Evidence level: B
导言
34 轮元学习适配器在四脏气虚(心/肝/脾/肺/肾)零样本迁移 P2-EXP-EM-044 已落地;#32 轮 MMF-SVA 在 1993/2002/2017/2023 四版辨证标准的零样本外推 P3-EXP-042 也已成型。两条链路共同留下了最后一个关键盲点:训练分布之外的”扩展证型”是否能被元学习器识别。
本轮以心阴虚(heart-yin deficiency)与肾阳虚(kidney-yang deficiency)作为「未训练过的目标域」,检验元学习适配器在虚证体系内跨病机维度(气虚 → 阴虚/阳虚)外推的能力边界。这一检验比版本外推更深一层:版本外推是”条目权重变化”(同一证型),证型外推是”病机轴变化”(不同证型),要求元学习器从「证型族」层面抽象出可迁移的判别模式,而非过拟合到心气虚的具体谐波指纹。
正文
1. 为什么选心阴虚与肾阳虚
在中医辨证体系中,气虚/阴虚/阳虚构成三轴:气虚以功能衰减为主,阴虚以津液不足为主,阳虚以温煦失职为主。心阴虚与肾阳虚分别代表了”心系虚证”的阴虚极化与”肾系虚证”的阳虚极化,与心气虚在症状条目、HRI 谐波通道(PMID 33903610, 36294440)上的重叠率估计 < 20%(以现行 2002 / 2017 / 2023 标准条目逐条勾选衡量)。这是验证元学习器”病机轴外推”的最低准入门槛:既不能太相似(沦为版本外推的特例),也不能太异质(沦为不可学)。心阴虚与心气虚属同一器官系,便于把器官先验约束为可解释变量;肾阳虚与心气虚分属不同器官系,便于把跨器官迁移当作系统结构挑战。
2. 元学习器的形式化重构
延续 P3-EXP-041/044 的 MMF-SVA(Multi-Modal Fusion with Syndrome-Version Adapter)框架,本轮新增两个结构化组件:
S1 病机轴编码器(Pathogenesis Axis Encoder, PAE):把辨证条目按”气虚 / 血虚 / 阴虚 / 阳虚 / 痰湿 / 血瘀”六轴做硬聚类编码。训练域(心气虚 + 四脏气虚)只覆盖”气虚”轴;目标域(心阴虚 / 肾阳虚)首次覆盖”阴虚”与”阳虚”两轴。PAE 把条目结构从 one-hot 升级为多标签分布,使元学习器能识别”症状在病机轴上的坐标”而非仅记忆条目 ID。
S2 跨器官原型网络(Cross-Zang Prototypical Network, CZPN):把心气虚训练得到的「心-气虚原型向量」与四脏气虚的「肝/脾/肺/肾-气虚原型向量」联合为”五脏-气虚原型集合”。在零样本外推时,识别”心-阴虚”的最近邻不是「心-气虚原型」,而是从”心-气虚原型”沿”心-阴虚轴向量”位移后的虚拟原型(参照 #33 轮 adapter-prompt hybrid 的提示位移思想)。肾阳虚则从”肾-气虚原型”沿”阳虚轴向量”位移。这一”原型位移”思路在 MAML 类方法学综述(PMID 38464210)已有原型网络可塑性的形式化基础。
3. 训练-测试域分割的最小可行性
训练域(seen):心气虚(4 标准版本 × 200 例 = 800 例)+ 肝/脾/肺/肾气虚(各版本 200 例 × 4 脏 × 4 标准 = 3200 例,来源 P3-EXP-044 阶段 2),合计 4000 例。
目标域(unseen):
– 心阴虚 200 例(按 2023 年版 ZY/T 001 标准由 ≥ 2 名高年资医师盲法标注)
– 肾阳虚 200 例(同上)
– 合计 400 例,全部 0-shot(不进入任何梯度更新)。
对照域(baseline):
– 直接 fine-tune 50-shot 心阴虚→肾阳虚(基线 1)
– 随机权重初始化从头训练(基线 2)
– MAML 在 seen 域做 5-shot meta-train,目标域 0-shot(基线 3,文献成熟外推对照)
4. 假设与可证伪条件
- H1(病机轴可迁移):MMF-SVA-PAECZPN 在心阴虚 0-shot 上的 macro-F1 ≥ MAML 基线 0.75 倍,且 Cohen’s Kappa ≥ 0.6。
- H2(跨器官可迁移):MMF-SVA-PAECZPN 在肾阳虚 0-shot 上的 macro-F1 ≥ 心阴虚 0.85 倍(允许跨器官衰减但不得 < 0.6)。
- H3(病机轴不可迁移):若 H1/H2 任一不成立,则否定 PAE + CZPN 的外推能力假设,需退回到 per-syndrome 微调或 per-axis 独立训练;同时把心阴虚/肾阳虚作为新一轮 P3-EXP 单独建库的独立课题。
5. 评价指标与统计
主指标:macro-F1(解决证候样本不均衡)、Cohen’s Kappa(标注一致性等价)、AUROC(与现有 P3-EXP-041/044 一致便于 meta-comparison)。
次指标:Hamming loss(多标签分解误差)、Brier score(概率校准)、推理延迟(端侧部署可行性)。
统计框架:每个 0-shot 实验重复 5 随机种子取 95% CI;配对 t 检验(MMF-SVA-PAECZPN vs 三个基线);Bonferroni 校正三基线比较;多假设检验以 Holm-Bonferroni 阶梯控制 FWER ≤ 0.05。
6. HRI 谐波通道映射的器官先验
参考 #22-#30 轮背书的五脏谐波文献(PMID 33903610, 36294440, 42305919):
- 心:HRI C2/C3 谐波主导(心气虚文献成熟)
- 肾:HRI C7/C8 谐波主导(肾气虚文献较薄但有间接证据)
- 心阴虚:理论上应在 C2/C3 主导基础上叠加 C6/C9(阴虚火旺-交感激活代理)
- 肾阳虚:理论上应在 C7/C8 主导基础上叠加 C1/C2(温煦失职-副交感衰减代理)
以上谐波通道映射是理论假设而非已证实证据,仅作为元学习器的输入特征空间设计依据;不进入结论。
7. 失败案例归因预案
参考 #34 轮末计划 #134 失败归因,本轮设置三类失败模式归因:
1. PAE 病机轴编码器失效(条目 → 病机轴映射误差)
2. CZPN 跨器官原型漂移(位移向量估计偏差)
3. HRI 数据校正失败(依赖 P3-EXP-040 ComBat 仿射校正链)
归因结果反向指导下一轮:若 1 类失败主导则 PAE 退回到 per-syndrome soft-label;若 2 类失败主导则 CZPN 改为五脏独立原型集合不做跨器官位移;若 3 类失败主导则需先回退到 P3-EXP-040 重做。
证据与局限
支持证据:MAML/ProtoNet 在医学影像跨域少样本学习已成熟(PMID 38464210 综述);大模型 LoRA PEFT 在中医 LLM(NSPL 2024 / TCM-SD 等)显示跨任务可塑性(PMID 41243110 综述 + 间接医学 LLM 应用 PMID 39123298);HRI 五脏谐波独立相关性证据(PMID 33903610, 36294440, 42305919)支持器官先验可用;辨证标准漂移(PMID 14727501 / 19134453)证明”训练域条目”与”目标域条目”非完全独立。
局限:心阴虚/肾阳虚样本量 400 例属小样本研究天花板;PAE 六轴聚类依赖中医师共识;CZPN 跨器官位移向量假设未在脏气虚训练中学到(仅见过同器官气虚);HRI 谐波通道先验是理论外推非已证实;本轮无法覆盖阴虚火旺/阳虚水泛等更细分亚型;合成扰动无法代表真实未训练证型的全部变异;本轮只检验”虚证内部”外推,”实证(痰湿/血瘀)外推”留待下一轮。
相关阅读
- P3-EXP-041 心气虚四版标准元学习(P3-EXP-041 阶段 2-3)
- P3-EXP-042 元学习零样本 2025+ 未来版本外推(#32 轮)
- P3-EXP-043 元学习 + adapter-prompt 混合架构(#33 轮)
- P3-EXP-044 心气虚→四脏气虚迁移(#34 轮)
- P3-EXP-040 改良 ComBat 仿射校正在 HRI 多中心多设备矩阵(#29 轮)
-
36 轮待办:P3-EXP-036 跨标准互转映射实证回打分
-
38-#42 轮待办:方法学联级扩展(若本轮 H1/H2 通过)
字数:约 1450 字
PMID 引用数:8 个(核心证据 4 条:38464210, 41243110, 39123298, 33903610;扩展证据 4 条:36294440, 42305919, 14727501, 19134453)
审查状态:draft — 待人工 intake 后方能进入 779。
注:本轮 SearXNG 直接搜索 “meta-learning + TCM syndrome” 与 “yin/yang deficiency diagnostic criteria” 返回结果被搜索引擎歧义过滤为 Meta/Facebook/TCM 电视台等无关噪声;本轮证据骨架完全由前几轮背书 PMID 与元学习/医学影像综述组成,符合”宁可补证据边界、反证、实验设计,也不要为了产量扩题或把假说写成结论”的五池原则。
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。
Meta-Learning Adapters for Zero-Shot Extrapolation to Untrained Syndrome Extensions of Heart Yin Deficiency and Kidney Yang Deficiency
Evidence level: B
Introduction
The #34 round meta-learning adapters for four-organ qi-deficiency (heart/liver/spleen/lung/kidney) have been successfully deployed in zero-shot transfer P2-EXP-EM-044; similarly, the #32 round MMF-SVA has matured for zero-shot extrapolation across four versions of diagnostic criteria (1993/2002/2017/2023) in P3-EXP-042. However, these two pathways leave a final critical blind spot: Can the meta-learner recognize “extended syndrome patterns” outside the training distribution?
This round uses heart-yin deficiency (heart-yin deficiency) and kidney-yang deficiency (kidney-yang deficiency) as “untrained target domains” to test the extrapolation capability boundaries of meta-learning adapters within the deficiency syndrome framework across pathogenesis dimensions (qi-deficiency → yin/yang deficiency). This test goes deeper than version extrapolation: version extrapolation involves “item weight shifts” (same syndrome), whereas syndrome extrapolation involves “pathogenesis axis shifts” (different syndromes). It requires the meta-learner to abstract transferable discriminative patterns at the “syndrome family” level rather than overfitting to specific harmonic fingerprints of heart qi-deficiency.
Main Body
1. Rationale for Selecting Heart-Yin and Kidney-Yang Deficiency
In the TCM diagnostic system, qi-deficiency/yin-deficiency/yang-deficiency form a tri-axis: qi-deficiency primarily reflects functional attenuation, yin-deficiency reflects fluid volume depletion, and yang-deficiency reflects failed warming function. Heart-yin deficiency and kidney-yang deficiency represent the yin-deficiency polarization of “heart system deficiencies” and the yang-deficiency polarization of “kidney system deficiencies,” respectively. The estimated overlap rate in symptom items and HRI harmonic channels (PMID 33903610, 36294440) with heart qi-deficiency is < 20% (measured by item-by-item selection against current 2002/2017/2023 standards). This establishes the minimum threshold for validating meta-learner “pathogenesis axis extrapolation”: it must not be too similar (becoming a trivial case of version extrapolation) nor too heterogeneous (becoming unlearnable). Heart-yin deficiency and heart qi-deficiency share the same organ system, facilitating the use of organ priors as interpretable variables; kidney-yang deficiency and heart qi-deficiency belong to different organ systems, facilitating the treatment of cross-organ migration as a structural challenge.
2. Formal Reconstruction of the Meta-Learner
Building on the MMF-SVA (Multi-Modal Fusion with Syndrome-Version Adapter) framework from P3-EXP-041/044, this round introduces two structured components:
S1 Pathogenesis Axis Encoder (PAE): Clusters diagnostic items into six hard clusters based on “qi-deficiency / blood-deficiency / yin-deficiency / yang-deficiency / phlegm-dampness / blood-stasis.” The training domain (heart qi-deficiency + four-organ qi-deficiencies) covers only the “qi-deficiency” axis; the target domain (heart-yin / kidney-yang) covers the “yin-deficiency” and “yang-deficiency” axes for the first time. PAE upgrades item structure from one-hot to multi-label distributions, enabling the meta-learner to identify “symptom coordinates on the pathogenesis axis” rather than merely memorizing item IDs.
S2 Cross-Zang Prototypical Network (CZPN): Jointly constructs a “five-organ qi-deficiency prototype set” by combining the “heart-qiyang deficiency prototype vector” derived from training with “liver/spleen/lung/kidney-qiyang deficiency prototype vectors.” During zero-shot extrapolation, the nearest neighbor for “heart-yin deficiency” is not the “heart-qiyang prototype” but a virtual prototype displaced along the “heart-yin axis vector” from the “heart-qiyang prototype” (referencing the prompt displacement idea in #33 round adapter-prompt hybrid). Kidney-yang deficiency is displaced along the “yang-deficiency axis vector” from the “kidney-qiyang prototype.” This “prototype displacement” approach has formalized foundations in prototype network plasticity within MAML methodological reviews (PMID 38464210).
3. Minimal Feasible Training-Test Domain Split
Training Domain (seen): Heart qi-deficiency (4 versions × 200 cases = 800 cases) + Liver/Spleen/Lung/Kidney qi-deficiency (200 cases × 4 organs × 4 versions = 3,200 cases from P3-EXP-044 Stage 2), totaling 4,000 cases.
Target Domain (unseen):
– Heart-yin deficiency: 200 cases (blindly annotated by ≥2 senior physicians per 2023 ZY/T 001 standard)
– Kidney-yang deficiency: 200 cases (same method)
– Total 400 cases, all 0-shot (no gradient updates).
Baseline Domain:
– Direct fine-tuning: 50-shot heart-yin → kidney-yang (Baseline 1)
– Random weight initialization from scratch (Baseline 2)
– MAML: 5-shot meta-training on seen domain, 0-shot on target domain (Baseline 3, literature-based mature extrapolation control).
4. Hypotheses and Falsifiability Conditions
- H1 (Pathogenesis Axis Transferable): MMF-SVA-PAECZPN achieves macro-F1 ≥ 0.75 × MAML baseline on heart-yin 0-shot, with Cohen’s Kappa ≥ 0.6.
- H2 (Cross-Organ Transferable): MMF-SVA-PAECZPN achieves macro-F1 on kidney-yang 0-shot ≥ 0.85 × heart-yin performance (allowing cross-organ attenuation but not < 0.6).
- H3 (Pathogenesis Axis Non-Transferable): If H1 or H2 fails, the extrapolation capability assumption for PAE + CZPN is negated; a retreat to per-syndrome fine-tuning or per-axis independent training is required. Heart-yin and kidney-yang should then be treated as independent topics for new P3-EXP database construction.
5. Evaluation Metrics and Statistics
Primary Metrics: macro-F1 (addressing class imbalance), Cohen’s Kappa (equivalent to annotation consistency), AUROC (consistent with P3-EXP-041/044 for meta-comparison).
Secondary Metrics: Hamming loss (multi-label decomposition error), Brier score (probability calibration), inference latency (edge deployment feasibility).
Statistical Framework: Each 0-shot experiment repeats with 5 random seeds to calculate 95% CI; paired t-tests (MMF-SVA-PAECZPN vs. three baselines); Bonferroni correction for three baseline comparisons; Holm-Bonferroni step-down control for multiple hypothesis testing to ensure FWER ≤ 0.05.
6. Organ Priors via HRI Harmonic Channel Mapping
Referencing five-organ harmonic literature validated in rounds #22–#30 (PMID 33903610, 36294440, 42305919):
- Heart: HRI C2/C3 harmonics dominate (mature literature for heart qi-deficiency).
- Kidney: HRI C7/C8 harmonics dominate (kidney qi-deficiency literature is thinner but supported by indirect evidence).
- Heart-Yin: Theoretically should superimpose C6/C9 on top of C2/C3 dominance (yin-fire excess → sympathetic activation proxy).
- Kidney-Yang: Theoretically should superimpose C1/C2 on top of C7/C8 dominance (failed warming → parasympathetic attenuation proxy).
These harmonic channel mappings are theoretical hypotheses rather than confirmed evidence, serving only as input feature space design guidelines for the meta-learner and not entering the final conclusions.
7. Failure Case Attribution Plan
Referencing failure attribution plan #134 from the end of round #34, this round sets up three failure mode attributions:
1. PAE Pathogenesis Axis Encoder failure (mapping error from items to pathogenesis axis).
2. CZPN Cross-Organ Prototype Drift (displacement vector estimation bias).
3. HRI Data Correction Failure (relying on P3-EXP-040 ComBat affine correction chain).
Attribution results will guide the next round: if Type 1 dominates, revert PAE to per-syndrome soft labels; if Type 2 dominates, modify CZPN to use independent five-organ prototype sets without cross-organ displacement; if Type 3 dominates, revert to P3-EXP-040 for re-execution.
Evidence and Limitations
Supporting Evidence: MAML/ProtoNet maturity in cross-domain few-shot learning for medical imaging (PMID 38464210 review); LLM LoRA PEFT cross-task plasticity in TCM LLMs (NSPL 2024 / TCM-SD etc.) (PMID 41243110 review + indirect medical LLM applications PMID 39123298); independent HRI five-organ harmonic correlation evidence (PMID 33903610, 36294440, 42305919) supporting organ priors; diagnostic criterion drift (PMID 14727501 / 19134453) proving training and target item sets are not fully independent.
Limitations: Sample size of 400 cases for heart-yin/kidney-yang represents the small-sample research ceiling; PAE six-axis clustering relies on TCM physician consensus; CZPN cross-organ displacement vectors are not learned during zang qi-deficiency training (only same-organ qi-deficiency observed); HRI harmonic channel priors are theoretical extrapolations not yet confirmed; this round does not cover finer subtypes like yin-fire excess or yang-water flooding; synthetic perturbations cannot represent all variations of real untrained syndromes; this round only tests extrapolation “within deficiencies,” leaving “real syndromes (phlegm-dampness/blood-stasis) extrapolation” for the next round.
Related Reading
- P3-EXP-041 Heart Qi-Deficiency Four-Version Standard Meta-Learning (P3-EXP-041 Stage 2-3)
- P3-EXP-042 Meta-Learning Zero-Shot 2025+ Future Version Extrapolation (#32 Round)
- P3-EXP-043 Meta-Learning + Adapter-Prompt Hybrid Architecture (#33 Round)
- P3-EXP-044 Heart Qi-Deficiency → Four Organ Qi-Deficiency Migration (#34 Round)
- P3-EXP-040 Improved ComBat Affine Correction in HRI Multi-Center Multi-Device Matrix (#29 Round)
-
36 Round To-Do: Empirical Rescore of Cross-Standard Mutual Conversion Mapping (P3-EXP-036)
-
38–#42 Round To-Do: Methodological Hierarchical Extension (if H1/H2 pass)
Word Count: ~1450 words
PMID Citation Count: 8 (Core Evidence: 38464210, 41243110, 39123298, 33903610; Extended Evidence: 36294440, 42305919, 14727501, 19134453)
Review Status: draft — pending human intake before entry into 779.
Note: Direct SearXNG searches for “meta-learning + TCM syndrome” and “yin/yang deficiency diagnostic criteria” returned results filtered by search engines as irrelevant noise (Meta/Facebook/TCM TV etc.). This round’s evidence skeleton relies entirely on PMIDs from previous rounds and meta-learning/medical imaging reviews, adhering to the “five-pool principle”: supplement evidence boundaries, refute, design experiments, avoid expanding topics for output volume, and do not write hypotheses as conclusions.
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