Evidence level: B
本文属研究原型设计讨论,不构成诊断、治疗或健康建议;文中”建议生成”指机器学习研究管线的一个待验证环节。
一句话结论
“反事实解释”(counterfactual explanation)看起来是给个性化建议量身定做的工具——
模型说”如果把某特征改到某值,预测就会翻转”。但最小改动反事实天然偏向
模型最敏感的特征,而最敏感的特征往往正是测得最不准的特征。不加约束的
建议闭环,会系统性地建议人去调节”最测不准的量”。
背景:从热力图到建议的一步之差
前两轮已完成两块基石:#74 用 SHAP 热力图对照经络理论,#76 立了”测量链分辨率
下限”——同一测量链的噪声底会让跨设备一致性指标摆动 0.2 以上,且热力图漏报
截距项 φ₀ 时会出现”方向对、幅度全错”的静默失败。#77 想走的最后一步是:
反事实扰动 → HRI 调节目标 → 自动生成养生建议。
这一步恰好踩在前面两块基石的接缝上。反事实的最小改动路径 = 分数面梯度
最大的方向 = SHAP 值最大的特征;而脉搏谐波里 SHAP 敏感度最高的特征,与
噪声底最大的特征,来自同一条测量通道。敏感度与噪声同源,这不是巧合,
是同一物理量的两副面孔。
四道闸门(每道都可证伪)
第一道:方向一致性。 反事实说”特征 A 调高会改善目标”,必须回头查训练
数据:A 与目标的经验相关方向真是这样吗?模型可能学到了伪相关——反事实
沿着伪相关走,建议就是错的,而且错得很有说服力。检验时必须报随机基线:
符号一致的零参数基线是 0.5,”15 条里 10 条一致”离抓阄只有 1.33 倍。
第二道:重打分验证。 SHAP 的加性和只是局部近似。反事实承诺的预测变化
必须用原模型重新打分来验证——承诺变化与实测变化比值低于 0.7 的,该条
反事实作废。这一道专治”图好看但外推失效”。
第三道:可行性白名单。 特征分三层:可干预(起居、呼吸、穴位刺激)、
不可干预(年龄、性别)、测量伪迹(设备标定相关)。后两层禁止出现
在建议里。可行性约束的严格表述已有成熟框架(Ustun et al. 2019 的
actionable recourse),不是事后过滤,而是硬约束优化。
第四道:步长下限。 建议的调节量若小于设备噪声底,做了也测不出变化——
建议在物理上不可执行。这类特征只能输出”有关联但当前分辨率不足以监测”。
四道全过才允许生成建议。第一道失败是整线终止信号:若模型反事实方向
与经验方向的一致率只剩 0.6(基线 0.5 附近),说明整套建议建立在伪相关上,
继续投入就是往漏的桶里加水。
为什么阈值 0.7 不放在 0.5 上
第一道闸门的终止线设 0.7,而不是”及格线” 0.5。原因与 R27 的老结论一致:
止损闸门的两侧代价不对称。一致率在 0.5–0.7 之间的模型,看着”比随机好一点”,
但用它生成健康建议的期望损失(错误建议进入闭环)远大于收益;把终止线
挪到 0.7,误杀的是”勉强可用”的模型,保住的是”不输出建立在沙上的建议”。
多保留的 15% 检出力由后续数据自然兜底,错误的健康建议没有兜底。
证据边界
本文是设计讨论,无任何实测。四道闸门的阈值全部是预注册值,待既有数据
(约 300 条 HRI 记录)上跑通后才能评估。文献锚点:Cao 等 2026(PMID
42548561)已把 SHAP+ICE 轨迹用于术前鉴别诊断的反事实推理,Alalawi 2026
(PMID 42510061)测了五种模型间反事实解释的一致性——两者都证明这类方法
在医学场景可行,但都停在”解释”层,没有一条进入”建议生成”层。
我们找不到任何一条直接研究”脉诊谐波反事实建议的可行性约束”的文献,
如实标证据等级 B。
待验证
① 既有数据上四道闸门的基线通过率(阶段 0,纯本地);② 方向一致率若落在
0.6–0.7 之间,是否有分层特征(按谐波阶次、按采集批次)能挽救部分建议;
③ 跨模型方向一致性(五种模型投票)能否替代单模型经验方向检验。
免责声明
本文是研究原型设计讨论,所有阈值均为预注册值,不构成医疗建议,不能替代执业医师的判断与诊疗;文中出现的”建议生成”仅指机器学习研究管线的一个待验证环节,不针对任何具体个人的健康状况。如有健康问题,请及时就医并咨询专业医疗人员。
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。
Counterfactual Advice Generation Actionability Constraints
Evidence level: B
This article is a design discussion and does not constitute diagnosis, treatment, or health advice; the term “recommendation generation” refers to an unverified stage in the machine learning research pipeline.
One-Sentence Conclusion
“Counterfactual explanations” appear to be a tool tailored for personalized recommendations—models stating “if feature A were adjusted to value X, the prediction would flip.” However, minimum-change counterfactuals inherently bias toward the model’s most sensitive features, and the most sensitive features are often those measured with the least accuracy. An unconstrained recommendation loop would systematically advise individuals to adjust “the quantities measured least accurately.”
Background: The Gap Between Heatmaps and Recommendations
The previous two rounds established two foundational elements: #74 used SHAP heatmaps to cross-reference with meridian theory, and #76 established the “measurement chain resolution floor”—where noise floors within the same measurement chain cause cross-device consistency metrics to swing by over 0.2, and omitting the intercept term φ₀ in heatmaps results in “correct direction but completely wrong magnitude” silent failures. #77 aims for the final step: counterfactual perturbation → HRI adjustment targets → automatically generated health recommendations.
This step precisely bridges the gap between the two foundational elements. The minimum-change path of counterfactuals equals the direction of steepest ascent on the score surface, which corresponds to the feature with the largest SHAP value; meanwhile, the feature in the pulse harmonic with the highest SHAP sensitivity and the feature with the largest noise floor originate from the same measurement channel. Sensitivity and noise share the same origin; this is not a coincidence but two faces of the same physical quantity.
Four Gates (Each Falsifiable)
Gate 1: Directional Consistency. Counterfactuals state “adjusting feature A upward will improve the target”; one must verify against training data: does the empirical correlation direction between A and the target truly align? The model may have learned spurious correlations—following counterfactuals along spurious paths yields incorrect recommendations that sound plausible. Validation must report random baselines: the zero-parameter baseline with consistent signs is 0.5, and “10 out of 15 consistent” is only 1.33 times better than random chance.
Gate 2: Rescore Verification. SHAP’s additive and local approximations are limited. The predicted change promised by counterfactuals must be verified by rescoreing with the original model—the ratio of promised change to actual measured change below 0.7 invalidates that counterfactual. This gate specifically addresses “the heatmap looks good but extrapolation fails.”
Gate 3: Feasibility Whitelist. Features are stratified into three layers: actionable (lifestyle, breathing, acupoint stimulation), non-actionable (age, gender), and measurement artifacts (device calibration-related). The latter two layers must not appear in recommendations. The strict formulation of feasibility constraints has established frameworks (Ustun et al. 2019’s actionable recourse); this is not post-hoc filtering but hard-constraint optimization.
Gate 4: Step Size Floor. If the recommended adjustment magnitude is smaller than the device noise floor, the change cannot be detected—such recommendations are physically unexecutable. These features can only output “associated but current resolution insufficient for monitoring.”
All four gates must pass before generating recommendations. Failure at Gate 1 is the full-line termination signal: if the consistency rate between model counterfactual direction and empirical direction drops to 0.6 (near the baseline of 0.5), it indicates the entire recommendation system rests on spurious correlations; continuing investment is like adding water to a leaking bucket.
Why Threshold 0.7, Not 0.5
Gate 1’s termination line is set at 0.7 rather than the “passing grade” of 0.5. This aligns with R27’s old conclusion: the costs on either side of the stop-loss gate are asymmetric. Models with consistency rates between 0.5–0.7 appear “better than random,” yet the expected loss (incorrect recommendations entering the loop) far exceeds the gain when generating health advice; shifting the termination line to 0.7 sacrifices “marginally usable” models to preserve “no recommendations built on sand.” The extra 15% sensitivity retained is naturally covered by subsequent data, whereas erroneous health recommendations lack such coverage.
Evidence Boundaries
This article is a design discussion with no empirical testing. Thresholds for all four gates are pre-registered values, to be evaluated once run on existing data (approximately 300 HRI records). Literature anchors: Cao et al. 2026 (PMID 42548561) applied SHAP+ICE trajectories for counterfactual reasoning in preoperative differential diagnosis, and Alalawi 2026 (PMID 42510061) measured consistency of counterfactual explanations across five models—both demonstrate feasibility in medical contexts, but both stop at the “explanation” layer with none reaching the “recommendation generation” layer. No direct study on “feasibility constraints for pulse diagnosis harmonic counterfactual recommendations” exists; evidence level marked B.
Pending Verification
① Baseline pass rates of the four gates on existing data (Stage 0, purely local); ② If directional consistency falls between 0.6–0.7, whether stratified features (by harmonic order, by acquisition batch) can salvage partial recommendations; ③ Whether cross-model directional consistency (five-model voting) can replace single-model empirical direction validation.
Disclaimer
This article is a discussion of a research prototype; all thresholds are pre-registered values, and it does not constitute medical advice or replace the judgement and diagnosis of a licensed physician. “Recommendation generation” here refers only to a to-be-verified stage of a machine-learning research pipeline and does not apply to any specific individual’s health condition. If you have health concerns, please seek medical care and consult a qualified clinician.
Leave a Reply