Evidence level: B
藏医三因体质判别模型的”金标准”标签本身有噪声——量表自评与专家脉诊不一致的灰色地带,可能正是模型学错方向的地方。
正文
藏医把人体能量归纳为隆(rlung,风)、赤巴(mkhris pa,火)、培根(bad kan,水土)三因,体质辨识是”因人施治”的第一步。现代研究把这一步搬进了机器学习:采集桡动脉脉搏波,用谐波拟合提取振幅比例、相位角、变异系数等频域指标,再训练分类器判别体质类型。这类研究的前提假设是:用来训练模型的”体质标签”是正确的。但这个假设在藏医体质判别场景下可能并不成立——这正是”标签噪声”问题。
标签噪声不是理论上的吹毛求疵。2020 年一项针对深度学习模型的研究(PMID 32749222)发现,在干净标注的数据集上,仅 2% 的标签错误就能让模型 AUC 显著下降;而公开数据集 NIH 和 CheXpert 的标签本身就需要”蒸馏”后才能使用。医学影像领域如此,脉搏判别领域只会更敏感——因为脉诊的”真值”依赖专家主观判断,而专家之间的分歧是系统性的。2023 年一项对 11 位 ICU 顾问医生独立标注的研究(PMID 36810915)给出了惊人的数据:同一批病例,专家间一致性仅有 Fleiss’ κ = 0.383(勉强算”一般”),外部验证时 pairwise Cohen’s κ 掉到 0.255(”极低”);更关键的是,研究者发现多数投票(majority vote)并不能得到最优模型,只有筛选出”可学习的标注”(learnable annotations)才能构建出最优分类器。另一项 2023 年发表于 Radiology AI 的研究(PMID 37035427)直接指出:”金标准”标签在低一致性类别上不可靠,多数投票的结果高度依赖”哪个专家组参与标注”。
把这套逻辑搬到藏医三因体质上,问题就清晰了。当前的主流流程是:给受试者发三因量表(自评)+ 请藏医专家做脉诊判定,二者一致就入组,不一致就丢弃或勉强归类。这种”用干净标签训练”的做法看似严谨,实则把最关键的信号丢掉了——不一致的灰色地带恰恰可能包含真实信息:如果谐波指标与量表维度不完全重合(量表测的是生活方式/体型/情绪倾向,脉搏测的是当下的血管动力学状态),那么”量表说培根、脉诊说赤巴”的受试者,其脉搏谐波特征很可能不是噪声,而是一种真实的中间态或混合态。
脉搏谐波判别本身的技术基础已经相当扎实。2015 年中国中医科学院团队用谐波拟合 + PCA/LASSO 判别脂肪肝与肝硬化患者脉搏(PMID 27088124),监督学习准确率超过 93%;2020 年台湾团队用 71 维时域/空间/频域特征 + PCA + 神经网络分类 10 种脉象(PMID 33378877),精度达 98.2%;2022 年 Sensors 上的一篇研究用 40 个谐波指标(振幅比例、变异系数、相位角标准差)做脉搏波形质量分类,随机森林 AUC = 0.96(PMID 36433203);2023 年上海中医药大学用径向脉搏波 + 7 种机器学习模型分类多囊卵巢综合征(PMID 37957660),459 名受试者测试准确率 72%、AUC 0.72。这些研究证明:谐波特征携带可判别的生理信息,且机器学习能稳定提取。
藏医三因的近亲阿育吠陀已经走在了前面。2017 年印度 CSIR 团队用 147 名健康人的表型数据做无监督聚类,自然涌现出三个与 Vata/Pitta/Kapha 三极体质对应的簇,监督建模达到高特异性与敏感度(PMID 28981546);2026 年一篇系统综述进一步梳理了 AI 与数字传感器用于 Prakriti 评估的进展,明确指出传统评估受限于主观性与可重复性(PMID 42546498)。
据此提出本实验设计:三因量表 + 双专家独立脉诊判定 → 按”双判一致 vs 不一致”分层 → 构建两套标签集(A:仅入组双判一致者;B:纳入全部,不一致者以量表或专家判定其一为标签)→ 分别建模做判别,对比 AUC 与校准度 → 对不一致者单独做频谱特征探索(降维可视化 + 与一致者特征空间的距离度量)。参考 36810915 的”可学习标注”策略:先算双专家的 κ 与标签可学习性,再决定哪部分标注能进训练集。若不一致者的谐波特征形成独立的中间簇,说明三因不是三个离散盒子而是连续谱——这既修正了体质分类学的标签策略,也反过来为”三因共体”的传统说法提供了现代频域证据。
证据与局限
已验证: 脉搏谐波指标携带可判别的生理信息,ML 判别精度可达 93%-98%(PMID 27088124, 33378877, 36433203, 37957660);标签噪声显著损害医学 ML 模型(2% 噪声即掉 AUC),多数投票非最优(PMID 32749222, 36810915, 37035427);阿育吠陀体质 ML 判别成立且无监督聚类自然涌现三极体质簇(PMID 28981546, 42546498)。
部分验证: 藏医三因与阿育吠陀三 dosha 结构同源但判定体系不同,跨体系类推成立但有边界。
待验证: 藏医三因双判不一致者的脉搏谐波是否真的形成独立中间簇——这是本实验设计要回答的核心问题,尚无直接数据。
相关阅读
证据等级: B
参考文献: PMID 27088124, PMID 33378877, PMID 36433203, PMID 37957660, PMID 32749222, PMID 36810915, PMID 37035427, PMID 28981546, PMID 42546498
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。
Control Experiments on ‘Label Noise’ in Tibetan Medicine’s Three Dosha Constitution Pulse Harmonic Discrimination: Comparing Separability Between Scale-Expert Consistent and Inc…
Evidence level: B
The “gold standard” labels inherent to the Tibetan medicine tri-dosha constitution model contain noise—the gray area where self-assessment scales diverge from expert pulse diagnosis. This discrepancy may precisely indicate where the model is learning incorrectly.
Main Text
Tibetan medicine categorizes human energy into three factors: rlung (wind), mkhris pa (fire), and bad kan (earth-water). Constitution identification constitutes the first step of “treating according to the individual.” Modern research has translated this step into machine learning: collecting radial artery pulse waves, extracting frequency-domain metrics such as amplitude ratios, phase angles, and coefficients of variation via harmonic fitting, and then training classifiers to discriminate constitution types. The premise assumption underlying such research is that the “constitution labels” used for training are accurate. However, this assumption may not hold in the context of Tibetan medicine constitution discrimination—this is precisely the issue of “label noise.”
Label noise is not mere theoretical nitpicking. A 2020 study on deep learning models (PMID 32749222) found that on clean-labeled datasets, only a 2% label error can cause a significant drop in model AUC; furthermore, labels in public datasets like NIH and CheXpert require “distillation” before use. As is the case in medical imaging, the pulse discrimination domain would be even more sensitive—because the “truth” of pulse diagnosis relies on subjective expert judgment, and inter-expert divergence is systematic. A 2023 study involving independent annotation by 11 ICU consultant physicians (PMID 36810915) provided startling data: for the same set of cases, inter-expert consistency was only Fleiss’ κ = 0.383 (barely “fair”), dropping to pairwise Cohen’s κ = 0.255 (“very low”) during external validation; more critically, researchers found that majority voting does not yield the optimal model, and only by filtering for “learnable annotations” can an optimal classifier be constructed. Another study published in Radiology AI in 2023 (PMID 37035427) directly stated that “gold standard” labels are unreliable in low-consistency categories, and majority voting results depend heavily on “which expert panel participated in the annotation.”
Applying this logic to Tibetan medicine tri-dosha constitution clarifies the issue. The current mainstream workflow involves administering a tri-dosha scale (self-assessment) + requesting Tibetan medicine experts for pulse diagnosis; cases are enrolled only if both agree, otherwise discarded or forced into a category. This practice of “training with clean labels” appears rigorous but discards the most critical signal—the gray area of inconsistency may actually contain genuine information: if harmonic indicators do not fully overlap with scale dimensions (scales measure lifestyle/body type/emotional tendencies, while pulses measure current vascular dynamics), then subjects where “the scale says bad kan but the pulse diagnosis says mkhris pa” likely possess pulse harmonic features that are not noise, but rather a true intermediate or mixed state.
The technical foundation for pulse harmonic discrimination is already quite solid. In 2015, a team from the Chinese Academy of Traditional Medicine Sciences used harmonic fitting + PCA/LASSO to discriminate pulse waves between fatty liver and cirrhosis patients (PMID 27088124), achieving supervised learning accuracy exceeding 93%; in 2020, a Taiwanese team used 71 time/space/frequency domain features + PCA + neural networks to classify 10 pulse types (PMID 33378877), reaching 98.2% accuracy; a 2022 study in Sensors used 40 harmonic indicators (amplitude ratios, coefficient of variation, phase angle standard deviation) for pulse waveform quality classification with a Random Forest AUC = 0.96 (PMID 36433203); and in 2023, Shanghai University of Traditional Chinese Medicine used radial pulse waves + 7 machine learning models to classify polycystic ovary syndrome (PMID 37957660), achieving 72% accuracy and AUC 0.72 among 459 subjects. These studies demonstrate that harmonic features carry discriminable physiological information and that machine learning can stably extract them.
The close relative of Tibetan medicine, Ayurveda, has moved ahead. In 2017, an Indian CSIR team used phenotypic data from 147 healthy individuals for unsupervised clustering, naturally emerging three clusters corresponding to the Vata/Pitta/Kapha tri-dosha constitutions, with supervised modeling achieving high specificity and sensitivity (PMID 28981546); a 2026 systematic review further梳理ed (reviewed) the progress of AI and digital sensors for Prakriti assessment, explicitly noting that traditional assessment is limited by subjectivity and reproducibility (PMID 42546498).
Based on this, we propose the following experimental design: Tri-dosha scale + independent pulse diagnosis by two experts → stratify by “double-judgment consensus vs. discrepancy” → construct two label sets (A: only those with double-consensus; B: include all, using either scale or expert judgment for discrepancies) → model separately and compare AUC and calibration → perform spectral feature exploration on discrepancies alone (dimensionality reduction visualization + distance metrics against consistent cases). Referencing the “learnable annotation” strategy from PMID 36810915: first calculate κ and label learnability between the two experts, then decide which annotations enter the training set. If the harmonic features of inconsistent subjects form a distinct intermediate cluster, it indicates that tri-dosha are not three discrete boxes but a continuous spectrum—this corrects the labeling strategy in constitution taxonomy and simultaneously provides modern frequency-domain evidence for the traditional claim of “co-existing tri-dosha.”
Evidence and Limitations
Verified: Pulse harmonic indicators carry discriminable physiological information; ML discrimination accuracy reaches 93%-98% (PMID 27088124, 33378877, 36433203, 37957660); label noise significantly degrades medical ML models (2% noise causes AUC drop); majority voting is not optimal (PMID 32749222, 36810915, 37035427); Ayurvedic constitution ML discrimination is established and unsupervised clustering naturally emerges tri-dosha clusters (PMID 28981546, 42546498).
Partially Verified: Tibetan medicine tri-dosha and Ayurvedic three dosha share structural homology but differ in judgment systems; cross-system analogy holds but has boundaries.
To be Verified: Whether pulse harmonic features of Tibetan medicine tri-dosha double-judgment inconsistent subjects truly form an independent intermediate cluster—this is the core question this experimental design aims to answer, for which there is currently no direct data.
Related Reading
- Tibetan Medicine Tri-Dosha Constitution and Pulse Harmonics
- Blind Discrimination of Tri-Dosha Pulse Harmonics
- Research Pool: Biological Frequency Code
Evidence Level: B
References: PMID 27088124, PMID 33378877, PMID 36433203, PMID 37957660, PMID 32749222, PMID 36810915, PMID 37035427, PMID 28981546, PMID 42546498
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