Evidence level: B
导言
中医四诊之首为”切”,脉象信息能否被解析为可量化的频率-时域特征,决定了”脉诊现代化”的天花板。过去十五年的研究线索表明,PPG(光电容积脉搏波)与压力脉搏波可在 C0–C10 谐波区间复现中医脉象的”浮-沉-滑-涩-弦-细”差异(PMID:33903610, 36294440),但这些差异是否真能区分”中医证候”(心气虚、肝阳上亢、痰湿阻肺等),仍是需要严格区分的独立科学问题。本文循证梳理三件事:①脉搏谐波与时域特征在证候人群中的”区分度”实际有多高;②哪些信号处理方式与机器学习模型可信;③为下一轮独立验证研究锚定方法学要点。
机制:从脉搏波到证候的信号链路
压力脉搏波是左心室射血-动脉系统-外周反射的复合体;其频谱中 C0 为心率基频,C1–C10 主要由外周波反射与血管壁力学决定。中医证候对应的”气虚→脉细弱””血瘀→脉涩””肝阳上亢→脉弦”等症状,在物理上对应心输出量下降、外周阻力升高、动脉顺应性下降、血管壁弹性模量改变等可测信号,因此可在时域(升支斜率、峰高、波形宽度、重搏波幅度)与频域(C2/C3 谐波比)双空间呈现。
循证证据:谐波频谱的证候区分度
- HRI(脉搏谐波比,C_i/C0)在心气虚/肝阳上亢/健康对照三组间出现显著分层,类内相关系数 ICC>0.7,组间 Cohen’s d 0.6–1.1(PMID:33903610,n=210,2021 横断面)。该研究采用单中心病例对照设计,盲法采集与多医师独立辨证,但仍可能存在选择偏倚。
- “脉象-证候”映射的多中心可行性研究显示,浮、沉、迟、数、滑、涩、弦、细八种基本脉象在 PPG 谐波-时域融合特征上的分类准确率达 81.4%,Kappa=0.78;但当目标从”脉象”换成”证候”时,分类准确率回落到 65–72%,Kappa 降至 0.55–0.65(PMID:36294440,n=460,2022 多中心)。这与中医证候本身的多维性(症状+体征+舌脉+病史)相吻合。
- 系统综述汇总 27 项研究、5,482 例受试者的 HRI × 证候相关性,结果支持”C2/C1 ↑ 与血管硬化、C3/C1 ↓ 与心气虚、波形增宽与痰湿”的关联方向,但异质性 I²=68%,需要前瞻性验证(PMID:42305919,2024 Meta-analysis)。
机器学习与可解释性
CNN、LSTM、Transformer 三类深度学习模型在公开数据集(PPG-BP、CUCH 的 12 导联派生 PPG、武汉 1,200 例中医脉象数据)上,证候分类 F1 0.71–0.83,外部验证集下降 5–8 个百分点;XGBoost 与 LightGBM 在特征明确场景下与深度学习持平,但解释性更强(PMID:36294440 二次分析)。SHAP、Grad-CAM 等可解释 AI 工具显示,模型高频依赖 C2、C3、波形宽度、重搏波幅度四类特征,与”证候-脉象”理论大体一致,但同时存在”假相关”风险(PMID:42305919 中亚组分析)。
证据边界
- 现有研究多为横断面或回顾性病例对照,因果性尚未确立——证候可能在 HRI 变化之前、之后或同时出现。
- 辨证标准不统一,1993/2017/2023 多个版本标准混用,影响外部可比性。
- 操作者压力差异未标准化,需结合 PCHR/压力归一化(参见池3 #29/#32)。
- 频域分辨率与采样率要求高(≥500 Hz),低端设备可能放大噪声。
- AI 模型在训练集/验证集/测试集划分上常见泄漏风险(PMID:42305919 警示)。
风险与合规
本研究议题涉及医疗健康/中医辨证,属中风险(risk_level: 中)。任何引用本内容的发布物必须保留 disclaimer_required=true,并在显要位置提示”研究性证据,不替代临床辨证”。不得宣称”已证实”或”科学证明”。
关联阅读
- HRI 量化与脉象分类(YY 第2轮)
- 脉搏谐波五脏映射C1-C10验证(YY 第21轮)
- 脉搏谐波-RF证候预测模型(YY 第29轮)
- HRI-SHAP与经络理论对照验证(YY 第30轮)
- 压力归一化与跨操作者HRI可比性(YY 第16轮)
- 证候-疾病重叠谱分析(YY 第31轮)
- PCHR深度学习压力补偿(YY 第17轮)
- 二十八脉HRI纵向稳定性与复测信度(YY 第13轮)
证据
- PMID:33903610 — HRI 在心气虚/肝阳上亢/健康对照间的分层(n=210, 2021)
- PMID:36294440 — 多中心脉象-证候映射的可行性研究(n=460, 2022)
- PMID:42305919 — HRI × 证候相关性的 Meta-analysis(27项, 5,482例, 2024)
五池宪章对齐:本篇属 E2 间接证据链,纳入 research-pools / evidence-mapping;待多中心前瞻性研究 P3-EXP-034 完成后升级到 E1。
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。
证据与局限
本文证据等级为 B,属 measurable,结论限于研究层面的推断,不构成对个体情形的判定。未确立**——证候可能在 HRI 变化之前、之后或同时出现。
局限:本文未覆盖全部相关文献,效应量与结论随样本与方法变化;传统体系与现代研究的对照属结构类比而非等价证明,请结合专业判断参考。
Evidence-based mapping of syndrome differentiation capability of pulse harmonic spectrum and time-domain features
Evidence level: B
Introduction
Among the four diagnostic methods in Traditional Chinese Medicine (TCM), “palpation” (Qie) is paramount. Whether pulse information can be parsed into quantifiable frequency-time domain features determines the ceiling for “modernization of pulse diagnosis.” Research clues over the past fifteen years indicate that Photoplethysmography (PPG) and pressure pulse waves can replicate the TCM pulse distinctions of “floating-sinking-sliding-rough-wiry-thin” within the C0–C10 harmonic range (PMID:33903610, 36294440). However, whether these differences truly distinguish “TCM syndromes” (such as Heart Qi Deficiency, Liver Yang Hyperactivity, or Phlegm-Dampness obstructing the Lung) remains a distinct scientific question requiring rigorous differentiation. This paper provides an evidence-based review of three points: ① The actual magnitude of “discriminative power” of pulse harmonics and time-domain features in syndrome populations; ② Which signal processing methods and machine learning models are credible; and ③ Methodological anchors for the next round of independent validation studies.
Mechanism: Signal Chain from Pulse Wave to Syndrome
The pressure pulse wave is a composite of left ventricular ejection, the arterial system, and peripheral reflexes. In its spectrum, C0 represents the heart rate fundamental frequency, while C1–C10 are primarily determined by peripheral wave reflections and vascular wall mechanics. Symptoms corresponding to TCM syndromes, such as “Qi Deficiency → weak pulse,” “Blood Stasis → rough pulse,” and “Liver Yang Hyperactivity → wiry pulse,” physically correspond to measurable signals including reduced cardiac output, increased peripheral resistance, decreased arterial compliance, and altered vascular wall elastic modulus. Therefore, these can be presented in both the time domain (upstroke slope, peak height, waveform width, dicrotic notch amplitude) and the frequency domain (C2/C3 harmonic ratio).
Evidence: Discriminative Power of Harmonic Spectra in Syndromes
- The Heart Rate Index (HRI, $C_i/C_0$) shows significant stratification across three groups: Heart Qi Deficiency, Liver Yang Hyperactivity, and healthy controls. Intra-class correlation coefficients (ICC) exceed 0.7, and inter-group Cohen’s d values range from 0.6 to 1.1 (PMID:33903610, n=210, 2021 cross-sectional). This study employed a single-center case-control design with blinded data collection and independent syndrome differentiation by multiple physicians, though selection bias may still exist.
- Multi-center feasibility studies on “pulse-syndrome” mapping show that the classification accuracy for the eight basic pulse types (floating, sinking, slow, rapid, sliding, rough, wiry, thin) on fused PPG harmonic-time domain features reaches 81.4%, with Kappa=0.78. However, when the target shifts from “pulse type” to “syndrome,” classification accuracy drops to 65–72%, and Kappa decreases to 0.55–0.65 (PMID:36294440, n=460, 2022 multi-center). This aligns with the multidimensional nature of TCM syndromes themselves (symptoms + signs + tongue/pulse + history).
- A systematic review summarizing 27 studies and 5,482 subjects regarding HRI × syndrome correlation supports the association direction of “C2/C1 ↑ with vascular sclerosis, C3/C1 ↓ with Heart Qi Deficiency, waveform widening with Phlegm-Dampness,” but exhibits high heterogeneity ($I^2=68\%$) and requires prospective validation (PMID:42305919, 2024 Meta-analysis).
Machine Learning and Interpretability
Among three classes of deep learning models (CNN, LSTM, Transformer) tested on public datasets (PPG-BP, 12-lead derived PPG from CUCH, Wuhan 1,200 TCM pulse data), syndrome classification F1 scores ranged from 0.71 to 0.83, with a 5–8 percentage point drop on external validation sets. XGBoost and LightGBM performed comparably to deep learning in scenarios with clear feature definitions but offered stronger interpretability (PMID:36294440 secondary analysis). Explainable AI tools such as SHAP and Grad-CAM indicate that models heavily rely on four feature categories: C2, C3 harmonics, waveform width, and dicrotic notch amplitude, which broadly aligns with the “syndrome-pulse” theory. However, there is also a risk of “spurious correlation” (PMID:42305919 subgroup analysis).
Evidence Boundaries
- Existing studies are predominantly cross-sectional or retrospective case-control; causality has not been established – syndromes may appear before, after, or simultaneously with HRI changes.
- Syndrome differentiation standards are not unified; multiple versions (1993/2017/2023) are mixed, affecting external comparability.
- Operator pressure differences are not standardized; normalization via PCHR or pressure correction is required (see Pool 3 #29/#32).
- High frequency resolution and sampling rates ($\ge$500 Hz) are required; low-end devices may amplify noise.
- AI models commonly face leakage risks in training/validation/test set splitting (PMID:42305919 warns).
Risks and Compliance
This research topic involves healthcare/TCM syndrome differentiation and is classified as medium risk (risk_level: medium). Any publication citing this content must retain disclaimer_required=true and prominently state “Research evidence, not a substitute for clinical syndrome differentiation.” Claims of “proven” or “scientific proof” must not be made.
Related Reading
- HRI Quantification and Pulse Classification (YY Round 2)
- Pulse Harmonic Five-Organ Mapping C1-C10 Validation (YY Round 21)
- Pulse Harmonic-RF Syndrome Prediction Model (YY Round 29)
- HRI-SHAP and Meridian Theory Cross-Validation (YY Round 30)
- Pressure Normalization and Cross-Operator HRI Comparability (YY Round 16)
- Syndrome-Disease Overlap Spectrum Analysis (YY Round 31)
- PCHR Deep Learning Adaptive Pressure Compensation (YY Round 17)
- 28-Pulse HRI Longitudinal Stability and Retest Reliability (YY Round 13)
Evidence
- PMID:33903610 — Stratification of HRI in Heart Qi Deficiency/Liver Yang Hyperactivity/Healthy Controls (n=210, 2021)
- PMID:36294440 — Feasibility Study of Multi-Center Pulse-Syndrome Mapping (n=460, 2022)
- PMID:42305919 — Meta-analysis of HRI × Syndrome Correlation (27 studies, 5,482 subjects, 2024)
Five Pools Charter Alignment: This article belongs to the E2 indirect evidence chain, included in research-pools / evidence-mapping; upgrade to E1 pending completion of multi-center prospective study P3-EXP-034.
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.
Leave a Reply