指纹是认人的,不是认穴的:模型换个人还认不认得出

Written by

in

Language

Evidence level: B

一个在同一个人身上能认出”这是哪个穴”的分类器,换一个人、换一台机器之后还剩多少——这个”剩多少”,才是穴位特异性的真实强度。

正文

上一篇给出了判别力检验方案:多穴位、参数化刺激、交叉验证、看 AUC 是否显著高于 0.5。方向对,但判据里藏着一个陷阱——AUC 不是一个数字,是三个数字的合成:同一个人隔周再来、换一个没进过训练的受试者、换一台扫描仪。三种衰减机制不同,压成一个”是否显著”,等于用平均值盖住三种失败。

一、先把上限算出来

90 个实验、1008 名被试的元分析显示,任务态 fMRI 平均组内相关系数(ICC)只有 0.397;在两个独立队列重测 11 种常用任务,ICC 落在 0.067–0.485。作者的结论很硬:这些指标”目前不适合用于脑生物标志物发现”(PMID 32489141)。

这不是背景,是天花板。误差对相关的衰减是:观测相关 ≈ 真相关 × √(信度₁×信度₂)。信度取 0.4,系数就是 0.4——真相关 0.5 的信号观测到只剩 0.2,80% 功效下要近两百人。

针刺这边有一份直接的重测数据:6 人、每人 6 次相同扫描,条件含胆经光明(GB37)、膀胱经昆仑(UB60)、非穴点与手指敲击对照。结果分两层:组平均层面,三个条件的得气评分与信号变化彼此相似;个体层面,针刺诱发的信号变化显著比手指敲击更多变(PMID 17157035)。

两层各打一枪。第一枪打特征:单次扫描的脑图不够格当指纹。第二枪更麻烦——组平均层面三个条件都相似,那”跨穴判别”要找的组间差异,得先证明它存在。

二、跨被试不是打折,是换编码

直觉是跨被试差因为噪声更大。但有研究给出更难处理的答案:同一批数据上分别训练被试内与跨被试分类器,两者依赖的体素权重模式并不一致,最优脑区也不同(PMID 20347995)。衰减不是同一个信号变弱,是模型被迫去学另一套东西。

另一项工作直接拆了”多体素模式”四个字:解码器空间上并不精确(做空间平滑,性能几乎不受影响);90% 冗余(随机抽 10% 的体素就足以恢复全部性能);性能还依赖拿来对比的条件(PMID 35219121)。跑出高 AUC 的解码器抓住的往往是大尺度网络状态,不是穴位细节——”指纹”比数据支持的要细得多。

三、”指纹”这个方向一开始就反了

功能连接组确实能做指纹:在几百人里认出同一个人,跨场次成功,跨任务态与静息态也能认(PMID 26457551)。结论扎实。

但注意方向。那里被分类的是人——个体差异是信号,越大越好认;这里被分类的是穴——个体差异是噪声,必须先压掉才看得见穴位。同一个数据集上,识别人可以接近饱和,识别刺激却可能刚过随机,这不矛盾,因为信噪比结构是反的。

所以”网络指纹”会让人预支两个还没验的结论:独一无二、终身稳定。

四、跨中心那一层最容易被误读

5 名受试者在 10 台扫描仪上各扫两次:重测信度不错,但跨中心信度一开始很低,站点与”站点×被试”方差贡献显著(PMID 17636563)。验证集若与发现集同中心、同机器、同术者,泛化成功只证明同中心可重复。反过来更危险:站点间的系统偏移会被读成”这个穴在不同人群里效应不同”,仪器效应就此翻译成生理学发现。

五、HRV 那根柱子,长度不够

63 人重复测量:6 分钟与 2 分钟时域指标 ICC 大于 0.7,2 分钟频域大于 0.5,10 秒更低(PMID 15286929)。更细的标定是:心率 60 秒、时域 240 秒、相对频域 300 秒(PMID 33856258)。针刺刺激段常只有几十秒到两分钟,而低频成分需至少两分钟——窗口不够,算出来的已不是定义里的那个量。

六、五条可执行判据

一,别只报 AUC,报穴-穴距离矩阵的重测相关:同一人两次扫描各算 K 个穴位两两距离矩阵,再求两矩阵相关。AUC 会被类别数与不平衡掩盖,距离矩阵不会。

二,加阳性对照:同批数据跑”识别个体”。若识别人接近饱和而识穴接近随机,那点信号属于人,不属于穴。

三,验证集必须换中心,退一步至少换扫描仪或换术者,否则”跨被试泛化”实际在测重复性。

四,样本量按信度算:给定 ICC 与可接受的 AUC 区间宽度反推 N。小样本功效问题在神经科学里被反复证明(PMID 23571845),交叉验证误差条也会大到无法支持结论(PMID 28655633)。

五,预注册等效性判据:若观测 AUC 的置信区间同时覆盖 0.5 与信度允许的上限,结论只能写”无法区分有无特异性”。

七、粗守形,上守神

《灵枢·九针十二原》:“粗守形,上守神。” 判别模型守的是”形”——体素模式的空间形状;针刺真正的差异可能在”神”——得气出现与消退的时序。形没抓住,不等于神不存在。

《素问·异法方宜论》早把个体差异放进治法:同一类病,东方砭石、西方毒药、北方灸焫、南方九针、中央导引按跷,“一病而治各不同,皆愈”。古人从不假设一套参数通用。

《灵枢·官能》更直说:“语徐而安静,手巧而心审谛者,可使行针艾。” 行针交给特定的一类人——这是公开承认术者方差。现代把它塞进”随机效应”,古人写成择人标准。差别在于:写出来,就有人去管它。

证据与局限

已验证: ①任务态 fMRI 的重测信度整体偏低(90 个实验元分析,平均 ICC 0.397;两个独立队列 0.067–0.485,PMID 32489141);②针刺诱发的 fMRI 信号变化在同一受试者不同场次间的变异显著大于主动对照任务,且组平均层面穴位与非穴可相近(PMID 17157035);③被试内与被试间分类器依赖不同的信息与脑区(PMID 20347995);④多体素解码器高度冗余、空间不精确、依赖对比条件(PMID 35219121);⑤功能连接可稳定识别个体,跨场次、跨状态成立(PMID 26457551);⑥跨中心 fMRI 信度显著受站点与站点×被试方差影响(PMID 17636563);⑦短时 HRV 的重测信度随记录时长下降,频域指标门槛最高(PMID 15286929, PMID 33856258)。

部分验证: “距离矩阵重测相关优于 AUC”目前是方法学推论,尚未在穴位数据上做过正面对照;跨被试衰减中”噪声”与”编码差异”各占多少,缺少可分解的实验。

待验证: ①同一批多穴位数据上,”识别人”与”识别穴”的准确率落差有多大;②在 ICC 已知的条件下,多穴位判别 AUC 的实际可检出下限;③换中心后穴位判别的 AUC 衰减幅度是否超过仪器效应本身;④距离矩阵重测相关与临床疗效是否存在关联。

相关阅读

证据等级: B
参考文献: PMID 32489141, PMID 17157035, PMID 20347995, PMID 35219121, PMID 26457551, PMID 17636563, PMID 15286929, PMID 33856258, PMID 23571845, PMID 28655633


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。

Fingerprints Identify People, Not Body Holes: Will the Model Still Recognize Them After Changing Individuals?

Evidence level: B

A classifier that can identify “which acupoint” on the same individual—how much of this capability remains when switching to a different person or machine? This “remainder” is the true strength of acupoint specificity.

Main Text

The previous article provided a discrimination validation scheme: multi-acupoint, parametric stimulation, cross-validation, and checking if the AUC is significantly above 0.5. The direction is correct, but the criterion hides a trap—AUC is not a single number; it is a synthesis of three numbers: re-testing the same person after a week, using a new untrained subject, or switching scanners. These three attenuation mechanisms differ; compressing them into a single “significance” verdict is like covering three types of failure with an average.

I. First, Calculate the Ceiling

A meta-analysis of 90 experiments with 1,008 participants showed that the average intraclass correlation coefficient (ICC) for task-state fMRI is only 0.397; in two independent cohorts re-testing 11 common tasks, ICC ranged from 0.067–0.485. The authors’ conclusion is firm: these metrics “are currently not suitable for discovering brain biomarkers” (PMID 32489141).

This is not background noise; it is a ceiling. The attenuation of correlation due to error is: Observed Correlation ≈ True Correlation × √(Reliability₁ × Reliability₂). With reliability set at 0.4, the coefficient is 0.4—a true correlation of 0.5 yields an observed value of only 0.2, requiring nearly two hundred participants for 80% power.

Regarding acupuncture, there is direct re-test data: 6 participants, each scanned 6 times under identical conditions, including GB37 (Lianggu), UB60 (Kunlun), non-acupoint points, and finger tapping controls. Results showed two layers: At the group average level, the qi sensation scores and signal changes across the three conditions were similar; At the individual level, signal changes induced by acupuncture varied significantly more than those from finger tapping (PMID 17157035).

Two shots fired at different targets. The first shot targets features: a single scan’s brain map is insufficient to serve as a fingerprint. The second shot is more troublesome—at the group average level, all three conditions are similar; therefore, the inter-group difference sought for “cross-acupoint discrimination” must first be proven to exist.

II. Cross-Subject Is Not Discounting, It Is Changing Encoding

Intuition suggests cross-subject performance drops due to higher noise. However, research offers a harder-to-process answer: when trained on the same dataset, both within-subject and between-subject classifiers rely on inconsistent voxel weight patterns, and the optimal brain regions differ (PMID 20347995). The attenuation is not the same signal weakening; it is the model forced to learn a different set of rules.

Another study directly deconstructed the phrase “multi-voxel pattern”: the decoder space is not precise (spatial smoothing has almost no impact on performance); 90% redundancy (randomly sampling 10% of voxels is sufficient to restore full performance); and performance depends on the conditions used for comparison (PMID 35219121). High-AUC decoders often capture large-scale network states rather than acupoint details—”fingerprints” are far finer than the data supports.

III. The “Fingerprint” Direction Was Wrong from the Start

Functional connectomes can indeed serve as fingerprints: identifying the same individual across hundreds of people, across sessions, and across task-state versus rest states is successful (PMID 26457551). The conclusion is solid.

But note the direction. There, the classifier targets people—individual differences are the signal; larger is better recognized. Here, the classifier targets acupoints—individual differences are noise that must be suppressed before acupoint features become visible. On the same dataset, recognizing people can approach saturation, while recognizing stimuli may barely exceed random chance; this is not contradictory because the signal-to-noise ratio structure is inverted.

Therefore, “network fingerprints” lead to two premature conclusions: uniqueness and lifelong stability.

IV. The Cross-Center Layer Is Most Easily Misread

Five participants scanned twice on ten different scanners: re-test reliability was good, but cross-center reliability was initially low, with significant variance contributions from site and site × subject (PMID 17636563). If the validation set matches the discovery set in center, machine, and operator, successful generalization only proves repeatability within the same center. The reverse is more dangerous: systematic shifts between sites may be misread as “this acupoint has different effects across different populations,” translating instrument effects into physiological findings.

V. The HRV Pillar Is Too Short

In a 63-person repeated-measurement study: ICC for time-domain metrics exceeded 0.7 at 6 minutes and 2 minutes, while frequency-domain metrics exceeded 0.5 only at 2 minutes (lower at 10 seconds) (PMID 15286929). Finer calibration shows: heart rate requires 60 seconds, time-domain metrics require 240 seconds, and relative frequency domains require 300 seconds (PMID 33856258). Acupuncture stimulation segments often last only tens of seconds to two minutes, while low-frequency components require at least two minutes—the window is insufficient; what is calculated is not the quantity defined in the theory.

VI. Five Executable Criteria

  1. Do not report only AUC; report the re-test correlation of the inter-acupoint distance matrix: calculate the pairwise distance matrix for K acupoints for two scans of the same person twice, then compute the correlation between the two matrices. AUC is masked by category count and imbalance; distance matrices are not.
  2. Add a positive control: run “individual recognition” on the same batch of data. If recognizing people approaches saturation but recognizing acupoints approaches random chance, that signal belongs to the person, not the acupoint.
  3. The validation set must change centers; as a minimum step, switch scanners or operators, otherwise “cross-subject generalization” actually measures repeatability.
  4. Calculate sample size based on reliability: reverse-calculate N given the acceptable AUC interval width and known ICC. Small sample power issues are repeatedly proven in neuroscience (PMID 23571845), and cross-validation error bars can be too large to support conclusions (PMID 28655633).
  5. Pre-register equivalence criteria: if the confidence interval of the observed AUC simultaneously covers 0.5 and the reliability-allowed upper limit, the conclusion can only be written as “unable to distinguish presence or absence of specificity.”

VII. Roughly Guard the Form, Above Guard the Spirit

The Classic of Internal Medicine · Nine Needles and Twelve Origins: “Roughly guard the form, above guard the spirit.” Discrimination models guard the “form”—the spatial shape of voxel patterns; the true difference in acupuncture may lie in the “spirit”—the temporal sequence of qi sensation appearing and subsiding. If the form is not grasped, it does not mean the spirit does not exist.

The Classic of Yellow Emperor · Treatments According to Different Regions placed individual differences into treatment methods early on: for the same type of disease, moxa stones in the East, toxic drugs in the West, moxibustion in the North, nine needles in the South, and guiding pressing in the Center, “one disease treated differently yet all recover.” The ancients never assumed a single set of parameters is universal.

The Classic of Internal Medicine · Functions and Abilities states even more directly: “Those who speak slowly and calmly, with skilled hands and careful hearts, may be entrusted with needling and moxa.” Needling is entrusted to a specific class of people—this is an openly admitted operator variance. Modern practice hides this in “random effects”; the ancients wrote it as a selection standard. The difference lies in: once written down, someone manages it.

Evidence and Limitations

Verified: ① Re-test reliability in task-state fMRI is generally low (meta-analysis of 90 experiments, average ICC 0.397; two independent cohorts 0.067–0.485, PMID 32489141); ② fMRI signal changes induced by acupuncture vary significantly more between sessions in the same subject than in active control tasks, and at the group average level, acupoints and non-acupoints can be similar (PMID 17157035); ③ Within-subject and between-subject classifiers rely on different information and brain regions (PMID 20347995); ④ Multi-voxel decoders are highly redundant, spatially imprecise, and dependent on comparison conditions (PMID 35219121); ⑤ Functional connectivity can stably identify individuals across sessions and states (PMID 26457551); ⑥ Cross-center fMRI reliability is significantly affected by site and site × subject variance (PMID 17636563); ⑦ Re-test reliability of short-term HRV decreases with recording duration, with frequency-domain metrics having the highest threshold (PMID 15286929, PMID 33856258).

Partially Verified: “Distance matrix re-test correlation is superior to AUC” is currently a methodological inference; no direct comparison has been made on acupoint data yet. The proportion of “noise” versus “encoding differences” in cross-subject attenuation lacks a decomposable experiment.

To Be Verified: ① On the same batch of multi-acupoint data, what is the accuracy gap between “recognizing people” and “recognizing acupoints”; ② Under known ICC conditions, what is the actual detectable lower limit for multi-acupoint discrimination AUC; ③ Whether the AUC attenuation after changing centers exceeds the instrument effect itself; ④ Whether distance matrix re-test correlation correlates with clinical efficacy.

Related Reading

Evidence Level: B
References: PMID 32489141, PMID 17157035, PMID 20347995, PMID 35219121, PMID 26457551, PMID 17636563, PMID 15286929, PMID 33856258, PMID 23571845, PMID 28655633


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 *