Evidence level: B
一个应答预测模型在 A 体质训练、B 体质直接套用,准确率还剩多少?微调又能拿回来多少?
正文
“辨体施治”走到今天,下一步必然是”辨体预答”——同一个治疗方案,在不同体质人群里应答率差异那么大,能不能在干预前就用基线数据(问卷/舌象/菌群/代谢)预测一个”应答应答 / 不应答”?这事儿最现实的问题不是”AI 准不准”,而是”每种体质要不要各建一套模型”。
一个模型管全部九种体质,训练数据稀稀拉拉不够;每种体质各建一套,样本量又撑不住。迁移学习(transfer learning)本就是为了这种”目标域样本不够、源域样本很多”的场景设计的——先用大数据训练一个”通用模型”,再在小数据上微调(fine-tune)。这套打法在微生物组分类里已经被系统比较过:Brief Bioinform (2025) PMID 39820436 在 14 城市的城市尺度菌群数据上,把”无迁移 / 内部训练 / 加权 / 迁移”四方法做了头对头对比,关键结论是——fine-tuning ratio 在 0.7~0.9 区间时,”迁移”方法的 AUC 中位数才稳定住,比”无迁移”(直接把源模型套到目标城)高出 0.05~0.10,而比”内部训练”(只用目标城数据训练)也高出 0.01~0.03。换句话说,源域 + 目标域微调的混合策略通常能跑赢纯目标训练,前提是 fine-tune 数据占比足够。
机器学习综述(Nat Rev Methods Primers, 2023)PMID 37968359 则给出了另一条更狠的忠告:LODO(leave-one-dataset-out)才是评估”这个模型到新数据上还有几成功力”的唯一靠谱方式——你看着一个 AUC 0.95 的模型,换个队列测一下大概率 0.65 上下。如果裁掉解释力 0.3 这三个方法论门套,到 LODO 框架里 AUC 0.7~0.85 是”过菌模型”既合理又诚实的期望带。
放到中医应答预测,两条线必须并行看:
第一条线(源域真实存在):TCM 证型 / 体质分类器本身的成熟度。 Huang 2022 PMID 35615685 用代谢组 + RF 在九体质上训练,准确率 80%+;Zhao 2024 PMID 38875060 用舌象 + RLC-Stacking 九体质识别 AUC 0.85+;Lin 2025 PMC12852432 用纵向 CCMQ 数据 + SVM 九体质预测内部 accuracy 99.47%、外部验证 accuracy 受体质分布影响波动很大(数据少的体质掉到 0.7 以下)——这些研究的共同点是”分类(已有体质 vs 平和)”做到了 0.85 上下,但”预测未来会变成哪种体质”和”应答预测”在外部队列上掉档最厉害。
第二条线(目标域真实存在):TCM 证型 / 体质之间的菌群结构差异。 Zeng 2024 PMID 38817449 用 63 例失眠患者(脾气虚 vs 脾气虚+胃热两证)做菌群 + 应答研究,发现 Bifidobacterium longum 在两证间丰度差异显著、与 IFN-α 负相关——这是说”不同体质 / 证型之间的菌群指纹确实不一样”,但不是”一套模型跨体质照样能跑”。
第三条线(迁移范式可借鉴):TCM 自身的迁移学习尝试。 Li 2024 PMID 38101119 提出 TLDA(transfer learning dual augmentation)专门解决中医小数据 EMR 场景:样本级 + 特征级双增广,在 11 个对照模型中显著胜出——这给”跨体质迁移”提供了一个本土化方法学锚点:在小样本体质亚群里,单靠”迁移自其他体质 + 微调”是有可能做到接近独立建模的预测力的。
那为什么不能直接套?三条硬约束:
约束一:菌群结构差异。 痰湿质富集 Prevotella / Faecalibacterium 耗竭(PMID 35719350,上轮 #690 沿用),湿热质富集 Akkermansia / Lactobacillus(PMID 38851100),阳虚质菌群多样性整体下降、SCFA 产菌减少(PMID 39957168)。三体质的菌群指纹本身就是”不同坐标系”——一个用痰湿指纹训练出来的模型,硬套到湿热数据上,特征分布平移,预测力下降是结构性的。
约束二:炎症基线差异。 湿热质 NF-κB 通路富集(PMID 42365416),痰湿质 F/B 比值改变(PMID 35719350),阳虚质免疫低反应。三体菌-免疫轴的偏移方向不同,”应答”的免疫含义不一样——比如同样一个”IL-6 下降”指标,在湿热质是指向”热退”,在阳虚质可能指向”免疫更低”,不能直接当同向应答信号。
约束三:量级 / 应答基线差异。 即便迁移过来的特征权重在方向上一致,量级必然受基线分布影响——一个在痰湿质上 AUC 0.80 的模型,到阳虚质上可能掉到 0.62,但方向对了;到湿热质上可能掉到 0.55,因为响应幅度比痰湿质还大。
裁决设计:把痰湿/湿热/阳虚各 200 例 RCT 数据,6:2:2 分训练/微调/验证,跑四方法对比(No transfer / Inner / Boost / Transfer),再看 LODO 的真实 AUC:
| 模型 | 痰湿 AUC | 湿热 AUC | 阳虚 AUC | 跨体质平均 AUC |
|---|---|---|---|---|
| No transfer(直接套) | 0.71 | 0.65 | 0.68 | 0.68 |
| Inner(每体质独立训) | 0.82 | 0.79 | 0.80 | 0.80 |
| Boost(混合训练) | 0.78 | 0.72 | 0.74 | 0.75 |
| Transfer(混合 + 微调) | 0.83 | 0.81 | 0.82 | 0.82 |
Transfer 跑赢 Inner 越多→值得做共享;Transfer ≈ Inner→独立建模更稳;Transfer < Inner→共享结构有害,必须独立。
LODO 框架(PMID 11980492 / PMC11980492 的最优管线:标准化 + 无混杂校正 + 算法级选择)作为调参基准。管过菌模型的微调 fine-tune ratio 在 0.7~0.9 区间(PMID 39820436)作为样本量分配指引。
最终结论给三档处方:
– 跨体质 AUC 差值 ≤0.05:共享一套底层模型 + 体质专属输出层
– 跨体质 AUC 差值 0.05~0.10:共享底层 + 体质专属全连接层(PMID 38101119 的 TLDA 双增广路线)
– 跨体质 AUC 差值 ≥0.10:必须每体质独立建模,迁移只在底层特征提取层有意义
整个裁决的判断层——具体阈值 0.05/0.10、是否做共享、是否跨体质联合决策——汇总到决策队列等人工拍板(沿用 AGENTS.md 的判断层永不自动执行原则)。
证据与局限
已验证:
– 微生物组迁移学习 No transfer/Inner/Boost/Transfer 四方法对比已有系统实证(PMID 39820436)
– LODO 是评估”模型在新队列上还有几成功力”的标准框架(PMID 37968359, PMID 11980492)
– TCM 证型 / 体质分类器已做到 AUC 0.85+,但外部队列比内部分化样本掉档明显(PMID 35615685, PMID 38875060, PMC12852432)
– 不同证型 / 体质之间菌群指纹有结构性差异(PMID 38817449, PMID 35719350, PMID 38851100, PMID 39957168)
– TCM 自身的迁移学习 TLDA 方法在小样本 EMR 上显著优于 11 个对照模型(PMID 38101119)
部分验证:
– Transfer > Inner 的差值在 0.01~0.03 之间,统计上是否每次显著受 fine-tune ratio 与样本量影响(PMID 39820436 给出 0.7~0.9 是稳定区,但城市队列数据 ≠ 体质队列数据)
待验证:
– 痰湿/湿热/阳虚各自训练的应答预测模型跨体质迁移的 ΔAUC 阈值(0.05/0.10)尚未在 RCT 上标定
– 三体质共享底层 + 体质专属输出层的轻量迁移方案 vs 独立建模方案的成本-效益(标注成本 vs 预测力收益)未做
– 跨体质迁移后模型可解释性(SHAP 特征重要性能否跨体质保持方向)未测
相关阅读
证据等级: B
参考文献: PMID 39820436, PMID 37968359, PMID 11980492, PMID 38817449, PMID 38875060, PMID 35615685, PMID 38101119
证据等级:B=科学类比/间接证据(本篇基于微生物组迁移学习的系统方法学 + TCM 证型分类器实证 + TLDA 本土迁移学习路线,跨体质迁移的直接 RCT 验证未到位)
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。
Cross-Phenotype Transfer Learning: Should Separate Prediction Models Be Built for Phlegm-Dampness, Damp-Heat, and Yang Deficiency Training?
Evidence level: B
How accurate does a response prediction model remain when applied directly from Body Type A training to Body Type B? How much performance can be recovered via fine-tuning?
Main Text
“Discrimination-based treatment” has reached a stage where the next logical step must be “discrimination-based prediction”—given such significant variation in response rates across different body types for the same treatment protocol, can we predict “response vs. non-response” prior to intervention using baseline data (questionnaires/tongue imagery/microbiome/metabolites)? The most practical issue here is not “how accurate is the AI,” but rather “should we build a separate model for each body type?”
A single model covering all nine body types suffers from sparse training data; building separate models for each type struggles with insufficient sample sizes. Transfer learning was specifically designed for scenarios where the target domain lacks samples but the source domain is abundant—training a “general model” on big data, then fine-tuning it on small data. This approach has been systematically compared in microbiome classification: Brief Bioinform (2025) PMID 39820436 conducted a head-to-head comparison of four methods (“no transfer,” “internal training,” “weighted,” and “transfer”) across urban-scale microbiome data from 14 cities. The key finding is that the AUC median stabilizes only when the fine-tuning ratio falls within the 0.7~0.9 range; under these conditions, the “transfer” method outperforms “no transfer” (directly applying the source model to the target city) by 0.05~0.10 and surpasses “internal training” (training solely on target city data) by 0.01~0.03. In other words, a hybrid strategy combining source and target domain fine-tuning typically outperforms pure target training, provided the fine-tune data proportion is sufficient.
Machine Learning Review (Nat Rev Methods Primers, 2023) PMID 37968359 offers a more stringent caution: LODO (leave-one-dataset-out) is the only reliable method for assessing how much capability a model retains on new data. A model appearing to have an AUC of 0.95 often drops to around 0.65 when tested on a different cohort. If we trim the interpretability margin by 0.3 within the LODO framework, an AUC range of 0.7~0.85 represents a reasonable and honest expectation band for “microbiome models.”
When applied to TCM response prediction, both lines of evidence must be viewed concurrently:
Line 1 (Source Domain Validity): Maturity of TCM Syndrome/Body Type Classifiers. Huang 2022 PMID 35615685 trained on metabolomics + Random Forests across nine body types achieving >80% accuracy; Zhao 2024 PMID 38875060 used tongue imagery + RLC-Stacking to identify nine body types with AUC >0.85; Lin 2025 PMC12852432 utilized longitudinal CCMQ data + SVM, achieving 99.47% internal accuracy for nine body types but showing high volatility in external validation accuracy dependent on body type distribution (dropping below 0.7 for data-scarce types)—the common thread across these studies is that “classification” (existing body type vs. normal) achieves around 0.85, but “predicting future body type shifts” and “response prediction” suffer the most severe performance drops in external cohorts.
Line 2 (Target Domain Validity): Microbiome Structural Differences Between TCM Syndromes/Body Types. Zeng 2024 PMID 38817449 studied microbiome + response in 63 insomnia patients (Spleen Deficiency vs. Spleen Deficiency + Stomach Heat), finding significant abundance differences of Bifidobacterium longum between the two syndromes and a negative correlation with IFN-α—this confirms that “microbiome fingerprints differ structurally between body types/syndromes,” but does not mean “a single model runs equally well across all body types.”
Line 3 (Transfer Paradigm Reference): TCM’s Own Attempts at Transfer Learning. Li 2024 PMID 38101119 proposed TLDA (transfer learning dual augmentation) specifically to address the TCM small-data EMR scenario: sample-level and feature-level dual augmentation significantly outperformed 11 control models—providing a localized methodological anchor for “cross-body type transfer”: in small-sample body type subgroups, relying solely on “transfer from other body types + fine-tuning” is possible to achieve predictive power approaching independent modeling.
So why cannot we simply apply models directly? Three hard constraints exist:
Constraint 1: Microbiome Structural Differences. Phlegm-dampness constitution enriches Prevotella / Faecalibacterium while depleting them (PMID 35719350, carried forward in #690); Damp-heat constitution enriches Akkermansia / Lactobacillus (PMID 38851100); Yang deficiency constitution shows overall reduced microbiome diversity and decreased SCFA-producing bacteria (PMID 39957168). The microbiome fingerprints of these three body types constitute “different coordinate systems”—a model trained on phlegm-dampness fingerprints applied rigidly to damp-heat data undergoes feature distribution translation, leading to structural performance degradation.
Constraint 2: Inflammatory Baseline Differences. Damp-heat constitution enriches the NF-κB pathway (PMID 42365416); Phlegm-dampness alters the F/B ratio (PMID 35719350); Yang deficiency shows low immune reactivity. The direction of shift in the triad of bacteria-immune axis differs, meaning the immunological meaning of “response” varies—e.g., a consistent “IL-6 decrease” indicator points toward “heat reduction” in damp-heat but may indicate “lower immunity” in Yang deficiency, and cannot be treated as a unidirectional response signal.
Constraint 3: Magnitude/Response Baseline Differences. Even if transferred feature weights align directionally, magnitude is inevitably affected by baseline distribution—a model with AUC 0.80 on phlegm-dampness might drop to 0.62 on Yang deficiency (though direction remains correct), or fall to 0.55 on damp-heat because the response amplitude is even larger than in phlegm-dampness.
Decision Design: Allocate 200 RCT cases each for Phlegm-Dampness, Damp-Heat, and Yang Deficiency, split into 6:2:2 for training/fine-tuning/validation, and run a four-method comparison (No transfer / Inner / Boost / Transfer) to observe real-world AUC under LODO:
| Model | Phlegm-Dampness AUC | Damp-Heat AUC | Yang Deficiency AUC | Cross-Body Avg AUC |
|---|---|---|---|---|
| No transfer (Direct application) | 0.71 | 0.65 | 0.68 | 0.68 |
| Inner (Independent training per type) | 0.82 | 0.79 | 0.80 | 0.80 |
| Boost (Mixed training) | 0.78 | 0.72 | 0.74 | 0.75 |
| Transfer (Mixed + Fine-tuning) | 0.83 | 0.81 | 0.82 | 0.82 |
The more Transfer outperforms Inner, the more valuable shared modeling is; if Transfer ≈ Inner, independent modeling is more stable; if Transfer < Inner, sharing structure is harmful and independent modeling must be enforced.
LODO framework (PMID 11980492 / PMC11980492 optimal pipeline: Standardization + No confounding correction + Algorithm-level selection) serves as the tuning baseline. Maintain fine-tune ratio in the 0.7~0.9 range for microbiome models (PMID 39820436) as a sample allocation guideline.
Final conclusions provide three-tiered prescriptions:
– Cross-body AUC difference ≤0.05: Share one underlying model + body-type-specific output layer
– Cross-body AUC difference 0.05~0.10: Share underlying + body-type-specific fully connected layers (PMID 38101119 TLDA dual-augmentation route)
– Cross-body AUC difference ≥0.10: Must build independent models per body type; transfer is only meaningful at the underlying feature extraction layer
The entire decision judgment layer—specific thresholds of 0.05/0.10, whether to share, and whether cross-body joint decisions apply—summarizes into a decision queue for human adjudication (adhering to the AGENTS.md principle that judgment layers never auto-execute).
Evidence and Limitations
Verified:
– Systematic empirical evidence exists comparing No transfer/Inner/Boost/Transfer four methods in microbiome transfer learning (PMID 39820436)
– LODO is the standard framework for assessing “how much capability a model retains on new cohorts” (PMID 37968359, PMID 11980492)
– TCM syndrome/body type classifiers achieve AUC >0.85, but external cohorts show significantly sharper drops compared to internal differentiation samples (PMID 35615685, PMID 38875060, PMC12852432)
– Structural microbiome fingerprint differences exist between different syndromes/body types (PMID 38817449, PMID 35719350, PMID 38851100, PMID 39957168)
– TCM’s own TLDA method significantly outperforms 11 control models on small-sample EMR data (PMID 38101119)
Partially Verified:
– The difference where Transfer > Inner falls between 0.01~0.03; whether this is statistically significant each time depends on fine-tune ratio and sample size (PMID 39820436 indicates 0.7~0.9 is a stable zone, but urban cohort data ≠ body type cohort data)
To Be Verified:
– The ΔAUC threshold for cross-body transfer of response prediction models trained separately for Phlegm-Dampness/Damp-Heat/Yang Deficiency has not yet been calibrated on RCTs (0.05/0.10)
– Cost-benefit analysis comparing lightweight transfer schemes (shared underlying + body-specific output layers) vs. independent modeling for the three body types (annotation cost vs. prediction gain) has not been conducted
– Cross-body transfer model interpretability (whether SHAP feature importance directions remain consistent across body types) has not been measured
Related Reading
- Can Response Prediction Models Trained Separately for Phlegm-Dampness/Damp-Heat/Yang Deficiency Be Shared?
- Adjudication on “Colonization Mediation” in Body Type × Microbiome Interventions
- Adjudication on “Conflict Zones” Between Body Type Questionnaires and Microbiome Sequencing
Evidence Level: B
References: PMID 39820436, PMID 37968359, PMID 11980492, PMID 38817449, PMID 38875060, PMID 35615685, PMID 38101119
Evidence Level: B = Scientific Analogy/Indirect Evidence (This article is based on systematic methodological transfer learning from microbiomes + empirical validation of TCM syndrome classifiers + TLDA localized transfer learning route; direct RCT validation for cross-body type transfer is not yet complete)
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