元学习适配器对”未训练过的新版中医辨证标准”的零样本外推能力评估

Written by

in

Language

Evidence level: B

导言

中医辨证标准从 1993 / 2002 / 2017 / 2023 已经历四版迭代(P3-EXP-036 / #30 轮互转映射协议),但第五版(行业预期 2025–2027)尚未公开。如果一个 AI 辨证模型在 2023 年版上训练,2025 年版发布时必须重新标注大量数据再训练——这就是”标准漂移灾难”。本文提出用元学习(meta-learning)让模型在 1993/2002/2017/2023 四版上学会”如何学辨证”,从而在 2025+ 未见版本上实现零样本(zero-shot)或仅需 5–20 个支持样本(few-shot)的快速适配。这是 P3-EXP-041(MMF-SVA 框架)的关键外推检验。

正文

1. 零样本外推的本质问题

零样本外推要求模型在没有见过目标版本条目的前提下,仅靠版本间的元知识(meta-knowledge)正确分类。元学习路径有三种主流方案:

  • MAML(Model-Agnostic Meta-Learning):在多个训练任务上寻找一个对所有任务都”易于微调”的初始化参数,使新任务只需少量梯度步即可收敛(arXiv:1703.03400)。
  • Prototypical Networks:把每个类别映射到一个嵌入空间的”原型向量”,新任务用最近邻分类,对未见类别有天然适配能力(arXiv:1703.05175)。
  • Domain Generalization:训练时不假设目标域可见,只最大化跨域不变特征,arXiv:2106.06908 已成功迁移到医学影像分类(CT/MRI 跨中心)。

在医学影像场景,MAML 已用于息肉分割和皮肤病灶分类(PMC13493216, Springer s42452-025-08188-3)。但所有这些研究的”任务”都是图像层级,对中医辨证这种”文本/数值混合标签 + 标准版本作为元特征”的目录入式任务,尚无直接验证。

2. 中医辨证版本漂移的独特性

与医学影像的”设备/中心漂移”不同,中医标准漂移有三个独有特征:

  1. 条目可增删:1993 版心气虚有 4 主症 + 3 兼症;2023 版可能加入”心肌能量代谢指标”作为微观条目(PMID:19134453 综述指出 2002 版起即开始纳入实验室指标)。
  2. 权重可重分配:同一症状(神疲乏力)在 1993 版权重 0.3,2023 版可能升至 0.5(P3-EXP-036 互转公式草案)。
  3. 阈值可漂移:C2/C3 谐波比值的”心气虚切点”在 2017 版 vs 2023 版可能有 0.05 的位移(#24 轮 P3-EXP-035 已记录此现象)。

这种漂移使传统零样本学习(标签空间固定)失效,必须用元学习让模型学到”漂移本身”的模式。

3. 元学习适配器 MMF-SVA 的零样本外推假设

基于 P3-EXP-041 提出的 MMF-SVA(Meta-Mapped Few-shot Standard-Version Adapter)框架,我们提出三个可证伪假设:

  • H1(完全零样本):仅靠 1993/2002/2017 三版训练,2023 版测试的辨证准确率 ≥ 0.70(基线 LoRA 微调为 0.55)。
  • H2(5-shot 元适应):在 2023 版提供 5 例支持样本,元学习适配器准确率 ≥ 0.85,而全参数微调需 ≥ 200 例才能达到同等水平。
  • H3(未见 2025+ 版本零样本):基于 1993/2002/2017/2023 四版训练的元学习器,对未发布的 2025+ 版本(用合成扰动模拟)的预测一致性 ≥ 0.75(Cohen’s Kappa)。

H3 是本轮核心命题。若 H3 被证伪,则标准版本更新仍需重新标注大量样本,元学习框架的实用价值降低。

4. 方法学证据链

直接证据:PubMed 命中 “meta-learning + TCM syndrome zero-shot” = 0。本轮只能拼接:
– 方法学间接链:MAML/Prototypical Networks 在 10+ 医学影像任务中已成熟(arXiv:2106.06908, Springer s42452-025-08188-3, PMC13493216)。
– 零样本临床链:PMC13511489 在 2D 乳腺 X 线未见病灶分类上达到 AUC 0.82;PMC11258407 用 LLM 零样本提取临床评估标准。
– HRI 基线:P3-EXP-040 ComBat 仿射校正已能跨 5 中心 × 3 设备校准 b/a 比(#29 轮),证明 HRI 数据的批次效应可控。
– 心气虚标准基线:PMID:14727501, 19134453 提供 1993/2002 版心气虚条目的实证对比。

5. 评估协议设计

由于真实 2025+ 版本未发布,本轮采用合成扰动验证法:

  • 从 2023 版条目集合出发,随机执行 3 类扰动:①新增 1–2 条”假想 2025 版”条目(如”心肌能量代谢”权重 0.4);②调整 2–3 条既有条目权重 ±0.15;③调整 1 个 C2/C3 谐波切点 ±0.05。
  • 在每个扰动版本上让模型零样本预测,统计 Cohen’s Kappa 与宏 F1。
  • 正向情景:若 1000 组随机扰动下 Kappa ≥ 0.75 占比 ≥ 80%,H3 成立。
  • 反向情景:若 ≥ 30% 的扰动导致 Kappa < 0.6,则元学习器对”条目增删型漂移”鲁棒性不足,需在 P3-EXP-043 引入”漂移感知正则化”。

6. 风险与边界

  • 过度拟合 4 个版本:若训练版本数过少,元学习可能学到版本特定的捷径,而非真正的”漂移元知识”。本轮通过 5-fold 留一版交叉验证缓解。
  • 合成扰动的真实性:合成扰动与真实标准修订存在差距,最终需待 2025+ 真版发布后做 post-hoc 验证。
  • HRI 数据依赖:所有输入依赖 P3-EXP-040 ComBat 校正链成功;若校正链失败,元学习器无法获得稳定的嵌入空间。
  • 外推边界:本框架仅适用于”条目级文本+谐波向量”输入,不适用于图像、舌象、问诊语音等其他模态。

证据与局限

直接证据:PMID 14727501(1993 版心气虚标准)、PMID 19134453(2002 版心气虚修订)、PMID 33903610 / 36294440 / 42305919(HRI-C2/C3 与心气虚关联的多中心研究证据)、PMID 27861248 / 27314975(HRI 跨设备信度基础)。

方法学奠基(arXiv,非 PMID):MAML (1703.03400)、Prototypical Networks (1703.05175)、Domain Generalization via MAML (2106.06908)、STL-DeepBDC (PMC13493216)、Zero-shot 2D mammography (PMC13511489)、Zero-shot criteria extraction (PMC11258407)、Evolving few-shot landscape (Springer s42452-025-08188-3)、Zero-shot endoscopic multi-label (Springer s11548-025-03439-5)。

局限:
– 真实 2025+ 版本尚未发布,本轮只能基于合成扰动推断;
– 元学习在中医辨证这一目录入式场景尚无第三方复现;
– HRI 信号稳定性依赖 P3-EXP-040 ComBat 校正链;
– 不适用于图像、语音等多模态输入。

相关阅读

  • P3-EXP-036 / tcm-syndrome-diagnostic-criteria-version-mapping-hri-implementation.md(互转映射协议草案,#30 轮)
  • P3-EXP-040 / modified-combat-affine-calibration-east-asian-ppg-multi-center-multi-device-anchor-rescore.md(ComBat 仿射校正,#29 轮)
  • P3-EXP-041 / tcm-syndrome-meta-learning-standard-version-adaptation.md(MMF-SVA 框架,#31 轮)
  • P3-EXP-035 / hri-c2-c3-heart-qi-deficiency-severity-dose-response.md(C2/C3 切点剂量-反应,#24 轮)
  • P3-EXP-042 / pool3-tcm-syndrome-meta-learning-zero-shot-2025-future-version-adaptation.md(本轮实验设计)

本文为五池③勘探产物 · 任务 #122 · 第 32 轮 · YY
证据等级 B · 风险等级 medium · review_status: draft · disclaimer_required: true

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。

Evaluation of Zero-Shot Extrapolation Capability of Meta-Learning Adapters for Untrained New Versions of Traditional Chinese Medicine Syndrome Differentiation Standards

Evidence level: B

Introduction

TCM syndrome differentiation standards have undergone four iterations since 1993, 2002, 2017, and 2023 (P3-EXP-036 / #30 round mutual mapping protocol), but the fifth version (industry expectation 2025–2027) remains unpublished. If an AI differentiation model is trained on the 2023 version, it must re-label a large volume of data and retrain upon the release of the 2025 version—this constitutes a “standard drift catastrophe.” This paper proposes using meta-learning to enable the model to learn “how to differentiate” across the four versions (1993/2002/2017/2023), thereby achieving zero-shot or few-shot (5–20 support cases) rapid adaptation on unseen 2025+ versions. This represents a key extrapolation validation for P3-EXP-041 (MMF-SVA framework).

Main Body

1. The Fundamental Problem of Zero-Shot Extrapolation

Zero-shot extrapolation requires the model to correctly classify based on meta-knowledge between versions without having seen specific entries from the target version. Three mainstream meta-learning pathways exist:

  • MAML (Model-Agnostic Meta-Learning): Searches for initialization parameters across multiple training tasks that are “easy to fine-tune” for all tasks, allowing convergence with few gradient steps on new tasks (arXiv:1703.03400).
  • Prototypical Networks: Maps each category to a “prototype vector” in an embedding space; new tasks utilize nearest-neighbor classification, offering natural adaptation capabilities for unseen categories (arXiv:1703.05175).
  • Domain Generalization: Trains without assuming the target domain is visible, maximizing cross-domain invariant features; arXiv:2106.06908 has successfully migrated this to medical image classification (CT/MRI across centers).

In medical imaging contexts, MAML has been applied to polyp segmentation and skin lesion classification (PMC13493216, Springer s42452-025-08188-3). However, all these studies operate at the image level; there is no direct validation for TCM differentiation, which involves a cataloging task with “text/numerical mixed labels + standard versions as meta-features.”

2. Unique Characteristics of TCM Standard Drift

Unlike medical imaging “device/center drift,” TCM standard drift possesses three unique characteristics:

  1. Entries Can Be Added or Removed: The 1993 version of Heart Qi Deficiency includes 4 main symptoms + 3 concurrent symptoms; the 2023 version may add “myocardial energy metabolism indicators” as micro-entries (PMID:19134453 review notes that laboratory indicators began inclusion from the 2002 version).
  2. Weights Can Be Reallocated: The same symptom (e.g., fatigue and weakness) may have a weight of 0.3 in the 1993 version but rise to 0.5 in the 2023 version (P3-EXP-036 mutual mapping formula draft).
  3. Thresholds Can Drift: The “Heart Qi Deficiency cutoff” for the C2/C3 harmonic ratio may shift by 0.05 between the 2017 and 2023 versions (#24 round P3-EXP-035 recorded this phenomenon).

Such drift invalidates traditional zero-shot learning (fixed label space), necessitating meta-learning to enable the model to learn the pattern of “drift itself.”

3. Zero-Shot Extrapolation Hypotheses for MMF-SVA Adapter

Based on the MMF-SVA (Meta-Mapped Few-shot Standard-Version Adapter) framework proposed in P3-EXP-041, we propose three falsifiable hypotheses:

  • H1 (Pure Zero-Shot): Using only training on the 1993/2002/2017 versions, differentiation accuracy on the 2023 test set ≥ 0.70 (baseline LoRA fine-tuning is 0.55).
  • H2 (5-shot Meta-Adaptation): Providing 5 support cases in the 2023 version, the meta-learning adapter achieves accuracy ≥ 0.85, whereas full-parameter fine-tuning requires ≥ 200 cases to reach the same level.
  • H3 (Zero-Shot on Unreleased 2025+ Versions): The meta-learner trained on the four versions (1993/2002/2017/2023) achieves prediction consistency ≥ 0.75 (Cohen’s Kappa) on unreleased 2025+ versions (simulated via synthetic perturbation).

H3 is the core proposition for this round. If H3 is falsified, standard version updates still require re-labeling large sample sets, reducing the practical value of the meta-learning framework.

4. Methodological Evidence Chain

Direct evidence: PubMed hits for “meta-learning + TCM syndrome zero-shot” = 0. This round can only assemble:
– Methodological Indirect Chain: MAML/Prototypical Networks are mature in 10+ medical imaging tasks (arXiv:2106.06908, Springer s42452-025-08188-3, PMC13493216).
– Zero-Shot Clinical Chain: PMC13511489 achieved AUC 0.82 on 2D mammography unseen lesion classification; PMC11258407 used LLM zero-shot extraction for clinical assessment criteria.
– HRI Baseline: P3-EXP-040 ComBat affine calibration has already calibrated b/a ratios across 5 centers × 3 devices (#29 round), proving HRI data batch effects are controllable.
– Heart Qi Deficiency Standard Baseline: PMID:14727501, 19134453 provide empirical comparisons of Heart Qi Deficiency entries from the 1993/2002 versions.

5. Evaluation Protocol Design

Since the real 2025+ version remains unpublished, this round adopts a synthetic perturbation validation method:

  • Starting from the 2023 version entry set, randomly execute three types of perturbations: ① Add 1–2 “hypothetical 2025 version” entries (e.g., “myocardial energy metabolism” weight 0.4); ② Adjust weights of 2–3 existing entries ±0.15; ③ Adjust one C2/C3 harmonic cutoff ±0.05.
  • Allow the model to make zero-shot predictions on each perturbed version and statistically analyze Cohen’s Kappa and macro F1.
  • Positive Scenario: If the proportion of Kappa ≥ 0.75 under 1000 random perturbations is ≥ 80%, H3 holds.
  • Negative Scenario: If ≥ 30% of perturbations result in Kappa < 0.6, the meta-learner lacks robustness against “entry addition/removal drift,” requiring “drift-aware regularization” to be introduced in P3-EXP-043.

6. Risks and Boundaries

  • Overfitting to Four Versions: If the number of training versions is too low, meta-learning may learn version-specific shortcuts rather than true “drift meta-knowledge.” This round mitigates via 5-fold leave-one-version-out cross-validation.
  • Authenticity of Synthetic Perturbation: Synthetic perturbation differs from real standard revisions; post-hoc validation is required once the true 2025+ version is released.
  • HRI Data Dependency: All inputs rely on the successful P3-EXP-040 ComBat calibration chain; if the chain fails, the meta-learner cannot obtain a stable embedding space.
  • Extrapolation Boundaries: This framework applies only to “entry-level text + harmonic vector” inputs and is not suitable for other modalities such as images, tongue diagnosis, or inquiry voice.

Evidence and Limitations

Direct Evidence: PMID 14727501 (1993 Heart Qi Deficiency standard), PMID 19134453 (2002 Heart Qi Deficiency revision), PMID 33903610 / 36294440 / 42305919 (multi-center study evidence linking HRI-C2/C3 to Heart Qi Deficiency), PMID 27861248 / 27314975 (HRI cross-device reliability foundation).

Methodological Foundations (arXiv, non-PMD): MAML (1703.03400), Prototypical Networks (1703.05175), Domain Generalization via MAML (2106.06908), STL-DeepBDC (PMC13493216), Zero-shot 2D mammography (PMC13511489), Zero-shot criteria extraction (PMC11258407), Evolving few-shot landscape (Springer s42452-025-08188-3), Zero-shot endoscopic multi-label (Springer s11548-025-03439-5).

Limitations:
– The real 2025+ version has not been released; this round can only infer based on synthetic perturbation;
– Meta-learning in TCM differentiation, a cataloging scenario, lacks third-party replication;
– HRI signal stability depends on the P3-EXP-040 ComBat calibration chain;
– Not applicable to multi-modal inputs such as images and voice.

Related Reading

  • P3-EXP-036 / tcm-syndrome-diagnostic-criteria-version-mapping-hri-implementation.md (mutual mapping protocol draft, #30 round)
  • P3-EXP-040 / modified-combat-affine-calibration-east-asian-ppg-multi-center-multi-device-anchor-rescore.md (ComBat affine calibration, #29 round)
  • P3-EXP-041 / tcm-syndrome-meta-learning-standard-version-adaptation.md (MMF-SVA framework, #31 round)
  • P3-EXP-035 / hri-c2-c3-heart-qi-deficiency-severity-dose-response.md (C2/C3 cutoff dose-response, #24 round)
  • P3-EXP-042 / pool3-tcm-syndrome-meta-learning-zero-shot-2025-future-version-adaptation.md (current experimental design)

This article is a product of Pool 5 Exploration · Task #122 · Round 32 · YY
Evidence Level B · Risk Level medium · review_status: draft · disclaimer_required: true


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 *