把尺子换成自己的:相对模式匹配能否消除三质判定的跨族裔漂移

Written by

in

Language

Evidence level: B

如果”绝对阈值”是别人给的尺子,”相对模式匹配”就是自带刻度——用反应方向加上组内 z 分数,能不能把族裔基线差异这把杂质滤掉?

正文

上一轮我们确认了一个扎心的事实:三质判定用的反应性指标(ΔLF/ΔHF/ΔRMSSD)自带族裔杂质,而且杂质大小还分三六九等。这轮问得更狠:既然绝对尺子不准,那换一把”相对尺子”行不行?具体说,就是反应方向(升/降/不变)+ 组内 z 分数这套”相对模式匹配”,能不能把跨族裔漂移整个抹掉?

先说清楚问题出在哪。族裔漂移的本质是位置参数漂移:同样一个”健康人”,非裔和欧裔的 HRV 基线就不是一个数。MESA 研究给出过最直白的证据——1175 名无心血管病人群的 10 秒 SDNN/rMSSD 参考值,按性别和种族分别给参考区间,”异常”阈值跟着族裔走(PMID 27396499)。系统综述更狠:非裔的迷走 HRV 反而比欧裔高,效应量接近一个标准差(Hedges g = 0.93)(PMID 25551201)。连反应测试都躲不开:Bogalusa 研究的 39 名青少年做全套心血管反应测试,白人的 LF/HF 反应显著高于非裔,pNN50 更低(PMID 9524048)。也就是说,用同一把绝对尺子量不同族裔,尺子本身就是歪的。

相对模式匹配的思路很朴素:把每个个体的反应先归一化到自己族裔/自己的分布里,只保留”我这族人里我算高还是算低、我是升还是降”这个相对信息,再拿去和三质标签对。这样位置参数被 z 分数消掉,剩下的是形状信息——就像风水里的”形势”看的是相对高下,不是海拔绝对高度。

这个思路在 HRV 分类里已经有直接支持。2024 年一篇 Frontiers 研究对抑郁/惊恐/健康三组人做应激-放松分类,个性化缩放(personalized scaling)后准确率全部突破 0.9,分类性能戏剧性提升,RRI 等特征的重要性还显著上升(PMID 39850069)。这说明”先用每个人的数据校准自己,再进分类器”这条路是通的。

但泼冷水的证据同样硬。1991 年一项 399 名儿童的两年随访早就发现:绝对反应值比变化值(差值)更稳定,作者明确建议除非有先验假设,否则优先用绝对值(PMID 1942018)。更麻烦的是 SWAN 睡眠研究:即便用了归一化单位(normalized units),非裔和华裔女性的 HF 仍显著高于欧裔、LF/HF 显著更低(PMID 24077772)——比例标准化消不掉族裔差异。这暴露了关键问题:族裔差异可能不只是”整体抬高或压低”(位置漂移),还有乘性/形状成分,单纯 z 分数只处理了前者。

机器学习公平性研究也提醒我们别把预处理当万能。2023 年一项 10.9 万人的 CVD 预测研究测试了三种减偏方法,结果去掉受保护属性(种族/性别)并不能显著减少模型偏倚(PMID 36706849)。还有一项跨族裔脑网络研究:白人与汉族的功能网络拓扑差异大到可以用 SVM 以 74% 准确率识别族裔(PMID 36657772)——生理信号里的族裔信息可能深到”归一化”都抹不平。

把这些拼起来,可检验的预测就清楚了:相对模式匹配能显著减少(但很可能不能完全消除)跨族裔漂移。z 分数能干掉位置漂移——这部分被 PMID 39850069 支持;但形状层面的族裔差异(PMID 24077772)和模型内部的族裔相关表征(PMID 36657772、36706849)仍会残留。工程上更聪明的做法可能是”相对判据 + 族裔分层阈值”双轨:先 z 分数去位置漂移,再在残余空间里做族裔感知校准。

还有个容易忽略的坑:z 分数要求”组”的定义本身站得住。如果”组”按族裔切,那本质上是把族裔差异直接搬进了预处理里,跨族裔泛化时新族裔的均值和方差从哪来就成了问题——要么用参考族裔的分布(又回到绝对尺子),要么现场采集(成本陡增)。这也是为什么相对模式匹配的完整形态应该包含”反应方向”这个离散特征:方向对均值和方差不敏感,跨样本迁移时比连续 z 分数更稳。所以真正值得做的检验不是”相对 vs 绝对”二选一,而是把”绝对阈值 / z 分数连续 / 方向+分档”三个模型摆在一起,看跨族裔 AUC 衰减谁最小。

对三质判定来说,这是”一把尺子走天下”还是”每人一把尺子”的路线之争。阿育吠陀经典描述体质用的本来就是相对语言(瓦塔是”相对快”的脉、卡法是”相对慢”的脉),相对模式匹配恰好更贴近传统的判读方式——只是科学上还没有人把”相对化到底能消除多少族裔漂移”这个数测出来。

证据与局限

已验证: 绝对 HRV 参考值与反应性指标均随族裔漂移(PMID 27396499、25551201、9524048);个性化缩放可显著提升 HRV 分类准确率(PMID 39850069);绝对反应值比变化值更稳定(PMID 1942018)。

部分验证: 归一化单位仍保留族裔差异,提示乘性/形状成分存在(PMID 24077772);去除受保护属性不能稳定减偏,预处理标准化非万能(PMID 36706849)。

待验证: 相对模式匹配(反应方向 + 组内 z 分数)对三质反应性指标的族裔漂移消除率到底是多少——需要多族裔同三质分型队列,头对头比较绝对阈值模型 vs 组内标准化模型的跨族裔 AUC 与重分类指数。

相关阅读

证据等级: B
参考文献: PMID 27396499, PMID 25551201, PMID 9524048, PMID 1942018, PMID 39850069, PMID 24077772, PMID 36657772, PMID 36706849


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。

Replacing the Ruler with Oneself: Can Relative Pattern Matching Eliminate Cross-Racial Drift in Three-Point Discrimination

Evidence level: B

If the “absolute threshold” is a ruler provided by others, “relative pattern matching” is a self-calibrated scale—can adding reaction direction plus within-group z-scores filter out the ethnic baseline differences acting as impurities?

Main Text

In the previous round, we confirmed a painful fact: the reactivity indices used in tridosha determination (ΔLF/ΔHF/ΔRMSSD) carry inherent ethnic impurities, with the magnitude of these impurities varying significantly. This round asks more directly: since the absolute ruler is inaccurate, can we switch to a “relative ruler”? Specifically, can the “reaction direction (increase/decrease/no change) + within-group z-score” framework of “relative pattern matching” eliminate the entire cross-ethnic drift?

First, let’s clarify where the problem lies. The essence of ethnic drift is location parameter drift: a “healthy person” has different HRV baselines for non-Hispanic Black versus non-Hispanic White populations. The MESA study provides the most direct evidence—reference values for 10-second SDNN/rMSSD in 1,175 cardiovascular disease-free individuals are stratified by sex and race, meaning the “abnormal” threshold follows the ethnicity (PMID 27396499). Systematic reviews are even more stark: vagal HRV is actually higher in non-Hispanic Black individuals than in non-Hispanic Whites, with an effect size approaching one standard deviation (Hedges g = 0.93) (PMID 25551201). Even reaction tests cannot escape this: in Bogalusa’s study of 39 adolescents undergoing full cardiovascular reaction testing, Whites showed significantly higher LF/HF responses and lower pNN50 compared to non-Hispanic Blacks (PMID 9524048). In other words, using the same absolute ruler to measure different ethnicities means the ruler itself is skewed.

The logic of relative pattern matching is straightforward: normalize each individual’s response within their own ethnicity/distribution, retaining only the relative information of “whether I am high or low among my group” and “whether I am rising or falling,” then match this against tridosha labels. This eliminates the location parameter via z-scores, leaving only shape information—much like Feng Shui‘s “situation” (Xing Shi), which observes relative height rather than absolute altitude.

This approach already has direct support in HRV classification. A 2024 Frontiers study classifying stress-relaxation states across depression, panic, and healthy groups found that personalized scaling pushed accuracy above 0.9 for all three groups, dramatically improving classification performance and significantly increasing the importance of features like RRI (PMID 39850069). This confirms that the path of “calibrating to one’s own data first, then entering the classifier” is viable.

However, evidence cooling the waters is equally robust. A 1991 two-year follow-up of 399 children already found that absolute reaction values are more stable than change values (differences), with authors explicitly recommending the use of absolute values unless prior assumptions exist (PMID 1942018). More complicating is the SWAN sleep study: even with normalized units, non-Hispanic Black and Chinese women still showed significantly higher HF and significantly lower LF/HF compared to non-Hispanic Whites (PMID 24077772)—proportional normalization does not eliminate ethnic differences. This exposes a critical issue: ethnic differences may not just involve “overall elevation or depression” (location drift) but also multiplicative/shape components, which simple z-scores only address the former.

Machine learning fairness research also warns against treating preprocessing as a panacea. A 2023 study of 109,000 people testing three debiasing methods for CVD prediction found that removing protected attributes (race/sex) did not significantly reduce model bias (PMID 36706849). Furthermore, a cross-ethnic brain network study showed that functional network topology differences between Whites and Han Chinese were so large that an SVM could identify ethnicity with 74% accuracy (PMID 36657772)—indicating that ethnic information in physiological signals may be deep enough that “normalization” cannot flatten it.

Putting these pieces together clarifies the testable prediction: relative pattern matching can significantly reduce (but likely cannot completely eliminate) cross-ethnic drift. Z-scores can eliminate location drift—supported by PMID 39850069—but ethnic differences at the shape level (PMID 24077772) and ethnic-related representations within the model (PMID 36657772, 36706849) will persist. A more engineering-wise approach might be a dual-track of “relative criteria + ethnic-stratified thresholds”: first remove location drift with z-scores, then perform ethnicity-aware calibration in the residual space.

Another often-overlooked pitfall: z-scores require the definition of the “group” itself to be sound. If the “group” is defined by ethnicity, it essentially embeds ethnic differences directly into preprocessing, raising questions about where to derive mean and variance for new ethnicities during cross-ethnic generalization—either by using a reference ethnicity’s distribution (returning to the absolute ruler) or collecting data on-site (drastically increasing costs). This is why the complete form of relative pattern matching should include the discrete feature of “reaction direction”: direction is insensitive to mean and variance, making it more stable during cross-sample transfer than continuous z-scores. Therefore, the worthier test is not a binary choice between “relative vs. absolute,” but comparing three models—”absolute thresholds / continuous z-scores / direction + binning”—to see which shows the smallest cross-ethnic AUC decay.

For tridosha determination, this is a debate between “one ruler for all” versus “a ruler for each person.” Ayurvedic classics describe constitutions using inherently relative language (Vata is a “relatively fast” pulse, Kapha is a “relatively slow” pulse), making relative pattern matching closer to traditional interpretive methods—though science has yet to measure exactly how much ethnic drift can be eliminated by “relativizing to the extreme.”

Evidence and Limitations

Verified: Absolute HRV reference values and reactivity indices all drift by ethnicity (PMID 27396499, 25551201, 9524048); personalized scaling significantly improves HRV classification accuracy (PMID 39850069); absolute reaction values are more stable than change values (PMID 1942018).

Partially Verified: Normalized units still retain ethnic differences, suggesting multiplicative/shape components exist (PMID 24077772); removing protected attributes does not consistently reduce bias, indicating preprocessing standardization is not a panacea (PMID 36706849).

To Be Verified: The exact elimination rate of relative pattern matching (reaction direction + within-group z-score) for ethnic drift in tridosha reactivity indices—requires multi-ethnic cohorts with identical tridosha phenotyping, head-to-head comparison of absolute threshold models vs. within-group standardized models regarding cross-ethnic AUC and reclassification indices.

Related Reading

Evidence Level: B
References: PMID 27396499, PMID 25551201, PMID 9524048, PMID 1942018, PMID 39850069, PMID 24077772, PMID 36657772, PMID 36706849


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 *