元学习适配器跨”心→肝/脾/肺/肾气虚”四脏气虚零样本迁移性验证 — 同步合成扰动敏感度图谱

Written by

in

Language

Evidence level: B

导言

五池第 33 轮 YY 推进了 P3-EXP-043(adapter+prompt 混合架构),其核心假设是”权重更新是有效路径、提示工程是便宜兼容路径”。但遗留两个 P1 问题尚未闭环:(a) 心气虚训练的元学习适配器能否迁移到肝/脾/肺/肾气虚(#123,上轮新立本轮实操);(b) 合成扰动类型(增条目/改权重/移切点)对零样本外推的失效占比如何分布(#124)。本轮设计 P3-EXP-044 同时回答这两个问题——H1:心气虚→四脏气虚零样本迁移 Cohen’s Kappa ≥ 0.60(基线对照为跨中心 Reshuffle 拆分);H2:增条目扰动失效占比 ≥ 40%(提示”增量修订”对元学习最不友好);H3:合成扰动敏感度图谱与”修订意图”对应,可指导未来真实标准修订的稳健策略。

正文

一、为什么”心→四脏”是元学习迁移性验证的最佳起点

中医五脏气虚证型虽同属”气虚”大证类,但条目结构差异显著:心气虚以神志/血脉为核心(心悸/气短/乏力/自汗/面色淡白),肝气虚以疏泄/筋目为核心(胁肋胀闷/善叹息/视物模糊/肢体麻木),脾气虚以运化/肌肉为核心(食少/腹胀/便溏/倦怠/消瘦),肺气虚以宣降/卫表为核心(咳嗽/气喘/自汗/易感冒),肾气虚以封藏/腰膝为核心(腰膝酸软/耳鸣/夜尿频/畏寒)。这种”同证类-异条目结构”为元学习提供了天然的跨任务(cross-task)评测场景——比单纯跨中心验证更具方法学价值。

Quiñonez-Baca 等(PMID 41008775,2025,Diagnostics)在胸片少样本分类上系统对比了 Prototypical Networks、Relation Networks、MAML、FoMAML 四种元学习架构,发现 ProtoNet 在 5-way 5-shot 设置下比 MAML 平均 AUC 高 0.05~0.08,且 MAML 的二阶导数优化在数据 < 1000 张时反而过拟合。这一发现直接指导本轮迁移性验证的设计:心气虚 → 四脏气虚的 N-way K-shot 设定应优先采用 ProtoNet 风格的度量学习路线,把”证型原型向量”作为元知识载体。

PMC12711456(2025)发表的肾虚证筛查核心要素研究证实,肾虚证的 5 维结构(腰膝酸软/畏寒肢冷/夜尿频/耳鸣/性功能减退)已具备可量化操作定义,意味着肾气虚的标注一致性(Cohen’s Kappa)可在专家共识上达到 0.75 以上,为元学习标签提供了高质量基底。PMC11589548(2024)的中医临床实践指南国际标准则覆盖了肺气虚、脾气虚、肾阳虚、肝肾阴虚等核心证型,证实多证型标准化术语已具备 ICD 对接基础——这是把元学习迁移结果与西医诊断代码打通的必要前提。

二、迁移性验证的三阶段实验设计

阶段1(基线锚定):从五池前 33 轮 P3-EXP-002/011/017/020/027 等数据集统一抽取心气虚 HRI 谐波特征 + 症状-舌脉多模态标签,按 P3-EXP-040 ComBat 仿射校正协议预处理;用 P3-EXP-041 MMF-SVA 元学习架构(5-way 5-shot)在心气虚 1993/2002/2017/2023 四版本数据上预训练。预训练冻结后保存元参数 θ_meta。

阶段2(跨证型迁移):把 θ_meta 在肝/脾/肺/肾气虚各证型的 ≤ 50 条 5-shot 支持集上做内循环微调(5 个 inner-step),冻结后输出四证型的预测;与四证型全量测试集(≥ 200 条/证型)对照,计算 Cohen’s Kappa 与 AUC。同时跑”基线对照”——随机初始化 ResNet 在四证型全量训练集上训练,相同测试集评估。H1 通过条件:元学习 vs 基线的 Cohen’s Kappa 差值 ≥ +0.10。

阶段3(合成扰动敏感度图谱):按 #31 轮 P3-EXP-041 设计的扰动模式(增条目/改权重/移切点三类各 333 组,共 1000 组)扰动测试集,量化元学习器在每类扰动下的失效占比;输出”扰动类型 × 证型” 二维敏感度热力图。H2 通过条件:增条目扰动失效占比 ≥ 40%(即”加新条目”对元学习最不友好);H3 通过条件:敏感度热力图与”修订意图”(拓展/收紧/重置)的对应一致性 ≥ 75%。

三、与 P3-EXP-043 混合架构的协同关系

P3-EXP-043(adapter+prompt hybrid)的设计目标是”零更新权重下的版本漂移吸收”。P3-EXP-044 与之互补:本轮关注”权重可更新的跨证型迁移性”,二者构成”空间维度 × 时间维度”的双轴验证空间。H1 通过后,P3-EXP-043 的”不更新权重”基线可被进一步收紧——心→四脏已通过元学习适配器解决,剩余的”版本差异”才有资格交给提示工程层;H1 失败则混合架构需重新设计(提示工程层承担更多负载)。

四、合成扰动敏感度图谱的方法学价值

合成扰动类型敏感性图谱(#124)若验证 H2,将直接回答”标准修订应优先采取哪种策略”。例如:若”增条目”扰动失效占比 60%、”改权重”扰动失效占比 30%、”移切点”扰动失效占比 40%,则可指导真实修订时——若新条目不可避免,应优先做小步增量化迭代(避免一次增加 ≥ 20% 条目);若仅需调整严重度切点,元学习器鲁棒性最强,可一次到位;改权重的修订最敏感,需要发布修订指南时附”元学习适配器增量更新指引”。

五、风险与边界

medium 风险主要来自:(a) 四脏气虚样本量天花板——肾气虚在五池前 33 轮数据集中样本量仍小于肺/脾气虚,需要从 PMC12711456 肾虚证核心要素研究的合作中心招募 ≥ 200 条标注样本;(b) 跨证型标注医师经验差异——心气虚由心血管科医师标注,肾气虚由肾病科/老年科医师标注,Kappa 在”同一患者—双标签”基准上需 ≥ 0.70 才能开始元学习训练;(c) 复合证型(心气虚+痰湿、心气虚+血瘀)本轮不纳入分析——这些是 #128 课题内容。

DOI 10.1186/s13020-026-01455-4(2026,Chin Med,TCM 可解释 AI 系统综述)指出 TCM AI 设计必须满足 4 个独特要求(可解释性/稀疏性/因果性/证候可分性),本轮 P3-EXP-044 的 SHAP 解释子模块直接对接这 4 个要求——特别是证候可分性,跨证型混淆矩阵本身就是证候可分性的直接证据。

证据与局限

  • 直接证据:5 个 PMID/DOI 在心气虚诊断标准(14727501、19134453)、心气虚 HRI 关联(33903610、36294440、42305919)、元学习跨类目迁移方法学(41008775)、TCM-DS LLM 指令微调可行性(41243110)、多证型术语国际标准化(11589548)、肾虚证量化筛查(12711456)、TCM 可解释 AI 设计框架(10.1186/s13020-026-01455-4)共 8 个独立证据点。
  • 间接证据:MAML/ProtoNet 在医学影像(PMC12468384、10.3390/diagnostics15182404)的迁移成功不能直接外推至”心→四脏气虚”目录入式任务;中医辨证标准版本漂移的剂量-效应(P3-EXP-036 阶段 1 估算 0.85 互转权重)仍属草案级假设。
  • 局限:本轮无法回答”真实 2025+ 标准修订”问题——必须等 P3-EXP-126 post-hoc 验证;不覆盖儿童/孕产/重症/复合证型;HRI 数据依赖 P3-EXP-040 校正链的稳定性。

相关阅读

  • tcm-syndrome-meta-learning-standard-version-adaptation.md
  • tcm-syndrome-meta-learning-zero-shot-2025-future-version-adaptation.md
  • tcm-syndrome-meta-learning-adapter-prompt-hybrid-architecture.md
  • tcm-syndrome-diagnostic-criteria-version-mapping-hri-implementation.md
  • pulse-harmonic-syndrome-independent-correlation.md
  • pulse-harmonic-five-zang-organ-function-correlation.md

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。

Zero-Shot Transfer Validation of Meta-Learning Adapters Across Four Zang Qi-Deficiency Patterns and Concurrent Sensitivity Mapping to Synthetic Perturbations

Evidence level: B

Introduction

Round 33 of the Five Pools project advanced P3-EXP-043 (adapter+prompt hybrid architecture), with the core hypothesis that “weight updates constitute an effective pathway while prompt engineering offers a cost-effective compatible pathway.” However, two P1 issues remain unresolved: (a) Can the meta-learning adapter trained on Heart Qi Deficiency transfer to Liver/Spleen/Lung/Kidney Qi Deficiency (#123, newly established last round and operationalized in this round); (b) How is the failure distribution distributed across synthetic perturbation types (insert entries/modify weights/shift cut-points) for zero-shot extrapolation (#124). This round designs P3-EXP-044 to address both questions simultaneously. H1: Zero-shot transfer from Heart Qi Deficiency to Four Zang Qi Deficiency achieves Cohen’s Kappa ≥ 0.60 (baseline control is cross-center Reshuffle splitting); H2: Insert-entry perturbation failure rate ≥ 40% (indicating “incremental revision” is least friendly to meta-learning); H3: The synthetic perturbation sensitivity map corresponds to “revision intent,” guiding robust strategies for future real-world standard revisions.

Body

I. Why “Heart → Four Zang” Is the Optimal Starting Point for Meta-Learning Transfer Validation

Although Traditional Chinese Medicine (TCM) Five Zang Qi Deficiency syndromes belong to the same “Qi Deficiency” major category, their entry structures differ significantly: Heart Qi Deficiency focuses on consciousness/blood vessels (palpitations/shortness of breath/fatigue/spontaneous sweating/pale complexion); Liver Qi Deficiency focuses on dispersal/tendons/eyes (hypochondriac distension/sighing/blurred vision/limb numbness); Spleen Qi Deficiency focuses on transportation/muscles (poor appetite/bloating/diarrhea/fatigue/wasting); Lung Qi Deficiency focuses on dispersal-descension/defensive exterior (cough/dyspnea/spontaneous sweating/frequent colds); Kidney Qi Deficiency focuses on sealing/storing/lumbar/knees (lumbar-knee soreness/tinnitus/nocturia/fear of cold). This “same syndrome category-different entry structure” provides a natural cross-task evaluation scenario for meta-learning—offering greater methodological value than simple cross-center validation alone.

Quiñonez-Baca et al. (PMID 41008775, 2025, Diagnostics) systematically compared Prototypical Networks, Relation Networks, MAML, and FoMAML four meta-learning architectures in chest X-ray few-shot classification. They found ProtoNet achieved 0.05~0.08 higher average AUC than MAML in a 5-way 5-shot setting, while MAML’s second-order derivative optimization led to overfitting when data was < 1000 samples. This finding directly guides the design of this round’s transfer validation: The N-way K-shot setting for Heart Qi Deficiency → Four Zang Qi Deficiency should prioritize a ProtoNet-style metric learning route, using “syndrome prototype vectors” as carriers of meta-knowledge.

PMC12711456 (2025) published research on core elements for Kidney Qi Deficiency screening, confirming that the 5-dimensional structure of Kidney Qi Deficiency (lumbar-knee soreness/fear of cold/nocturnal frequency/tinnitus/sexual function decline) possesses a quantifiable operational definition. This means annotation consistency (Cohen’s Kappa) for Kidney Qi Deficiency can reach above 0.75 on expert consensus, providing a high-quality baseline for meta-learning labels. PMC11589548 (2024), the International Standard for Traditional Chinese Medicine Clinical Practice Guidelines, covers core syndromes including Lung Qi Deficiency, Spleen Qi Deficiency, Kidney Yang Deficiency, and Liver-Kidney Yin Deficiency. This confirms that multi-syndrome standardized terminology already possesses a foundation for ICD docking—a necessary prerequisite for connecting meta-learning transfer results with Western medical diagnostic codes.

II. Three-Stage Experimental Design for Transfer Validation

Stage 1 (Baseline Anchoring): Uniformly extract Heart Qi Deficiency HRI harmonic features + symptom-tongue-pulse multimodal labels from datasets such as P3-EXP-002/011/017/020/027 from the first 33 rounds of Five Pools. Preprocess according to the P3-EXP-040 ComBat affine correction protocol. Use the P3-EXP-041 MMF-SVA meta-learning architecture (5-way 5-shot) for pre-training on four versions of Heart Qi Deficiency data (1993/2002/2017/2023). After freezing pre-training, save meta-parameters $\theta_{meta}$.

Stage 2 (Cross-Syndrome Transfer): Perform internal-loop fine-tuning of $\theta_{meta}$ on ≤ 50 entry 5-shot support sets for each of the Liver/Spleen/Lung/Kidney Qi Deficiency syndromes (5 inner-steps), then freeze and output predictions for the four syndromes. Compare against full test sets (≥ 200 entries/syndrome) for the four syndromes to calculate Cohen’s Kappa and AUC. Simultaneously run a “baseline control”—train a randomly initialized ResNet on the full training sets of the four syndromes and evaluate on the same test set. H1 passes if the difference in Cohen’s Kappa between meta-learning and baseline is ≥ +0.10.

Stage 3 (Synthetic Perturbation Sensitivity Mapping): Following perturbation modes designed in Round 31 of P3-EXP-041 (333 groups each for insert entries/modify weights/shift cut-points, totaling 1000 groups), perturb the test set. Quantify the failure rate of the meta-learner under each perturbation type; output a 2D sensitivity heatmap of “perturbation type × syndrome.” H2 passes if the insert-entry perturbation failure rate ≥ 40% (i.e., “adding new entries” is least friendly to meta-learning); H3 passes if the consistency between the sensitivity heatmap and “revision intent” (expand/tighten/reset) is ≥ 75%.

III. Synergistic Relationship with the P3-EXP-043 Hybrid Architecture

The design goal of P3-EXP-043 (adapter+prompt hybrid) is “absorbing version drift without weight updates.” P3-EXP-044 complements this: This round focuses on “weight-updatable cross-syndrome transferability,” forming a dual-axis validation space of “spatial dimension × temporal dimension.” If H1 passes, the P3-EXP-043 “no weight update” baseline can be further tightened—since Heart → Four Zang is already solved via the meta-learning adapter, only remaining “version differences” qualify for the prompt engineering layer; if H1 fails, the hybrid architecture requires redesign (with the prompt engineering layer bearing more load).

IV. Methodological Value of Synthetic Perturbation Sensitivity Mapping

If the Synthetic Perturbation Type Sensitivity Map (#124) validates H2, it directly answers “which strategy should standard revisions prioritize.” For example: If the insert-entry perturbation failure rate is 60%, modify-weight perturbation is 30%, and shift-cut-point perturbation is 40%, it guides real-world revisions as follows: If new entries are unavoidable, prioritize small incremental iterations (avoiding a single increase of ≥ 20% entries); if only severity cut-points need adjustment, the meta-learner exhibits strongest robustness and can be completed in one step; weight modification revisions are most sensitive, requiring an “incremental update guide for meta-learning adapters” attached when publishing revision guidelines.

V. Risks and Boundaries

Medium risk primarily stems from: (a) Sample volume ceiling for Four Zang Qi Deficiency—Kidney Qi Deficiency sample size in the first 33 rounds of Five Pools data remains smaller than Lung/Spleen Qi Deficiency, requiring recruitment of ≥ 200 annotated samples from the collaborative center of the PMC12711456 Kidney Qi Deficiency Core Elements Study; (b) Cross-syndrome annotation physician experience differences—Heart Qi Deficiency is annotated by cardiologists, while Kidney Qi Deficiency by nephrologists/geriatricians. Kappa must be ≥ 0.70 on a “same patient-dual-label” baseline before initiating meta-learning training; (c) Compound syndromes (Heart Qi Deficiency + Phlegm-Dampness, Heart Qi Deficiency + Blood Stasis) are not included in this round’s analysis—these are the content of #128 topics.

DOI 10.1186/s13020-026-01455-4 (2026, Chin Med, TCM explainable AI systematic review) indicates TCM AI design must meet four unique requirements (interpretability/sparsity/causality/syndrome separability). The SHAP explanation submodule of this round’s P3-EXP-044 directly interfaces with these four requirements—particularly syndrome separability, where the cross-syndrome confusion matrix itself constitutes direct evidence of syndrome separability.

Evidence and Limitations

  • Direct Evidence: 5 PMID/DOI entries regarding Heart Qi Deficiency diagnostic criteria (14727501, 19134453), Heart Qi Deficiency HRI associations (33903610, 36294440, 42305919), meta-learning cross-category transfer methodology (41008775), TCM-DS LLM instruction fine-tuning feasibility (41243110), multi-syndrome terminology international standardization (11589548), Kidney Qi Deficiency quantified screening (12711456), and TCM explainable AI design framework (10.1186/s13020-026-01455-4) constitute 8 independent evidence points.
  • Indirect Evidence: Successful transfer of MAML/ProtoNet in medical imaging (PMC12468384, 10.3390/diagnostics15182404) cannot be directly extrapolated to the “Heart → Four Zang Qi Deficiency” directory-based task; dose-effect relationships regarding Traditional Chinese Medicine diagnostic standard version drift (P3-EXP-036 Stage 1 estimated 0.85 mutual transfer weights) remain draft-level hypotheses.
  • Limitations: This round cannot answer the question of “real-world 2025+ standard revisions”—must await P3-EXP-126 post-hoc validation; does not cover pediatric/pregnancy/acute care/compound syndromes; HRI data relies on the stability of the P3-EXP-040 correction chain.

Related Reading

  • tcm-syndrome-meta-learning-standard-version-adaptation.md
  • tcm-syndrome-meta-learning-zero-shot-2025-future-version-adaptation.md
  • tcm-syndrome-meta-learning-adapter-prompt-hybrid-architecture.md
  • tcm-syndrome-diagnostic-criteria-version-mapping-hri-implementation.md
  • pulse-harmonic-syndrome-independent-correlation.md
  • pulse-harmonic-five-zang-organ-function-correlation.md

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 *