命门通道:本病的事实,跨病的幻觉——四诊消融排序的可搬运性检验

Written by

in

Language

Evidence level: C

在某一个病种里,”拿掉哪一路跌得最惨”是可以稳定重复的硬事实;但把这个排序搬到另一个病种上,它几乎立刻失效——而且失效的方式不是”稍微不准”,而是连第一名换谁都不一定。

一、临床经验的原始形态

四诊合参支持者常有一句经验断言:”这个病,望诊定乾坤””那个病,一问就明白”。翻译成可检验的语言,是这样一个命题:在某个病种里,四条通道对判别力的贡献存在一个明确的排序,其中排第一的那一路就是”命门通道”——把它拿掉,判别力跌幅最大;同时存在另一路”冗余通道”,拿掉几乎无感。

这句话的价值全在于”可搬运”。如果每个病种都各有各的命门,那这条经验只能用于事后解释,不能用于任何前置决策。所以真正该问的不是”有没有命门通道”,而是”命门通道的排序能不能跨病种搬运”。

二、把命题拆成可判的两半

把”四诊消融”写成实验:四个通道取 15 个组合(四路、三路四种、两路六种、单路四种),以判别力(AUC)为读数,定义某通道的”消融跌幅”= 全模型 AUC 减去去掉该通道后的 AUC。命门通道即跌幅最大者,冗余通道即跌幅最小者。

然后分两步判。第一步,同病种内重复抽样,看 top-1 是否稳定;第二步,换病种,看两套跌幅排序的秩相关(Kendall tau)能否过线。

结论是两半的结果截然相反。同病种内极其稳定:在一个望诊占绝对优势的构造里,每类样本量仅 100、独立重抽 40 次,跌幅第一名 40 次全中”望”,样本量加到 2000 仍是 40 次全中。也就是说,”本病的命门是谁”不是噪声,是客观事实。

但换病种立刻崩塌。六个病种原型(均衡型、望诊主导、问诊主导、切诊主导、两路冗余陷阱、高相关冗余幻象)两两做秩相关,15 对的 tau 落在 -0.667 到 +0.667 之间,中位恰是 0.000,没有一对达到 0.7。四种通道在不同原型里都当过第一名。这已经不是”搬运时略有损耗”,而是”这套排序跨病种没有任何秩序可言”。

三、最要命的那一层:单指标强不等于不可替代

临床直觉里,”某一路单用就准”与”某一路不可删”几乎是同一件事。这两件事在数据上是分开的,而且分开得很难看。

造一个反例:让望诊与闻诊的单通道 AUC 都很高(0.814 与 0.807),但二者互相高度相关(0.92);问诊与切诊单通道明显弱(0.635 与 0.608),但问诊携带的是别处没有的信息。结果:单通道最强是望诊,跌幅最大却是问诊(+0.0089),望诊的跌幅只排第二(+0.0065)。把相关性推到 0.98,望的跌幅降到 +0.0029、闻降到 +0.0003,而被判为”弱通道”的问反倒拿到 +0.0176——强者几乎可以整段删掉,弱者删不得。

这条反例的意义是方法论级别的:用单通道表现来推断不可替代性,方向是错的。同样的分裂在真实文献里已经出现过——2026 年一项乳腺 DCE-MRI 四模态多中心研究里,SHAP 归因把深度学习模态排到 48.6% 的第一位,但作者自己用留一模态消融发现:真正移除后跌幅最大确是这个模态(-0.033),而移除影像组学模态后 AUC 反而上升了 0.003;作者原话是”一个模态的归因权重不能保证它不可替代”。归因与消融是两把尺子,量出来的不是同一件事。

四、弱通道区段:可比的是强与弱,不可比的是弱与弱

即便承认”跌幅排序”这一读法,它也只在两极之间可判。在望诊主导原型里,强通道跌幅约 +0.19 且跨重抽标准差仅 0.014–0.031;三条弱通道的跌幅分别落在 +0.0007 到 +0.0023,标准差与均值同量级(0.0004–0.0034)。样本量从 100 加到 3000,弱三通道内部”哪个更弱”的最一致排序方案占比始终只有 20%–28%。

所以正确读法是:这套消融能可靠判出”谁明显强、谁明显弱”,不能判出”弱的那几个谁更弱”。把三弱通道的细排序写进结论,等于把噪声当结论。

五、单表代价与尺子问题

如果硬要一张跨病种通用的命门表,代价可以算出来:固定用望诊作万能命门,六个原型平均掉 10.0 个百分点,最惨的是问诊主导型(掉 22.1 个百分点),而且它只在 3/6 个病种里是本病单通道最强的那一路。全模型平均 AUC 0.8625,最优单一固定通道平均只有 0.7630。

还有一个常被忽略的坑:排序对尺子敏感。用”绝对跌幅”、”跌幅占全模型 AUC”、”跌幅占该通道单用余量”三种归一化,望诊主导型三序完全一致,冗余幻象型则从”问闻望切”变成”闻问望切”。换尺子会换答案,说明这个排序本身就不是不变量,而是”数据 + 模型 + 尺子”三者共同产出的一个读数。

六、双重死局

把上面几层合起来,得到一个比”排序不稳”更硬的结构:

若排序真有跨病种意义,前面 15 对 tau 全不过线已经否掉了它;

若坚持存在某种跨病种同构,那么在同一构造下五次独立重抽,通道间跌幅极差只有 0.005–0.016,与重抽标准差 0.002–0.005 同量级——排序本身就在噪声里,均衡型尤其如此。同构为真时排序无意义,排序有意义时同构为假,两边同时不成立。

结论收窄成一句:命门通道是一个病种内的局部事实,不是一条可搬运的规律。凡是要按”某病望诊定乾坤”做前置设计的,必须先在该病种内实测消融跌幅,并报出强/弱区段的判读边界与重抽标准差,否则结论无法从别的病种借来。

证据与局限

已验证:同病种内消融 top-1 可重复性极高(每类 n=100 即 40/40 全中);单通道表现与消融不可替代性可以系统性分裂,此现象在已发表的多中心研究中被作者独立指出(PMID 42710818);高预测精度不等于特征重要性排序稳定,极小扰动(<0.1% 特征移除)即可让监督模型排序失稳(PMID 41749845)。

部分验证:多模态融合相对单模态确有提升(鼻咽癌 MRI+全切片 C-index 融合 0.755 对单模态 0.636/0.736,外部队列 0.735 对 0.626/0.660,PMID 40409367;PET/CT+MRI 五模态融合 0.745 对单模态 0.539–0.664,PMID 37060444);缺失模态鲁棒性可训练(PMID 39709743,MIMIC-IV 三模态,消融各组件降幅不等);中医四诊多模态融合优于决策融合、可补偿单模态信息损失(PMID 36374441,n=540,舌象+面象+脉波+健康信息)。

待验证:上述全部结论来自线性判别与高斯构造的仿真,未触及真实四诊数据;真实临床数据的通道相关结构、缺失机制、标注噪声都远复杂于本文构造;”强弱区段判读边界”的具体阈值需在真实队列上标定;跨病种”族”的划分(哪些病种可以归为一族共享命门表)尚无经验依据。

相关阅读

证据等级: C
参考文献: PMID 42710818, PMID 41749845, PMID 40409367, PMID 37060444, PMID 39709743, PMID 36374441

L1 网络状态:该主题尚未进入 L1 对应网络。EDGE_TABLE_science-yixue.md 与 CORRESPONDENCE_NETWORK.md 检索「四诊 / 消融 / 通道 / 差异 / 排序」均零命中;最近邻为 E-061 段 KB-509(冗余度/协同度 PID,PMID 33266673,属呼吸-心率两通道,且原文已明写「冗余 = 同一目标有两条通路可预测;协同 = 只有联合才有信息,两者方向可以相反」),与本篇「四诊四通道消融排序的跨病种可搬运性」不是同一条边。已登记至 08-ops/science-yixue-l1-gaps.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。

Mingmen Channel: Facts of the Disease, Hallucinations Across Diseases—Transferability Test of Four-Examination Ablation Ranking

Evidence level: C

In a specific disease category, “removing which pathway causes the most severe decline” is a stable, repeatable fact; yet transferring this ranking to another disease category renders it almost immediately invalid—and the manner of invalidity is not merely “slightly inaccurate,” but that even the identity of the top-ranked pathway may change entirely.

I. The Raw Form of Clinical Experience

Proponents of the Four Examinations often make an empirical assertion: “For this disease, the Visual Examination determines the outcome,” or “For that disease, one question makes it clear.” Translated into testable language, this becomes the proposition: within a specific disease category, the four channels of discernment possess a distinct ranking of contribution, where the top-ranked channel is the “Life Gate Channel” (命门通道)—its removal causes the largest drop in discernment power; simultaneously, another channel is the “Redundant Channel” (冗余通道), whose removal results in almost no perceptible change.

The value of this statement lies entirely in its “portability.” If every disease category has its own unique Life Gate, this experience can only be used for post-hoc explanation and cannot inform any pre-decision making. Therefore, the true question is not “Does a Life Gate Channel exist?” but rather “Can the ranking of the Life Gate Channel be transferred across disease categories?”

II. Decomposing the Proposition into Two Testable Halves

The “dissolution of four diagnostics” is formulated as an experiment: four channels yield 15 combinations (four-way, three-way, two-way, and single-way), with discrimination power (AUC) as the readout. The “dissolution drop” of a channel is defined as the full-model AUC minus the AUC after removing that channel. The Mingmen channel is identified as the one with the largest drop, while redundant channels are those with the smallest drops.

The evaluation proceeds in two steps. First, within-disease resampling assesses whether the top-1 ranking remains stable. Second, cross-disease comparison examines whether the rank correlation (Kendall tau) between the two drop-orderings exceeds the threshold.

The results of these two halves are diametrically opposed. Within-disease stability is extreme: in a configuration where inspection (望诊) dominates, each category has only 100 samples and is independently resampled 40 times; the top drop consistently lands on inspection every time. Even when sample size increases to 2,000, the top drop remains on inspection all 40 times. Thus, “who is the Mingmen of this disease” is not noise but an objective fact.

However, cross-disease comparison immediately collapses. Across six disease prototypes (balanced, inspection-dominant, inquiry-dominant, palpation-dominant, two-way redundancy trap, high-correlation redundancy illusion), pairwise rank correlations yield 15 Kendall tau values ranging from -0.667 to +0.667, with a median of exactly 0.000; no pair reaches 0.7. All four channels have each been the top-ranked in different prototypes. This is no longer “slight loss during transfer” but rather “this ranking order lacks any coherence across diseases.”

III. The Most Critical Layer: Single-Indicator Strength Does Not Equal Irreplaceability

In clinical intuition, the claim that “a single pathway is accurate when used alone” and the assertion that “a single pathway cannot be removed” are often treated as equivalent. However, these two claims are distinct in data terms, and their separation is difficult to reconcile.

Consider a counterexample: suppose the single-channel AUC values for inspection (Wangzhen) and auscultation (Wenzhen) are both high (0.814 and 0.807, respectively), yet they are highly correlated (0.92). In contrast, the single-channel values for inquiry (Wenwen) and palpation (Qiezhen) are noticeably lower (0.635 and 0.608), but inquiry carries information not present elsewhere. As a result, the strongest single channel is inspection, yet the largest drop occurs in inquiry (+0.0089); the drop for inspection ranks second (+0.0065). When correlation is pushed to 0.98, the drop for inspection falls to +0.0029, auscultation to +0.0003, and inquiry—despite being judged a “weak channel”—actually gains +0.0176. Stronger pathways can be largely removed, while weaker ones cannot.

The significance of this counterexample is methodological: inferring irreplaceability from single-channel performance is the wrong direction. Such a split has already appeared in real literature. In a 2026 multi-modal, multi-center breast DCE-MRI study, SHAP attribution ranked the deep learning modality first at 48.6%, yet the authors’ own single-modality ablation revealed that the true largest drop occurred upon removal of this modality (−0.033), while removing the radiomics modality actually increased AUC by 0.003. As the authors stated, “the attribution weight of one modality does not may support its irreplaceability.” Attribution and ablation are two different measures; they do not quantify the same thing.

IV. Weak Channel Segments: Comparable are Strong vs. Weak, Incomparable are Weak vs. Weak

Even acknowledging the “decline ranking” reading, it is discernible only between extremes. In the prognosis-dominant prototype, strong channel declines average +0.19 with cross-re-sampling standard deviations of 0.014–0.031; the three weak channels decline between +0.0007 and +0.0023, with standard deviations and means of the same order (0.0004–0.0034). As sample size increases from 100 to 3,000, the most consistent ranking of “which weak is weaker” among the three weak channels accounts for only 20%–28% of cases.

Thus, the correct reading is: this ablation reliably distinguishes “who is clearly stronger” and “who is clearly weaker,” but cannot determine “which of the weak ones is weaker.” Writing fine-grained rankings of the three weak channels into the conclusion equates noise with finding.

V. Cost of Single-Table and the Ruler Problem

If one insists on a universal “Life Gate” table applicable across all disease categories, the cost can be calculated: fixing the gazing diagnosis (a non-invasive visual assessment of the patient’s state) as the universal Life Gate results in an average drop of 10.0 percentage points across the six archetypes, with the most severe impact occurring in inquiry-dominated models (a drop of 22.1 percentage points), which only serves as the strongest single-channel indicator in 3 out of 6 disease categories. The average AUC for the full model is 0.8625, whereas the optimal single fixed channel averages only 0.7630.

There is another often-overlooked pitfall: the ranking is sensitive to the ruler. Using three normalization methods—”absolute decline,” “decline relative to the full-model AUC,” and “decline relative to the remaining capacity of that specific channel”—the inquiry-dominated model yields three completely consistent rankings, while the redundant phantom model shifts from “ask, smell, look, touch” to “smell, ask, look, touch.” Changing the ruler changes the answer, indicating that this ranking itself is not an invariant but rather a reading produced jointly by “data + model + ruler.”

VI. The Double Deadlock

Combining the layers above yields a structure more rigid than “unstable ranking”:

If ranking truly holds cross-disease significance, the fact that all 15 tau pairs failed to cross the threshold already negates it;

If one insists on some cross-disease isomorphism, then under the same construction, five independent resamples show the inter-channel drop range is only 0.005–0.016, comparable in magnitude to the resampling standard deviation of 0.002–0.005—the ranking itself lies within the noise, especially for balanced types. When isomorphism holds, ranking is meaningless; when ranking matters, isomorphism fails; both cannot be true simultaneously.

The conclusion narrows to one sentence: the Mingmen channel is a local fact within a disease category, not a portable law. Any design that uses “diagnostic observation of a specific disease to determine the outcome” as a prerequisite must first measure and report the melt-down drop range within that disease category, along with the interpretation boundaries for strong/weak segments and the resampling standard deviation; otherwise, the conclusion cannot be borrowed from other disease categories.

Evidence and Limitations

Verified: Within-disease ablation of top-1 yields extremely high reproducibility (n=100 per class achieves 40/40 full matches); single-channel performance and ablation irreducibility can systematically diverge, a phenomenon independently noted by authors in published multi-center studies (PMID 42710818); high predictive precision does not equate to stable feature importance ranking, as minute perturbations (<0.1% feature removal) can destabilize supervised model ordering (PMID 41709845).

Partially Verified: Multi-modal fusion indeed improves upon single-modal approaches (nasopharyngeal carcinoma MRI + whole-slice C-index fusion: 0.755 vs single-modal 0.636/0.736; external cohort: 0.735 vs 0.626/0.660, PMID 40409367; PET/CT + MRI five-modal fusion: 0.745 vs single-modal 0.539–0.664, PMID 37060444); missing modality robustness is trainable (PMID 39709743, MIMIC-IV three-modal, ablation of each component yields unequal drops); Traditional consensus holds that multi-modal fusion of the four examinations outperforms decision-level fusion and can compensate for single-modal information loss (PMID 36374441, n=540, tongue image + facial image + pulse wave + health information).

Pending Verification: All conclusions above derive from linear discriminant and Gaussian-constructed simulations, not yet addressing real clinical four-examination data; the channel-related structure, missing mechanisms, and labeling noise in actual clinical data are far more complex than those constructed herein; specific thresholds for “strength-segment reading boundaries” must be calibrated on real cohorts; the grouping of diseases into “families” (which disease categories can be grouped under one shared Mingmen table) lacks empirical basis.

Related Reading

Evidence level: C
References: PMID 42710818, PMID 41749845, PMID 40409367, PMID 37060444, PMID 39709743, PMID 36374441

L1 Network status: This topic has not yet entered the L1-corresponding network. Searches for “four-diagnosis / ablation / channel / difference / ranking” in EDGE_TABLE_science-yixue.md and CORRESPONDENCE_NETWORK.md yield zero hits; the nearest neighbor is segment E-061 of KB-509 (redundancy/synergy PID, PMID 33266673), which pertains to the respiratory-heart rate dual-channel and explicitly states “redundancy = two pathways can predict the same target; synergy = information exists only through joint action, and their directions can be opposite.” This is not the same edge as the cross-disease transportability of “four-diagnosis, four-channel ablation ranking.” It has been logged to 08-ops/science-yixue-l1-gaps.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 *