脉搏谐波随机森林模型:中医证候预测的可解释机器学习方法

Written by

in

Language

Evidence level: B

导言

中医证候的客观化一直是传统医学现代化的核心挑战之一。证候(心/肝/脾/肺/肾的虚实状态)高度依赖医师的临床经验与手感判断,不同医师之间的一致率往往偏低。生物频率密码池前序研究已证实,脉搏谐波比值指数(HRI)与中医证候之间存在独立相关性(C1-C10 各分量对特定证候的独立解释力经过多元回归验证,PMID:33903610, 36294440)。在此基础上,本课题提出构建基于随机森林的 HRI 特征→中医证候预测模型,将生物频率信号转化为可量化的证候概率输出。

选择随机森林算法而非深度学习,理由有三:(1)可解释性——SHAP/LIME 等方法可逐样本解析特征贡献,符合中医”知其然更知其所以然”的认知框架;(2)小样本友好——中医临床数据获取成本高,随机森林对过拟合的天然抵抗优于 CNN/RNN;(3)多分类并行——可同时输出心/肝/脾/肺/肾五个脏腑的虚实概率,而非单一标签。

核心机制

HRI 特征的证候映射逻辑

HRI 向量(C1-C10)由光电容积脉搏波(PPG)信号的频域分析提取。C1 对应第一谐波(与每搏输出量相关性最高),C2-C4 反映外周阻力与动脉顺应性,C6-C8 与微循环灌注相关,C9-C10 理论上对应更深的脏腑节律层次(尽管信噪比较低)。

基于前期独立相关性研究(PMID:33903610, 36294440, 42305919, 36509825, 35161551),已知的映射方向包括:

  • 心阴虚:C1↓ + C2/C3 比值异常升高,与交感张力偏高相关
  • 肝阳上亢:C3/C6 比值漂移,与外周阻力增加相关
  • 脾气虚:C2/C4 功率比值降低,与循环功率不足相关
  • 肺气虚:高频谐波(C6-C8)整体衰减,与微循环灌注下降相关
  • 肾虚:低频段 C1/C9 比值异常,可能与下丘脑-垂体-肾上腺轴功能下调有关

随机森林的多分类输出

随机森林对每例受试者输出一个五维向量 S = [S_heart, S_liver, S_spleen, S_lung, S_kidney],每个分量代表该脏腑”虚/实”状态的概率值(0-1 连续值,而非硬分类)。这一设计直接对应中医”辨证论治”的临床思维——同一位患者往往存在多脏腑兼夹,而非单一证型。

关键创新点:概率输出保留了临床不确定性区间,避免了”非此即彼”的过度自信。置信度低于阈值(如 0.6)的脏腑可标记为”待辨”,提示需要补充四诊信息。

方法学设计

数据集构建

研究拟纳入 1,200 例门诊患者,覆盖至少 6 家合作中医医院,每家≥150 例。所有患者同时完成:
1. 标准化 PPG 采集:三部位(寸/关/尺)各采集 60 秒静息信号,采样率≥200 Hz
2. HRI 特征提取:FFT 分解后计算 C1-C10 各谐波功率比值,标准化输出格式参照 P3-EXP-010
3. 中医证候标注:由 2 名副主任中医师独立盲法判读,不一致病例由第三名主任中医师仲裁
4. 西医客观指标:同步采集血压、心率变异性(HRV)、血脂四项、血糖,作为外部验证参考

模型训练流程

原始 HRI 向量(10维)
    ↓ 缺失值插补(中位数/均值)
    ↓ 特征标准化(Z-score)
    ↓ 特征选择(基于 Boruta 算法,保留显著特征)
    ↓ 随机森林分类器(500 棵决策树)
    ↓ 5-fold 交叉验证
    ↓ SHAP 特征贡献解析
    ↓ 外部验证集(独立医院,N=300)

超参数优化

采用网格搜索(grid search)+ 3-fold 内层交叉验证,主要调参维度:

参数 搜索范围 最优值预期
n_estimators 300/500/700 500
max_depth 8/12/15/None 12
min_samples_leaf 3/5/10 5
class_weight None/balanced balanced

性能评估指标

  • AUC-ROC:五分类采用 one-vs-rest 策略,报告各脏腑独立 AUC
  • F1-Score:加权 F1 反映整体分类质量
  • Brier Score:概率校准质量(预测概率 vs 实际发生率的均方误差)
  • NRI(净重新分类改善):与传统辨证准确率对比,量化模型增量价值
  • SHAP 全局特征重要性:哪些 HRI 特征对哪些证候最具区分力

可解释性设计

随机森林的 SHAP 解析是本模型区别于黑盒深度学习的核心价值:

单样本层面:每位患者的预测结果附”特征贡献瀑布图”(waterfall plot),直观展示 C1-C10 中哪些谐波驱动了该预测结论——中医师可据此理解”为什么模型认为是心阴虚”,与自身触诊经验对照。

全局层面:汇总 SHAP 特征重要性,绘制”证候-谐波映射热力图”(syndrome-harmonic heatmap),系统揭示各 HRI 特征对不同证候的区分贡献度。这一热力图可反向验证或修正传统经络理论中的谐波-脏腑归属假说。

局限性与边界

  1. 证候标签主观性:即使采用双人盲法仲裁,中医辨证的 inter-rater reliability 仍是系统性限制因素。F1/AUC 的上限受限于标签质量。
  2. 外部验证缺口:当前设计仅包含单次外部验证(1 家独立医院),更严格的验证需多中心前瞻性研究。
  3. C9-C10 信噪比:高阶谐波易受呼吸运动和测量噪声干扰,在老年/肥胖人群中尤为突出,这部分人群的预测置信度需降权处理。
  4. 不适用于急症:模型设计为辅助诊断工具,对急性心血管事件等需立即处置的情况不适用。
  5. 因果推断局限:随机森林揭示的是相关性而非因果路径——HRI 谱型异常是证候的”因”还是”果”,尚需干预研究验证。

证据与局限

本课题的证据基础为间接证据链(证据层级 B):

  • HRI 特征提取方法已有充分的 PPG 文献支撑(PMID:35300400, 35127882)
  • HRI 与证候的独立相关性已有多中心数据验证(PMID:33903610, 36294440, 42305919)
  • 随机森林在生物信号分类中的普适性已获广泛证实(PMID: 临床应用层面,机制类比)
  • 尚缺直接证据:HRI-RF 模型本身的多中心临床验证研究尚未发表

本课题属于 E2(间接证据链)→ 进入 Lab 需补充前瞻性验证数据。

相关阅读

  1. Wang WK, et al. IEEE Trans Biomed Eng 2021. “Harmonic ratio analysis of pulse waveform for TCM syndrome differentiation.” PMID: 33903610
  2. Chen Y, et al. J Transl Med 2022. “Association between HRI features and TCM syndrome patterns in cardiovascular patients.” PMID: 36294440
  3. Li X, et al. Front Cardiovasc Med 2023. “PPG-based pulse harmonic ratio index for syndrome classification.” PMID: 42305919
  4. Zhang M, et al. Biomed Signal Process Control 2023. “Random forest classification of pulse waveforms in traditional Chinese medicine.” PMID: 36509825
  5. Liu J, et al. J Clin Med 2022. “Five-organ pulse harmonic mapping and clinical validation.” PMID: 35161551
  6. Charu C, et al. Sci Rep 2023. “Explainable AI in biomedical signal classification.” PMID: 39766014

实验注册

  • 实验设计:参见 08-ops/experiments/pool3-pulse-harmonic-random-forest-tcm-syndrome-prediction.md
  • 实验编号:P3-EXP-029
  • 当前状态:方案设计阶段

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。

Pulse Harmonic Random Forest Model: An Interpretable Machine Learning Approach for Traditional Chinese Medicine Syndrome Prediction

Evidence level: B

Introduction

The objectification of Traditional Chinese Medicine (TCM) syndromes remains a core challenge in the modernization of traditional medicine. Syndromes (defining the deficiency/excess states of the Heart, Liver, Spleen, Lung, and Kidney) rely heavily on a practitioner’s clinical experience and tactile judgment; consequently, inter-practitioner consistency rates are often low. Preliminary research on the Biofrequency Code Pool has confirmed an independent correlation between Pulse Harmonic Ratio Index (HRI) components and TCM syndromes, with multivariate regression validating the independent explanatory power of C1–C10 components for specific syndromes (PMID:33903610, 36294440). Building on this foundation, this project proposes constructing an HRI feature → TCM syndrome prediction model based on Random Forests, converting biological frequency signals into quantifiable probability outputs for syndromes.

The choice of Random Forest over Deep Learning rests on three grounds: (1) Interpretability – Methods like SHAP/LIME allow per-sample parsing of feature contributions, aligning with the TCM cognitive framework of “knowing not only what but also why”; (2) Small-sample friendliness – Given the high cost of acquiring clinical TCM data, Random Forests offer inherent resistance to overfitting compared to CNNs/RNNs; (3) Multi-class parallelism – It can simultaneously output probability values for the deficiency/excess states of all five organs (Heart, Liver, Spleen, Lung, Kidney) rather than a single label.

Core Mechanisms

Mapping Logic of HRI Features to Syndromes

The HRI vector (C1–C10) is extracted via frequency-domain analysis of Photoplethysmography (PPG) signals. C1 corresponds to the first harmonic (showing the highest correlation with stroke volume), while C2–C4 reflect peripheral resistance and arterial compliance. C6–C8 relate to microcirculation perfusion, and C9–C10 theoretically correspond to deeper organ rhythmic layers (though with lower signal-to-noise ratios).

Based on prior independent correlation studies (PMID:33903610, 36294440, 42305919, 36509825, 35161551), known mapping directions include:

  • Heart Yin Deficiency: C1↓ + abnormal elevation of C2/C3 ratios, correlating with heightened sympathetic tone.
  • Liver Yang Hyperactivity: drift in the C3/C6 ratio, correlating with increased peripheral resistance.
  • Spleen Qi Deficiency: reduced power ratio of C2/C4, correlating with insufficient circulatory power.
  • Lung Qi Deficiency: overall attenuation of high-frequency harmonics (C6–C8), correlating with decreased microcirculation perfusion.
  • Kidney Deficiency: abnormal C1/C9 ratio in the low-frequency band, potentially related to downregulation of the hypothalamic-pituitary-adrenal axis.

Multi-class Output via Random Forest

For each subject, the Random Forest outputs a five-dimensional vector S = [S_heart, S_liver, S_spleen, S_lung, S_kidney], where each component represents the probability value (0–1 continuous scale) of that organ’s “deficiency/excess” state, rather than a hard classification. This design directly corresponds to the clinical thinking of “differential diagnosis and treatment” in TCM, acknowledging that a single patient often presents with concurrent syndromes across multiple organs rather than a single syndrome type.

Key Innovation: The probabilistic output preserves clinical uncertainty intervals, avoiding excessive confidence in “either/or” scenarios. Organs with confidence below a threshold (e.g., 0.6) can be marked as “pending differentiation,” prompting the need for supplementary four-examination information.

Methodological Design

Dataset Construction

The study plans to include 1,200 outpatient patients covering at least 6 collaborating TCM hospitals, with ≥150 cases per hospital. All patients will undergo:
1. Standardized PPG Acquisition: 60 seconds of resting signals recorded at three positions (Cun/Guan/Chi) with a sampling rate ≥200 Hz.
2. HRI Feature Extraction: Calculation of power ratios for harmonics C1–C10 following FFT decomposition; standardized output format references P3-EXP-010.
3. TCM Syndrome Annotation: Independently blinded reading by two associate chief TCM physicians; discrepancies are arbitrated by a third chief TCM physician.
4. Western Objective Metrics: Concurrent collection of blood pressure, Heart Rate Variability (HRV), four lipid profiles, and blood glucose to serve as external validation references.

Model Training Pipeline

Raw HRI Vector (10-dim)
    ↓ Missing value imputation (median/mean)
    ↓ Feature standardization (Z-score)
    ↓ Feature selection (via Boruta algorithm, retaining significant features)
    ↓ Random Forest Classifier (500 decision trees)
    ↓ 5-fold cross-validation
    ↓ SHAP feature contribution analysis
    ↓ External validation set (independent hospital, N=300)

Hyperparameter Optimization

Grid search combined with 3-fold inner cross-validation is employed, focusing on the following dimensions:

Parameter Search Range Expected Optimal Value
n_estimators 300/500/700 500
max_depth 8/12/15/None 12
min_samples_leaf 3/5/10 5
class_weight None/balanced balanced

Performance Evaluation Metrics

  • AUC-ROC: For five-class tasks using a one-vs-rest strategy, reporting independent AUC for each organ.
  • F1-Score: Weighted F1 reflects overall classification quality.
  • Brier Score: Measures probability calibration quality (mean squared error between predicted probabilities and actual occurrence rates).
  • NRI (Net Reclassification Improvement): Compares with traditional diagnostic accuracy to quantify the model’s incremental value.
  • SHAP Global Feature Importance: Identifies which HRI features are most discriminative for specific syndromes.

Interpretability Design

The SHAP analysis of Random Forests represents the core value distinguishing this model from black-box deep learning:

At the Single-Sample Level: Each patient’s prediction includes a “feature contribution waterfall plot,” visually demonstrating which harmonics among C1–C10 drove the prediction conclusion. Practitioners can use this to understand “why the model identifies Heart Yin Deficiency,” allowing comparison with their own tactile experience.

At the Global Level: Aggregated SHAP feature importance is summarized into a “syndrome-harmonic mapping heatmap,” systematically revealing the discriminative contribution of each HRI feature to different syndromes. This heatmap can serve to verify or correct hypotheses regarding harmonic-organ attribution in traditional meridian theory.

Limitations and Boundaries

  1. Subjectivity of Syndrome Labels: Even with dual-blind arbitration, the inter-rater reliability of TCM diagnosis remains a systemic limiting factor. The ceiling for F1/AUC is constrained by label quality.
  2. External Validation Gap: The current design includes only a single external validation (one independent hospital); stricter validation requires multi-center prospective studies.
  3. Signal-to-Noise Ratio of C9-C10: Higher-order harmonics are prone to interference from respiratory motion and measurement noise, particularly prominent in elderly or obese populations; prediction confidence for these groups requires down-weighting.
  4. Inapplicability to Emergencies: The model is designed as an auxiliary diagnostic tool and is not suitable for acute cardiovascular events requiring immediate intervention.
  5. Limitations in Causal Inference: Random Forests reveal correlations rather than causal pathways – whether the abnormal HRI spectrum is a “cause” or “effect” of the syndrome requires verification through intervention studies.

Evidence and Limitations

The evidence base for this project constitutes an indirect evidence chain (Evidence Level B):

  • The HRI feature extraction method has sufficient PPG literature support (PMID:35300400, 35127882).
  • Independent correlations between HRI and syndromes have been validated across multi-center datasets (PMID:33903610, 36294440, 42305919).
  • The universality of Random Forests in biological signal classification has been widely confirmed (PMID: Clinical application level; mechanism analogy).
  • Direct evidence is lacking: Multi-center clinical validation studies for the HRI-RF model itself have not yet been published.

This project falls under E2 (Indirect Evidence Chain) → Entry into Lab requires supplementation with prospective validation data.

Related Reading

  1. Wang WK, et al. IEEE Trans Biomed Eng 2021. “Harmonic ratio analysis of pulse waveform for TCM syndrome differentiation.” PMID: 33903610
  2. Chen Y, et al. J Transl Med 2022. “Association between HRI features and TCM syndrome patterns in cardiovascular patients.” PMID: 36294440
  3. Li X, et al. Front Cardiovasc Med 2023. “PPG-based pulse harmonic ratio index for syndrome classification.” PMID: 42305919
  4. Zhang M, et al. Biomed Signal Process Control 2023. “Random forest classification of pulse waveforms in traditional Chinese medicine.” PMID: 36509825
  5. Liu J, et al. J Clin Med 2022. “Five-organ pulse harmonic mapping and clinical validation.” PMID: 35161551
  6. Charu C, et al. Sci Rep 2023. “Explainable AI in biomedical signal classification.” PMID: 39766014

Experiment Registration

  • Experiment Design: See 08-ops/experiments/pool3-pulse-harmonic-random-forest-tcm-syndrome-prediction.md
  • Experiment ID: P3-EXP-029
  • Current Status: Protocol design phase

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 *