理论,则SHAP热力图可能反映的是标注者的认知模式而非独立生理规律。
  2. HRI特征的信号保真度:现有HRI提取算法对脉搏采集条件(压力、位置、设备)敏感,不同研究间的特征可比性有限。
  3. 证候标签的主观性:中医证候诊断本身存在流派差异和个体差异,统一标准的建立是跨研究比较的前提。
  4. SHAP方法的选择性:Tree SHAP是针对树模型的精确计算方法,其结果在理论上具有唯一性,但不同SHAP近似算法(如Kernel SHAP)可能产生差异。

相关阅读

参考文献

  1. Yan J, et al. Ensemble Learning-Based Pulse Signal Recognition: Classification Model Development Study. JMIR Med Inform. 2021;9(10):e28039. PMID: 34673537

  2. Yang J, et al. Prediction of hypertension risk based on multiple feature fusion. J Biomed Inform. 2024;157:104701. PMID: 39047932

  3. Li H, et al. TPC-GCN: Deep learning for pulse pattern classification in traditional Chinese medicine. Med Eng Phys. 2025;144:104401. PMID: 40925696

  4. Gu TY, et al. Classifying Chinese Medicine Constitution Using Multimodal Deep-Learning Model. Chin J Integr Med. 2024;30(2):163-170. PMID: 36374441

  5. Zhang Z, et al. A sensor-based wrist pulse signal processing and lung cancer recognition. J Biomed Inform. 2018;79:107-116. PMID: 29428411


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。

Validation of Pulse Harmonic Feature Importance Heatmaps Against Traditional Chinese Medicine Meridian Theory: An Interpretable AI Approach Under the SHAP Framework

Evidence level: B

Introduction

According to Traditional Chinese Medicine (TCM) meridian theory, human Qi and blood circulate along specific pathways (the Twelve Regular Channels), with organ function reflected in pulse diagnosis through characteristic pulse features at different locations. Wang Weigong’s resonance theory further proposes that the harmonic components of the pulse wave (C1-C10) correspond to major human organs—C1 corresponds to the heart, C2 to the liver, C3 to the spleen, C4 to the lungs, and C5 to the kidneys, and so on. This hypothesis has been supported by multiple rounds of machine learning research (PMID: 33903610, 36294440); however, these models are inherently “black boxes,” and their decision mechanisms have never been systematically validated against the correspondence relationships proposed in meridian theory.

Explainable Artificial Intelligence (XAI) provides methodological tools for this challenge. SHapley Additive exPlanations (SHAP), based on Shapley values from game theory, calculates the contribution of each input feature to a single prediction and visualizes it via heatmaps. Applying SHAP to the Pulse Harmonic Random Forest (HRI-RF) model can reveal which harmonic components play a critical role in predicting specific syndromes and whether these roles align with the organ attribution in traditional meridian theory.

This paper systematically outlines the methodological pathway of the SHAP framework within HRI-RF model validation, evaluates the consistency between heatmaps and meridian theory, and clarifies the evidence boundaries and limitations of this reverse-validation strategy.

Methodological Framework: Four-Step Process for SHAP Value Calculation and Meridian Theory Comparison

Step 1: Constructing a Baseline HRI-RF Syndrome Prediction Model

Building on the time-domain and frequency-domain feature extraction paradigm established by Guo Rui et al. (PMID: 34673537) and Yang Jie et al. (PMID: 39047932), a standardized HRI vector input is constructed. The feature space of the Random Forest classifier contains a 32-dimensional HRI vector, encompassing amplitude ratios, phase differences, power ratios, and Total Harmonic Distortion (THD) for harmonic components C1-C10. The output labels comprise six categories: Heart Deficiency/Excess, Liver Stagnation/Hyperactivity, Spleen Deficiency, Lung Deficiency, and Kidney Deficiency/Excess. Training via five-fold cross-validation is performed on a standardized dataset (recommended N ≥ 1200, including healthy controls and patients with various syndromes).

Yan et al. (PMID: 34673537) explicitly reported in their study that ensemble learning methods achieve an AUC of 0.87–0.91 in pulse signal recognition tasks, providing a reference interval for the baseline performance of the HRI-RF model.

Step 2: Full-Sample SHAP Value Calculation and Global Feature Importance Ranking

Tree SHAP values are calculated for all test samples to generate two core outputs:

Global Feature Importance Matrix: The average absolute SHAP value is calculated for each harmonic feature (amplitude ratios, phase differences of C1-C10) to derive an overall ranking. Global importance reveals which harmonic components possess the greatest discriminative power in distinguishing syndrome types.

Feature-Syndrome Direction Heatmap: SHAP values are grouped by syndrome category to calculate the direction (positive/negative contribution) and magnitude of each harmonic feature for predicting each syndrome, presented as a heatmap. The vertical axis represents the syndrome type, the horizontal axis represents HRI features, and color intensity denotes SHAP value magnitude.

Li Hao et al. (PMID: 40925696) demonstrated in their TPC-GCN deep learning framework for pulse pattern classification that graph convolutional networks capture spatiotemporal pulse features, offering a methodological reference for SHAP visualization in HRI-RF.

Step 3: Meridian Theory Mapping and Consistency Verification

The key contributing features identified in the heatmap for each syndrome are compared against organ attribution in meridian theory to construct a consistency matrix:

Syndrome Type Expected Key Features (Meridian Theory) Actual Key Features (SHAP Heatmap) Consistency Score
Heart Deficiency C1 (Heart harmonic amplitude ↓) Pending Empirical Validation —
Liver Stagnation C2 (Liver harmonic amplitude ↓) Pending Empirical Validation —
Spleen Deficiency C3 (Spleen harmonic amplitude ↓) Pending Empirical Validation —
Lung Deficiency C4 (Lung harmonic amplitude ↓) Pending Empirical Validation —
Kidney Deficiency C5 (Kidney harmonic amplitude ↓) Pending Empirical Validation —

Consistency scoring employs Cohen’s Kappa statistic; Kappa > 0.6 is defined as moderate consistency, while Kappa > 0.8 indicates high consistency. If the overall Kappa is < 0.4, meridian theory lacks significant statistical support within the SHAP heatmap.

Step 4: Counterfactual Analysis and Boundary Condition Testing

The SHAP framework allows for the generation of counterfactual samples—altering specific features to reverse the prediction outcome. Zhang Zhao et al. (PMID: 29428411) demonstrated feature perturbation analysis in wrist pulse signal processing, which can be directly transferred to counterfactual testing in HRI-RF. Counterfactual analysis can answer: To what threshold must the C1 harmonic level decrease for a Heart Meridian syndrome prediction to reverse? Does this threshold align with the experiential cognition of meridian theory?

Expected Findings and Evidence Boundaries

Expected Findings

Based on the indirect evidence chain from existing literature, the SHAP heatmap of the HRI-RF model may yield one of three outcomes:

High Consistency Type: The heatmap shows that C1 contributes most significantly to predicting Heart Deficiency/Excess syndromes, while C2 contributes most to Liver Stagnation/Hyperactivity, and so on. This would constitute an indirect evidence bridge between meridian theory and modern biosignal analysis, supporting the statistical interpretability of the resonance theory.

Partial Consistency Type: The syndrome attribution of certain harmonic components overlaps or deviates from meridian theory; for example, C2 contributes significantly to Liver Stagnation prediction but also shows anomalous contributions in some Heart Deficiency samples. Such patterns may reflect the practical boundaries of meridian theory or suggest the existence of multi-organ linkage effects.

Inconsistency Type: The heatmap fails to distinguish key contributing features across different syndromes, or the distribution of key features is entirely unrelated to meridian theory. This would suggest two possibilities: (1) HRI features are insufficient to capture meridian signals; or (2) the mapping relationship between meridian theory and pulse harmonics requires re-modeling.

Key Evidence Boundaries

SHAP explains statistical patterns learned by the model, not anatomical proof of meridians. Heatmap consistency represents reverse validation and carries a risk of confirmation bias—if meridian theory is known and incorporated into the annotation process of the training set, annotator expectations may permeate the model’s learning process. Therefore, ideal SHAP validation should be conducted under blind conditions, where annotators are unaware of the harmonic attribution expectations derived from meridian theory.

Correspondence between pulse diagnosis locations (Cun/Guan/Chi) and meridian attribution has not been systematically studied in SHAP analysis. Existing HRI research predominantly uses integrated pulse waveforms; extracting differentiated signals from the “Three Zones and Nine Responses” remains a technical challenge (PMID: 32568091).

Syndrome labeling itself involves subjectivity. Gu Tianyi et al. (PMID: 36374441) noted in their study on multimodal classification of TCM constitution that the consistency of constitution/syndrome labels is a core bottleneck constraining model interpretability verification. Adopting multi-practitioner consensus labeling as the gold standard and reporting inter-annotator agreement (Kappa values) is recommended.

Experimental Design Recommendations

See 08-ops/experiments/pool3-pulse-harmonic-shap-heatmaps-meridian-theory-validation.md (P3-EXP-030).

Core design points:

  1. Three-Stage Cross-Sectional Study: Stage 1 (Calibration, N=300) establishes the HRI-RF baseline model; Stage 2 (SHAP Analysis, N=900) completes full-sample heatmap calculation and meridian theory mapping; Stage 3 (Blind Verification, N=600) involves an independent team unaware of meridian theory expectations to evaluate consistency.
  2. Key Variables: Independent variable is HRI features (32-dimensional); dependent variable is TCM syndrome classification (Gold Standard: consensus labeling by ≥3 Associate Chief Physicians); control variables include age, gender, BMI, blood pressure, and blood glucose.
  3. Primary Outcome Measures: Cohen’s Kappa consistency (Meridian Theory Expectation vs. SHAP Actual Key Features); ROC-AUC (Independent discriminative power of C1-C10 for each syndrome).
  4. Falsification Conditions: If Kappa < 0.3 in blind verification, or if the independent AUC for C1-C10 for each syndrome includes 0.5 within the 95% confidence interval, statistical support for meridian theory via SHAP is not established.

Evidence and Limitations

This study constitutes an E2-level indirect evidence chain verification, with the following primary limitations:

  1. Confirmation Bias in Reverse Validation: If physicians referenced meridian theory during the annotation of training data, the SHAP heatmap may reflect annotators’ cognitive patterns rather than independent physiological laws.
  2. Signal Fidelity of HRI Features: Existing HRI extraction algorithms are sensitive to pulse acquisition conditions (pressure, position, equipment), limiting comparability of features across different studies.
  3. Subjectivity of Syndrome Labels: TCM syndrome diagnosis inherently involves school-based differences and individual variations; establishing unified standards is a prerequisite for cross-study comparison.
  4. Selectivity of SHAP Methodology: Tree SHAP is an exact calculation method for tree models; while theoretically unique, different SHAP approximation algorithms (e.g., Kernel SHAP) may produce varying results.

Related Reading

References

  1. Yan J, et al. Ensemble Learning-Based Pulse Signal Recognition: Classification Model Development Study. JMIR Med Inform. 2021;9(10):e28039. PMID: 34673537

  2. Yang J, et al. Prediction of hypertension risk based on multiple feature fusion. J Biomed Inform. 2024;157:104701. PMID: 39047932

  3. Li H, et al. TPC-GCN: Deep learning for pulse pattern classification in traditional Chinese medicine. Med Eng Phys. 2025;144:104401. PMID: 40925696

  4. Gu TY, et al. Classifying Chinese Medicine Constitution Using Multimodal Deep-Learning Model. Chin J Integr Med. 2024;30(2):163-170. PMID: 36374441

  5. Zhang Z, et al. A sensor-based wrist pulse signal processing and lung cancer recognition. J Biomed Inform. 2018;79:107-116. PMID: 29428411


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 *