Evidence level: B
导言
中医脉诊倚赖指下触感与医师经验,传承依赖口耳心传,难以标准化和规模化。这一困境在数字时代催生了一类新型工具——脉象仪(电子脉诊传感器)。然而,十余种商业脉象仪各自采集、各自存储,输出波形数据格式互不相通。一份来自某国产脉象仪的原始信号,无法直接输入另一台设备的分析算法,也无法汇入多中心研究数据库。
核心问题不是算法不好,而是数据格式不统一。 没有标准化输出格式,就没有可比较性,就没有规模化的临床研究,也就没有真正意义上的AI辅助脉诊。本篇梳理这一领域的现状,分析统一输出的技术路径,并提出初步的标准化草案。
现状:数据格式的”战国时代”
当前市场上的中医脉象仪按传感器技术可分为三类:
| 技术路线 | 代表设备 | 输出格式 | 主要缺陷 |
|---|---|---|---|
| 压力传感器阵列 | 多家国产脉象仪 | 16位RAW + CSV | 无统一时基,注释字段各异 |
| PPG(光电容积脉搏波) | 智能手表/指环类 | PPG波形 + 心率 | 无谐波分解,频率信息丢失 |
| 超声多普勒 | 少数研究级设备 | 时域波形 + 频谱图 | 依赖专用软件,无开放接口 |
2015年,Nanyue Wei等人用机器学习分析脂肪肝与肝硬化患者的脉诊信号,发现健康志愿者与患者的脉搏波形存在显著差异,证实了脉诊信号可量化区分疾病(PMID: 27088124)。但这项研究的算法无法直接移植到另一台设备上——因为信号采集端根本不兼容。
2023年,台湾学者Hsiao等人在《Journal of Traditional and Complementary Medicine》发表了迄今最系统的脉搏谐波分析综述,回顾了28篇相关研究,指出”解读脉搏波数据的最大挑战在于仪器异质性”(PMID: 38481553)。
2026年,Wu等人用压力传感器采集乙肝/丙肝患者的桡动脉脉搏波,进行谐波分解(HA),发现特定谐波分量与肝脏病变显著相关,再次佐证了频率域特征的诊断价值——但其实验室自行开发的采集系统的数据格式,在论文中只字未提(PMID: 42487941)。
这正是问题所在: 每篇高质量论文都在用自己的数据格式发表结论,而整个领域却没有一个共享的数据基础设施。
标准化缺失的代价
1. 研究不可复现
缺乏统一格式,研究者无法获取他人的原始数据进行独立验证。PMID: 37969185的研究(用腕式血压计波形诊断肾功能不全)虽提出了156个谐波特征,但该算法依赖特定设备的采样率(通常200–500 Hz)和前置滤波参数,换一台设备就可能失效。
2. AI模型无法迁移
深度学习模型在特定设备数据上训练好后,跨设备迁移准确率急剧下降。这是因为模型学到的不只是”脉搏特征”,还有特定设备的噪声特征和采样偏差。
3. 临床转化受阻
一款脉象仪产品如果在医院使用,患者的脉象历史数据会被锁定在该品牌系统内。医生想对比三个月前在其他医院的检测结果?几乎不可能。
技术路径:HRI + 元数据框架
标准化输出格式的核心设计原则是“信号与解释分离”——原始波形数据本身保持原样,元数据(设备参数、采集条件、被测者信息)以标准化字段附加其后。
谐波比率指数(HRI)作为核心输出
不同于直接输出原始波形,更可行的路径是在设备端计算标准化的频率域特征,以谐波比率指数(HRI)作为跨设备的可比输出单元。
HRI定义:
HRI_n = A_n / A_1 (第n次谐波幅度 / 基波幅度)
其中A₁为心率频率对应的基波幅度,n=1–10覆盖主要的脏腑谐波频段(王唯工理论C1–C10)。
Hsiao等人的综述(PMID: 38481553)已确认HRI在区分不同生理状态方面的有效性。Wu等人(PMID: 42487941)的研究进一步在肝脏疾病患者中验证了特定HRI分量的异常。
HRI作为标准化输出的优势在于:它是设备无关的——无论采样率是200 Hz还是500 Hz,只要满足Nyquist定理,计算出的HRI值具有可比性。
元数据框架
每条输出记录应包含以下标准化元数据字段:
| 字段 | 说明 | 格式 |
|---|---|---|
device_id |
设备唯一标识 | UUID |
sensor_type |
传感器类型 | enum: pressure/PPG/ultrasound |
sampling_rate_hz |
采样率 | float |
filter_chain |
滤波参数(JSON) | JSON |
position |
采集部位 | enum: cun/guan/chi_left/right |
subject_id |
被测者ID(匿名) | hash |
acquisition_time |
采集时间(UTC) | ISO8601 |
HRI_vector |
10维HRI向量 | float[10] |
raw_waveform |
原始波形(可选传输) | float[] |
核心挑战
1. 商业利益壁垒
设备厂商缺乏开放格式的动力——锁定数据就是锁定客户。推动标准化需要行业协会或监管机构(国家药监局/中医药管理局)介入。
2. 基线频率的个体差异
HRI依赖于心率作为基波参考,但不同个体的静息心率差异显著(60–100 bpm),导致HRI向量的可比性受心率漂移影响。需要引入心率归一化步骤。
3. 中医脉象术语的映射
“浮脉””沉脉””滑脉”等传统术语对应怎样的HRI特征范围?目前没有足够大的标注数据库来建立这一映射。这是未来研究的核心数据需求。
实验设计建议
鉴于本课题本质上是标准化基础设施问题,核心验证不是临床试验,而是互操作性测试:
| 阶段 | 内容 | 指标 |
|---|---|---|
| 阶段一 | 招募5台不同品牌脉象仪,各采集30例健康志愿者的脉搏数据 | 计算每台设备的HRI向量 |
| 阶段二 | 建立”标准参考波形”(合成信号或高精度校准设备) | 各设备HRI向量与参考值的偏差 < 10% |
| 阶段三 | 多中心开放数据库建设(5家医院,N=500) | 数据库规模、跨设备查询成功率 |
证据与局限
支持证据(B级):
– Hsiao et al. 2023(PMID: 38481553):系统综述确认脉搏谐波分析的临床价值及仪器异质性问题
– Wu et al. 2026(PMID: 42487941):HBV/HCV患者的桡动脉谐波分解,验证HRI的疾病区分能力
– Kuo et al. 2023(PMID: 37969185):腕式血压计波形+156个谐波特征,肾功能不全诊断
– Wei et al. 2015(PMID: 27088124):机器学习+脉搏信号,脂肪肝/肝硬化分类
主要局限:
– 目前尚无正式的中医脉象仪行业数据标准,任何提案都需要行业协会背书
– HRI的心率归一化问题尚未解决,跨个体比较仍存在系统性偏差
– 中医脉象术语与HRI特征的映射数据库尚未建立
相关阅读
- Hsiao KY, et al. “Peripheral pulse harmonic analysis and its clinical application: A systematic review.” J Tradit Complement Med. 2023;14(2):223-236. PMID: 38481553 [Free PMC]
- Wu CH, et al. “Harmonic analysis of radial pulse in traditional Chinese medicine: physiological alterations associated with hepatitis B and hepatitis C infections.” Front Med (Lausanne). 2026;13:1848946. PMID: 42487941 [Free PMC]
- Kuo YC, et al. “Diagnosis of kidney insufficiency by using the pressure waveforms of wrist-type sphygmomanometers.” Am J Transl Res. 2023;15(10):6015-6025. PMID: 37969185 [Free PMC]
- Nanyue W, et al. “Pulse Diagnosis Signals Analysis of Fatty Liver Disease and Cirrhosis Patients by Using Machine Learning.” ScientificWorldJournal. 2015;2015:859192. PMID: 27088124 [Free PMC]
本篇属于池子③ · 生物频率密码 · 课题#17(勘探第11轮)
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。
Standardization Dilemma of Traditional Chinese Medicine Pulse Diagnosis Instruments: Why Data Cannot Interoperate
Evidence level: B
Introduction
Traditional Chinese medicine (TCM) pulse diagnosis relies on tactile perception and physician experience, with transmission depending on oral tradition and mental inheritance, making standardization and scaling difficult. This dilemma has spawned a new category of tools in the digital age—pulse monitors (electronic pulse diagnosis sensors). However, over a dozen commercial pulse monitors collect data independently, store it separately, and output waveform data in mutually incompatible formats. Raw signals from one domestic pulse monitor cannot be directly input into the analysis algorithms of another device, nor can they be integrated into multi-center research databases.
The core issue is not poor algorithms, but non-uniform data formats. Without standardized output formats, there is no comparability, no scalable clinical research, and consequently, no genuine AI-assisted pulse diagnosis. This paper reviews the current status of this field, analyzes technical pathways for unified output, and proposes a preliminary standardization draft.
Status Quo: The “Warring States Era” of Data Formats
Current TCM pulse monitors can be categorized into three types based on sensor technology:
| Technology Route | Representative Devices | Output Format | Primary Defect |
|---|---|---|---|
| Pressure Sensor Arrays | Multiple domestic pulse monitors | 16-bit RAW + CSV | No unified time base; varying annotation fields |
| PPG (Photoplethysmography) | Smartwatches/Ring sensors | PPG waveform + Heart rate | No harmonic decomposition; frequency information lost |
| Ultrasound Doppler | Few research-grade devices | Time-domain waveform + Spectrogram | Relies on proprietary software; no open interfaces |
In 2015, Nanyue Wei et al. used machine learning to analyze pulse signals from patients with fatty liver and cirrhosis, finding significant differences between healthy volunteers and patients’ pulse waveforms, confirming that pulse signals can be quantified to distinguish diseases (PMID: 27088124). However, the algorithm from this study cannot be directly ported to another device because the signal acquisition endpoints are fundamentally incompatible.
In 2023, Taiwanese scholars Hsiao et al. published the most systematic review of pulse harmonic analysis in the Journal of Traditional and Complementary Medicine, reviewing 28 related studies and noting that “the greatest challenge in interpreting pulse wave data lies in instrument heterogeneity” (PMID: 38481553).
In 2026, Wu et al. used pressure sensors to collect radial artery pulse waves from patients with hepatitis B and C, performed harmonic analysis (HA), and found that specific harmonic components are significantly associated with liver pathology, further corroborating the diagnostic value of frequency-domain features—yet their laboratory-developed acquisition system’s data format was not mentioned at all in the paper (PMID: 42487941).
This is precisely the problem: Each high-quality paper publishes conclusions using its own data format, while the entire field lacks a shared data infrastructure.
The Cost of Missing Standardization
1. Non-reproducibility of Research
Lack of a unified format prevents researchers from obtaining others’ raw data for independent verification. The study by PMID: 37969185 (using wrist-type sphygmomanometer waveforms to diagnose renal insufficiency) proposed 156 harmonic features, but its algorithm relies on specific device sampling rates (typically 200–500 Hz) and pre-filtering parameters; switching devices may render it ineffective.
2. Inability to Migrate AI Models
Deep learning models trained on data from specific devices see accuracy drop sharply when migrated across devices. This is because the models learn not only “pulse features” but also device-specific noise characteristics and sampling biases.
3. Blocked Clinical Translation
If a pulse monitor product is used in a hospital, patient pulse history data becomes locked within that brand’s system. If a doctor wishes to compare results from three months ago obtained at another hospital, it is nearly impossible.
Technical Pathway: HRI + Metadata Framework
The core design principle for standardized output formats is “Separation of Signal and Interpretation”—the raw waveform data remains unchanged, while metadata (device parameters, acquisition conditions, subject information) is appended via standardized fields.
Harmonic Ratio Index (HRI) as Core Output
Unlike direct output of raw waveforms, a more feasible path is to calculate standardized frequency-domain features at the device level, using the Harmonic Ratio Index (HRI) as a comparable output unit across devices.
HRI Definition:
HRI_n = A_n / A_1 (Amplitude of the nth harmonic / Amplitude of the fundamental wave)
Where A₁ is the amplitude corresponding to the heart rate frequency, and n=1–10 covers the main organ harmonic bands (based on Wang Weigong theory C1–C10).
Hsiao et al.’s review (PMID: 38481553) has already confirmed the effectiveness of HRI in distinguishing different physiological states. Wu et al. (PMID: 42487941) further validated specific HRI component abnormalities in patients with liver disease.
The advantage of HRI as a standardized output is that it is device-agnostic: regardless of whether the sampling rate is 200 Hz or 500 Hz, as long as the Nyquist theorem is satisfied, the calculated HRI values are comparable.
Metadata Framework
Each output record should include the following standardized metadata fields:
| Field | Description | Format |
|---|---|---|
device_id |
Unique device identifier | UUID |
sensor_type |
Sensor type | enum: pressure/PPG/ultrasound |
sampling_rate_hz |
Sampling rate | float |
filter_chain |
Filter parameters (JSON) | JSON |
position |
Acquisition site | enum: cun/guan/chi_left/right |
subject_id |
Subject ID (anonymized) | hash |
acquisition_time |
Acquisition time (UTC) | ISO8601 |
HRI_vector |
10-dimensional HRI vector | float[10] |
raw_waveform |
Raw waveform (optional transmission) | float[] |
Core Challenges
1. Commercial Interest Barriers
Device manufacturers lack the motivation to adopt open formats—locking data locks customers. Standardization requires intervention by industry associations or regulatory bodies (National Medical Products Administration / Administration of Traditional Chinese Medicine).
2. Individual Differences in Baseline Frequency
HRI relies on heart rate as a fundamental reference, but resting heart rates vary significantly among individuals (60–100 bpm), causing the comparability of HRI vectors to be affected by heart rate drift. A heart rate normalization step needs to be introduced.
3. Mapping of TCM Pulse Terminology
What HRI feature ranges correspond to traditional terms like “Floating pulse,” “Deep pulse,” or “Slippery pulse”? Currently, there is no sufficiently large annotated database to establish this mapping. This represents a core data requirement for future research.
Experimental Design Recommendations
Given that this topic is essentially a standardization infrastructure issue, the core validation is not a clinical trial but an interoperability test:
| Phase | Content | Metric |
|---|---|---|
| Phase I | Recruit 5 pulse monitors of different brands; each collects pulse data from 30 healthy volunteers | Calculate HRI vector for each device |
| Phase II | Establish a “Standard Reference Waveform” (synthetic signal or high-precision calibration device) | Deviation between each device’s HRI vector and reference value < 10% |
| Phase III | Multi-center open database construction (5 hospitals, N=500) | Database scale; cross-device query success rate |
Evidence and Limitations
Supporting Evidence (Level B):
– Hsiao et al. 2023 (PMID: 38481553): Systematic review confirming clinical value of pulse harmonic analysis and instrument heterogeneity issues
– Wu et al. 2026 (PMID: 42487941): Radial artery harmonic analysis in HBV/HCV patients, validating HRI’s disease discrimination capability
– Kuo et al. 2023 (PMID: 37969185): Wrist-type sphygmomanometer waveforms + 156 harmonic features for renal insufficiency diagnosis
– Wei et al. 2015 (PMID: 27088124): Machine learning + pulse signals for fatty liver/cirrhosis classification
Major Limitations:
– No formal industry data standard for TCM pulse monitors currently exists; any proposal requires backing from an industry association
– The heart rate normalization issue for HRI remains unresolved; systematic bias persists in cross-individual comparisons
– A mapping database between TCM pulse terminology and HRI features has not yet been established
Further Reading
- Hsiao KY, et al. “Peripheral pulse harmonic analysis and its clinical application: A systematic review.” J Tradit Complement Med. 2023;14(2):223-236. PMID: 38481553 [Free PMC]
- Wu CH, et al. “Harmonic analysis of radial pulse in traditional Chinese medicine: physiological alterations associated with hepatitis B and hepatitis C infections.” Front Med (Lausanne). 2026;13:1848946. PMID: 42487941 [Free PMC]
- Kuo YC, et al. “Diagnosis of kidney insufficiency by using the pressure waveforms of wrist-type sphygmomanometers.” Am J Transl Res. 2023;15(10):6015-6025. PMID: 37969185 [Free PMC]
- Nanyue W, et al. “Pulse Diagnosis Signals Analysis of Fatty Liver Disease and Cirrhosis Patients by Using Machine Learning.” ScientificWorldJournal. 2015;2015:859192. PMID: 27088124 [Free PMC]
This paper belongs to Pool 3 · Biological Frequency Codes · Project #17 (Exploration Round 11)
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