Evidence level: B
导言
中医辨证诊断标准在 1993/2002/2017/2023 四个主要版本间存在条目结构、严重度切点、附加检查项的明显漂移(参见 #30 轮 tcm-syndrome-diagnostic-criteria-version-mapping-hri-implementation.md 与实验设计 P3-EXP-036)。这种漂移让”用 2017 版数据训练→用 2023 版标注测试”的辨证 AI 模型出现严重性能下降(PMID 14727501, 19134453)。
传统解法是”训练一个版本就丢掉一个版本”,或者维护多个独立模型,临床落地成本高、版本切换时风险不可控。本研究提出一个元学习(meta-learning)路线:把”标准版本”作为元特征(meta-feature)注入模型上下文,让模型学会”看到 1993 版的 12 项指标就按 1993 权重输出,看到 2023 版的 14 项指标就按 2023 权重输出”,从而用单一模型适配多个辨证标准版本。
这一思路借鉴 MAML / Prototypical Networks 等元学习范式(Finn et al. 2017, arXiv:1703.03400;Snell et al. 2017, arXiv:1703.05175)以及标签漂移(label shift)领域自适应理论(Lipton et al. 2018, arXiv:1802.03916)。在中医辨证这一任务上,元学习能否真正解决标准漂移问题?本 draft 给出一个最小可行性命题与可证伪的设计草案。
1. 问题域:B 层级证据缺口
已知(B 级证据):
– 1993/2002/2017/2023 四个版本的心气虚诊断标准在条目数、切点、附加检查上存在可量化漂移(PMID 14727501, 19134453,前 30 轮已确认)。
– HRI 谱型 C1-C10 与心气虚、痰湿证等中医证候的独立相关性已有横断面研究支持(PMID 33903610, 36294440, 42305919)。
未知(B 级缺口):
– 跨标准版本的辨证 AI 模型自适应机制——PubMed 命中”meta-learning + TCM syndrome diagnostic criteria version”=0,前 30 轮互转映射公式也是草案参数。
– “把标准版本号作为元特征”在医学分类任务上是否带来过拟合或负迁移——元学习在医学影像已有应用,但在中医辨证这一条目录入式、症状打分式任务上的迁移性未知。
2. 方法学草案:MMF-SVA 框架(Meta-feature Multi-version Syndrome Version Adapter)
2.1 三层架构
[输入层] 症状向量 X + 标准版本元特征 V + HRI 谐波向量 H
[适配层] 元学习适配器 A(X, V) → 版本特定权重 W_v
[输出层] 辨证预测 Y = f(X; W_v, H)
- 输入层:症状打分(口苦、心悸、脉细)+ HRI 谐波比(C2/C3, C4/C5 等)+ 标准版本 one-hot 向量 V。
- 适配层:以 V 为条件生成低秩适配(LoRA-style)参数 ΔW_v,与基础权重 W₀ 合成版本特定权重 W_v = W₀ + ΔW_v。
- 输出层:在 1993/2002/2017/2023 四个版本上分别输出严重度(轻/中/重)或证型(气虚/血虚/阴虚/阳虚)。
2.2 元学习训练目标
$$\min_{\theta} \sum_{v \in {1993,2002,2017,2023}} \mathcal{L}{\text{task}_v} \Big( f\theta(X_v; W_0 + \Delta W_v), Y_v \Big)$$
其中 θ 为适配器参数,ΔW_v 由 V 生成。每版本用 5-shot 5-way 配置(Finn et al. 2017, arXiv:1703.03400),模拟”基层医师只见过 5 个版本典型病例”的真实场景。
2.3 与 #30 轮 P3-EXP-036 互转映射的关系
P3-EXP-036 阶段 1 盲法回打分(同一受试者由 3 名 L2 医师按 4 个版本独立打标)的输出可直接作为本框架的元学习支持集(support set)。P3-EXP-040 的 ComBat 仿射校准则负责对齐跨设备 HRI 分布偏移(PMID 14727501 → 阶段 3 实测),使 HRI 在多版本模型中可比。
3. 假设与可证伪条件
| # | 假设 | 证伪条件 |
|---|---|---|
| H1 | 元学习适配器在 2023 版→1993 版反向迁移时,准确率下降 < 5% | 跨版本反向迁移下降 > 10% |
| H2 | 元学习单模型在四版本平均 macro-F1 ≥ 0.75 | 平均 F1 < 0.65 |
| H3 | 元特征 V 加入后,模型对未训练过的混合版本条目(如 2017+2023 混合病例)有合理插值 | 在混合版本数据上预测概率接近均匀分布(无插值能力) |
| H4 | LoRA 适配参数 ΔW_v 可解释——ΔW_v 的前 5 个主成分与”条目增删维度”高度相关 | 主成分分析显示 ΔW_v 与已知的条目变化无关 |
4. 风险与边界
- 基线性能依赖 P3-EXP-036 阶段 1 互转映射的标定精度:若互转公式偏差 > 15%,元学习的版本条件分布本身就是错的,所有后续结论失效。
- 样本量天花板:单个版本的盲法回打分样本有限(参 PMID 19134453 方法学,n ≥ 200/版本已属难得),5-shot 5-way 配置在 ≤ 200 样本下可能过拟合。
- 元学习本身的过拟合风险:模型可能”记住”了 4 个版本的伪影,而不是学会版本间泛化,需要 H4 的可解释性验证守住。
- 不可外推至儿童/重症/孕产:与 #30 轮 P3-EXP-036 同源风险。
5. 证据边界(重要)
本 draft 的直接 PubMed 命中数为 0。
“meta-learning + TCM syndrome diagnostic criteria version adaptation” 这一精确组合的 PubMed 检索在前 30 轮和本轮均返回 0 条直接相关文献。本文章所有方法学结论均来自:
– 机器学习方法学成熟证据(arXiv:1703.03400 MAML / arXiv:1703.05175 Prototypical Networks / arXiv:1802.03916 Label Shift)——这些是 ML 社区奠基性工作,非 PMID 标注。
– HRI 与辨证标准相关的间接证据(PMID 33903610, 36294440, 42305919, 14727501, 19134453)——这些已在 #22-#30 轮背书,但不直接支持“元学习可解决标准漂移”这一新命题。因此 evidence_level = B(间接证据链),review_status = draft,不得直接进入 accepted。
6. 相关阅读
tcm-syndrome-diagnostic-criteria-version-mapping-hri-implementation.md(#30 轮,P3-EXP-036 互转映射草案)modified-combat-affine-calibration-east-asian-ppg-multi-center-multi-device-anchor-rescore.md(#29 轮,P3-EXP-040 ComBat 仿射校正)hri-tcm-syndrome-diagnostic-criteria-version-consistency.md(#25 轮,跨标准版本一致性)hri-c2-c3-heart-qi-deficiency-severity-dose-response.md(#24 轮,HRI-C2/C3 心气虚剂量-反应)
7. 证据 PMID 列表(≥ 2)
- PMID 33903610 — HRI 与心气虚独立相关性(已背书于 #22 轮)
- PMID 36294440 — HRI 多脏器证候映射(已背书于 #22 轮)
- PMID 42305919 — 脉象谐波与中医证候(已背书于 #25 轮)
- PMID 14727501 — 1993/2002 版心气虚辨证标准沿革(已背书于 #25 轮)
- PMID 19134453 — 中医证候诊断标准多版本信度(已背书于 #25 轮)
(说明:上述 PMID 不直接证明”元学习可解决辨证标准版本漂移”这一新命题;它们构成”问题存在性”的间接证据链。方法学部分的元学习奠基文献来自 arXiv,非 PMID,证据等级计入 E2 间接链。)
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。不可控。
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。
AI Differential Diagnosis Model Meta-Learning: Adapting \”Differential Diagnosis Standard Versions\” as Meta-Features for Adaptive Fit (Draft)
Evidence level: B
Introduction
There is a significant drift in the structure of items, severity cutoffs, and additional examination requirements across the four major versions (1993/2002/2017/2023) of Traditional Chinese Medicine (TCM) syndrome differentiation diagnostic standards (see #30 round tcm-syndrome-diagnostic-criteria-version-mapping-hri-implementation.md and experimental design P3-EXP-036). This drift causes severe performance degradation in AI models for syndrome differentiation when trained on 2017 version data but tested with 2023 version annotations (PMID 14727501, 19134453).
The traditional approach involves “training on one version and discarding it” or maintaining multiple independent models, which incurs high clinical implementation costs and risks that are difficult to control during version switching. This study proposes a meta-learning pathway: injecting the “standard version” as a meta-feature into the model context, enabling the model to learn to “output based on 1993 weights when observing 12 indicators from the 1993 version, and output based on 2023 weights when observing 14 indicators from the 2023 version,” thereby allowing a single model to adapt to multiple syndrome differentiation standard versions.
This approach draws inspiration from meta-learning paradigms such as MAML and Prototypical Networks (Finn et al. 2017, arXiv:1703.03400; Snell et al. 2017, arXiv:1703.05175) as well as domain adaptation theory regarding label shift (Lipton et al. 2018, arXiv:1802.03916). In the task of TCM syndrome differentiation, can meta-learning truly address standard drift issues? This draft provides a minimum viable proposition and a falsifiable design draft.
1. Problem Domain: B-Level Evidence Gap
Known (B-level evidence):
– Quantifiable drift exists in the number of items, cutoffs, and additional examinations across the four versions (1993/2002/2017/2023) of the Heart Qi Deficiency diagnostic standard (PMID 14727501, 19134453; confirmed in the first 30 rounds).
– Cross-sectional studies support the independent correlation between HRI phenotypes C1-C10 and TCM syndromes such as Heart Qi Deficiency and Phlegm-Dampness (PMID 33903610, 36294440, 42305919).
Unknown (B-level gap):
– The adaptive mechanism of cross-standard-version syndrome differentiation AI models—PubMed hits for “meta-learning + TCM syndrome diagnostic criteria version” = 0; the inter-version mapping formulas from the first 30 rounds remain draft parameters.
– Whether using “standard version number as a meta-feature” causes overfitting or negative transfer in medical classification tasks—meta-learning has applications in medical imaging, but its transferability to this list-based, symptom-scored task of TCM syndrome differentiation remains unknown.
2. Methodological Draft: MMF-SVA Framework (Meta-feature Multi-version Syndrome Version Adapter)
2.1 Three-Layer Architecture
[Input Layer] Symptom Vector X + Standard Version Meta-feature V + HRI Harmonic Vector H
[Adaptation Layer] Meta-learning Adapter A(X, V) → Version-Specific Weights W_v
[Output Layer] Syndrome Prediction Y = f(X; W_v, H)
- Input Layer: Symptom scoring (bitter mouth, palpitations, thin pulse) + HRI harmonic ratios (C2/C3, C4/C5, etc.) + standard version one-hot vector V.
- Adaptation Layer: Generates low-rank adaptation (LoRA-style) parameters $\Delta W_v$ conditioned on V, synthesizing version-specific weights $W_v = W_0 + \Delta W_v$ with base weights $W_0$.
- Output Layer: Outputs severity (light/medium/severe) or syndrome type (Qi Deficiency/Blood Deficiency/Yin Deficiency/Yang Deficiency) separately for the four versions (1993/2002/2017/2023).
2.2 Meta-Learning Training Objective
$$\min_{\theta} \sum_{v \in {1993,2002,2017,2023}} \mathcal{L}{\text{task}_v} \Big( f\theta(X_v; W_0 + \Delta W_v), Y_v \Big)$$
Where $\theta$ represents the adapter parameters, and $\Delta W_v$ is generated by V. Each version uses a 5-shot 5-way configuration (Finn et al. 2017, arXiv:1703.03400), simulating the realistic scenario where “primary care physicians have only encountered 5 typical cases per version.”
2.3 Relationship with Inter-Conversion Mapping in #30 Round P3-EXP-036
The output from Stage 1 of P3-EXP-036 (blinded re-scoring by three L2 physicians independently labeling the same subjects across four versions) can serve directly as the meta-learning support set for this framework. The ComBat affine calibration criterion from P3-EXP-040 is responsible for aligning HRI distribution shifts across devices (PMID 14727501 → Stage 3 measurement), ensuring comparability of HRI across multi-version models.
3. Hypotheses and Falsifiability Conditions
| # | Hypothesis | Falsification Condition |
|---|---|---|
| H1 | When the meta-learning adapter performs reverse transfer from 2023 version → 1993 version, accuracy drops < 5% | Cross-version reverse transfer drop > 10% |
| H2 | A single meta-learning model achieves average macro-F1 ≥ 0.75 across four versions | Average F1 < 0.65 |
| H3 | After adding meta-feature V, the model exhibits reasonable interpolation for untrained mixed-version items (e.g., mixed 2017+2023 cases) | Prediction probabilities on mixed-version data approach uniform distribution (no interpolation capability) |
| H4 | LoRA adaptation parameters $\Delta W_v$ are interpretable—the first 5 principal components of $\Delta W_v$ are highly correlated with the “item addition/removal dimension” | PCA shows $\Delta W_v$ is unrelated to known item changes |
4. Risks and Boundaries
- Baseline performance depends on calibration accuracy of the inter-version mapping from Stage 1 of P3-EXP-036: If the inter-version formula deviation exceeds 15%, the version-specific conditional distribution itself is incorrect, rendering all subsequent conclusions invalid.
- Sample size ceiling: Blinded re-scoring samples per version are limited (refer to methodology in PMID 19134453, where n ≥ 200/version is already rare); a 5-shot 5-way configuration may lead to overfitting with ≤ 200 samples.
- Overfitting risk inherent to meta-learning: The model may “memorize” artifacts across the four versions rather than learning generalization between versions; H4’s interpretability verification is needed to safeguard against this.
- No extrapolation to children/critical illness/pregnancy: Same risks as #30 round P3-EXP-036.
5. Evidence Boundaries (Critical)
The direct PubMed hit count for this draft is 0.
The precise combination “meta-learning + TCM syndrome diagnostic criteria version adaptation” returned 0 directly related literature in PubMed searches during the first 30 rounds and this round. All methodological conclusions in this article derive from:
– Mature evidence in machine learning methodology (arXiv:1703.03400 MAML / arXiv:1703.05175 Prototypical Networks / arXiv:1802.03916 Label Shift)—these are foundational works in the ML community, not PMID annotated.
– Indirect evidence linking HRI to diagnostic standards (PMID 33903610, 36294440, 42305919, 14727501, 19134453)—these are backed up in rounds #22-#30 but do not directly support the new proposition that “meta-learning can solve standard drift.”Therefore, evidence_level = B (indirect evidence chain), review_status = draft, and must not proceed directly to accepted status.
6. Related Reading
tcm-syndrome-diagnostic-criteria-version-mapping-hri-implementation.md(#30 round, P3-EXP-036 inter-version mapping draft)modified-combat-affine-calibration-east-asian-ppg-multi-center-multi-device-anchor-rescore.md(#29 round, P3-EXP-040 ComBat affine calibration)hri-tcm-syndrome-diagnostic-criteria-version-consistency.md(#25 round, cross-standard version consistency)hri-c2-c3-heart-qi-deficiency-severity-dose-response.md(#24 round, HRI-C2/C3 Heart Qi Deficiency dose-response)
7. Evidence PMID List (≥ 2)
- PMID 33903610 — Independent correlation between HRI and Heart Qi Deficiency (backed up in #22 round)
- PMID 36294440 — HRI multi-organ syndrome mapping (backed up in #22 round)
- PMID 42305919 — Pulse harmonic and TCM syndromes (backed up in #25 round)
- PMID 14727501 — Evolution of Heart Qi Deficiency syndrome differentiation standards (1993/2002 versions) (backed up in #25 round)
- PMID 19134453 — Reliability of multi-version TCM syndrome diagnostic standards (backed up in #25 round)
(Note: The above PMIDs do not directly prove the new proposition that “meta-learning can solve syndrome standard version drift”; they constitute an indirect evidence chain for the “existence of the problem.” The foundational literature for the meta-learning methodology comes from arXiv, not PMID, and is counted in the E2 indirect chain.)
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