尺子都一样了,数据却不一样:族裔数据质量差异如何拖垮三质判别的公平性

Written by

in

Language

Evidence level: B

绝对阈值和组内标准化争的是”用哪把尺子”,但真正决定某个族裔判别 AUC 的,可能是那把尺子量到的数据本身有多脏。

正文

上一篇把”标准化之后还剩什么”拆到了形状成分。这一篇问一个更基础的问题:如果两把尺子量的根本不是同样干净的数据,尺子之争还有意义吗?

母文(#341)在讨论”绝对阈值 vs 组内标准化”时留下了一个口子,来自一项 7.7 万人的结直肠癌风险预测研究(PMID 39136986)。加入种族变量后,黑人群体上的 AUC 显著提升(P=0.006),但原因不是黑人有什么”生理上的特殊风险”,而是自报家族史这个关键预测因子,在黑人群体里的数据质量系统性更差。换句话说,模型本身未必有偏见,是数据有偏见;种族变量之所以有用,是因为它当了”数据质量”的代理。这条证据把问题从”选哪把尺子”重新定位到了”数据本身有多干净”。

缺失数据不是随机掉下来的。2025 年一项系统实验(PMID 40585432)用 150 个真实场景变体证明:当缺失与受保护属性、与结局变量相关时,即便缺失比例很小,也会显著放大分类器的组间差异。2023 年一项急诊分诊机器学习研究(PMID 38638298)直接做了操控实验——人为改变电子病历问题清单的缺失量,两族裔的 c 统计量都随之变化,但假阳性率的变化幅度在组间出现统计显著差异(黑人 +2.79 个百分点 vs 白人 +5.18 个百分点)。更根本的理论来自 Zhang 与 Long(PMID 35495871):”只保留完整病例”训练的算法在完整病例域里可能公平,但映射回真实目标人群时会对边缘群体表现出不成比例的偏倚——两个域的分布本来就不同。一句话,公平性不是模型的属性,而是模型加数据质量的联合属性。

HRV 恰好是数据质量差异最”物理”的领域。2026 年一项对照研究(PMID 40968160)用客观测量的皮肤色素(ITA°)做自变量,发现三款腕戴/胸戴设备的心率数据质量都随肤色加深而下降——最要命的一条不是误差(SlateSafety 只多了约 1 bpm),而是缺失数据的分布:深肤色受试者(ITA°<10°)贡献了 Apple 设备 50%、SlateSafety 设备 85% 的缺失数据,离群点也有 56%–62% 落在深肤色人群。2024 年一项荟萃分析(PMID 39388258)确认 PPG 派生指标的精度按色素分组分层;NEJM 上的经典证据(PMID 33326721)更直白——脉搏血氧在黑人患者中隐匿性低氧血症的发生率接近白人的三倍。2023 年《Annual Review of Medicine》的综述(PMID 35901314)把偏倚来源归成三类:设备物理设计、不具代表性的数据、以及把生物学因素与社会决定因素混为一谈。对 HRV 而言,这层物理偏倚会直接传导——PPG 是 PRV(脉搏率变异性)的源头,而三质判定的很多特征恰恰来自 PPG 通道。

即便信号拿到了,HRV 对伪迹也极其敏感。2019 年提出的稳健校正算法对房性/室性早搏的检测敏感度达 96.96%(PMID 31314618);而 2012 年的经典综述指出,编辑方法、参数、研究人群、序列长度不同,编辑对 HRV 结果的影响也不同(PMID 22654764)——校正本身就会引入人群依赖的偏差。2025 年一项时间序列插补基准测试(PMID 40896298)给出最后一刀:MCAR(完全随机缺失)下的插补精度显著优于 MAR(随机缺失)与 NMAR(非随机缺失),可现实数据很少是 MCAR——用随机删除评估出来的”干净”性能,会系统性高估真实世界的表现。

把这些拼起来,本篇的可检验预测就清楚了。预测一:在特征质量齐平的理想数据上,组内标准化的跨族裔 AUC 衰减确实小于绝对阈值模型(母文的预测一应当复现)。预测二:一旦按族裔注入异质噪声或缺失(MCAR 与 MAR 两种机制),组内标准化的优势会单调衰减,并在某个噪声水平上被绝对阈值反超——存在一个”交叉点”。预测三:MAR 机制造成的衰减显著大于 MCAR,因为”缺失概率取决于肤色相关的信号质量”正是数据质量差异的真实形态。预测四:组内标准化在高缺失率族裔上均值/方差估计更不稳(呼应上一篇的”伪形状”),该族裔 AUC 的置信区间更宽、校准曲线更抖。

这里有个反直觉的推论值得强调:组内标准化在高噪声族裔上可能比绝对阈值更糟。z 分数用该族裔自己被污染的均值和方差当基准——噪声抬高了方差,归一化等于把噪声”合法化”地压回标准尺度;缺失率高又让估计更不稳。于是它很可能把”数据质量问题”误当成”群体差异”校准掉,制造虚假的公平外观。若再用插补补缺失,插补本身又会引入新的组间性能差(PMID 40585432、35495871)。

裁决实验可这样设计:取多族裔、同一三质金标准队列,先建立”特征质量齐平”基线,再按族裔人工注入异质噪声与缺失(MCAR/MAR 两种机制,5%–30% 梯度),两模型分别训练,报告分族裔 AUC、校准指标与净重分类指数,画出”噪声注入量 → AUC 衰减”梯度曲线并定位交叉点,并用置换检验确认差异非随机。

工程含义相当直接:三质判定产品的优先级应该翻转。先把数据质量拉平——统一的信号质量门控、族裔平衡的采集、统一且经验证的伪迹校正协议——再谈尺子。数据质量不齐时比较两把尺子,本质是在脏数据上做公平性化妆:无论选哪把,数据质量最差的族裔都会同步受损。

证据与局限

已验证: 种族调整可因修正”关键特征数据质量差异”而提升特定族裔 AUC(PMID 39136986);缺失与受保护属性/结局相关时会放大组间差异,即使缺失量很小(PMID 40585432);仅用完整病例训练的算法在目标人群上可能对边缘群体系统性偏倚(PMID 35495871);操控缺失量会改变分族裔预测性能及其组间差(PMID 38638298);皮肤色素显著影响 PPG 数据质量,缺失数据与离群点不成比例地集中在深肤色人群(PMID 40968160、39388258);脉搏血氧在黑人患者中隐匿性低氧血症率近三倍(PMID 33326721);设备物理设计、不代表性数据与社会决定因素混淆是偏倚三大来源(PMID 35901314);HRV 对早搏/伪迹高度敏感且校正方法影响人群依赖(PMID 31314618、22654764);MCAR 下插补精度显著优于 MAR/NMAR(PMID 40896298)。

部分验证: 组内标准化可消位置漂移(由定义与既有分解框架支持),但”在高噪声族裔上是否被绝对阈值反超”尚无直接研究,属方法学外推;数据质量差异对模型公平性的影响已有实证(PMID 39136986、40585432、38638298),但尚未在三质 HRV 场景直接检验。

待验证: 三质分型下,按族裔注入异质噪声/缺失(MCAR 与 MAR)时,绝对阈值与组内标准化两模型的该族裔 AUC 是否同步受损、衰减梯度形状如何、是否存在交叉点;组内标准化在高缺失率族裔的 AUC 方差是否更大;插补策略选择是否独立贡献组间性能差。

相关阅读

证据等级: B
参考文献: PMID 39136986, PMID 40585432, PMID 35495871, PMID 38638298, PMID 40968160, PMID 39388258, PMID 33326721, PMID 35901314, PMID 31314618, PMID 22654764, PMID 40896298


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。

When Rulers Are the Same but Data Differ: How Ethnic Data Quality Disparities Undermine Fairness in Three-Stage Judgment

Evidence level: B

Absolute thresholds and within-group standardization debate “which ruler to use,” but what truly determines an ethnic group’s discrimination AUC may be how dirty the data measured by that ruler actually is.

Main Text

The previous article decomposed “what remains after standardization” into shape components. This piece asks a more fundamental question: If two rulers measure fundamentally different levels of data cleanliness, does the debate over which ruler to use matter?

The original paper (#341) left an opening regarding “absolute thresholds vs. within-group standardization,” stemming from a colorectal cancer risk prediction study with 77,000 participants (PMID 39136986). After adding a race variable, the AUC for the Black population increased significantly (P=0.006), but not because of any “physiological special risk” inherent to Black people, but rather because the data quality for the key predictor ‘self-reported family history’ was systematically worse in the Black population. In other words, the model itself may not be biased; the data is. The utility of the race variable arises because it acts as a proxy for “data quality.” This evidence repositions the problem from “which ruler to choose” to “how clean the data itself is.”

Missing data does not drop randomly. A 2025 systematic experiment (PMID 40585432) using 150 real-world scenario variants proved that when missingness correlates with protected attributes or outcome variables, even small missing rates significantly amplify inter-group differences in classifiers. A 2023 emergency triage machine learning study (PMID 38638298) directly conducted a manipulation experiment—artificially altering the missingness in electronic health record questionnaires—and found that while c-statistics changed for both ethnic groups, the magnitude of change in false positive rates showed a statistically significant inter-group difference (Black +2.79 percentage points vs. White +5.18 percentage points). More fundamental theory comes from Zhang and Long (PMID 35495871): algorithms trained on “complete cases only” may be fair within the complete case domain but exhibit disproportionate bias against marginalized populations when mapped back to the true target population—the distributions of the two domains are inherently different. In short, fairness is not a property of the model alone, but a joint property of the model and data quality.

HRV is precisely the domain where data quality differences are most “physical.” A 2026 comparative study (PMID 40968160) used objectively measured skin pigmentation (ITA°) as an independent variable and found that the data quality of heart rate signals from three wrist-worn/chest-worn devices all declined as skin color deepened. The most critical factor was not error (SlateSafety only added ~1 bpm), but the distribution of missing data: subjects with darker skin (ITA°<10°) contributed 50% of missing data for Apple devices and 85% for SlateSafety devices, with outliers falling in the 56%–62% range for darker-skinned populations. A 2024 meta-analysis (PMID 39388258) confirmed that the precision of PPG-derived metrics stratifies by pigment groups; classic evidence in NEJM (PMID 33326721) is even more direct—esoteric-tradition hypoxemia rates in Black patients are nearly three times those in White patients. A 2023 review in Annual Review of Medicine (PMID 35901314) categorizes bias sources into three types: physical device design, non-representative data, and conflating biological factors with social determinants. For HRV, this layer of physical bias transmits directly—PPG is the source of PRV (Pulse Rate Variability), and many features for tridosha determination originate from the PPG channel.

Even with signal acquisition, HRV is extremely sensitive to artifacts. A robust correction algorithm proposed in 2019 achieved a detection sensitivity of 96.96% for atrial/ventricular premature beats (PMID 31314618); however, a classic 2012 review pointed out that the impact of editing on HRV results varies depending on the editing method, parameters, study population, and sequence length (PMID 22654764)—the correction itself introduces population-dependent bias. A 2025 time series interpolation benchmark test (PMID 40896298) provides the final cut: imputation accuracy under MCAR (completely random missingness) is significantly superior to MAR (random missingness) and NMAR (non-random missingness), yet real-world data is rarely MCAR—evaluating “clean” performance via random deletion systematically overestimates real-world performance.

Putting these pieces together clarifies the testable predictions for this article. Prediction 1: On ideal data with flat feature quality, the inter-ethnic AUC decay for within-group standardization is indeed smaller than for absolute threshold models (the original paper’s Prediction 1 should be reproduced). Prediction 2: Once heterogeneous noise or missingness (MCAR and MAR mechanisms) is injected by ethnicity, the advantage of within-group standardization will monotonically decay and be surpassed by the absolute threshold at a certain noise level—a “crossover point” exists. Prediction 3: The decay caused by MAR mechanisms is significantly greater than MCAR, because “missing probability depending on signal quality related to skin color” is the true morphological form of data quality disparities. Prediction 4: Within-group standardization yields less stable mean/variance estimates in high-missing-rate ethnic groups (echoing the “pseudo-shape” from the previous article), resulting in wider confidence intervals for that group’s AUC and a more jittered calibration curve.

Here is an counter-intuitive inference worth emphasizing: Within-group standardization may be worse than absolute thresholds in high-noise ethnic groups. The z-score uses the mean and variance of that specific ethnic group, which are contaminated by noise—noise inflates variance, and normalization effectively “legitimizes” the compression of noise back to a standard scale; high missingness further destabilizes estimates. Thus, it likely misattributes “data quality issues” as “group differences” for calibration, creating a false appearance of fairness. If imputation is then used to fill missing data, the imputation itself introduces new inter-group performance gaps (PMID 40585432, 35495871).

The adjudication experiment can be designed as follows: take a multi-ethnic, same-tridosha gold-standard cohort; first establish a baseline with “flat feature quality,” then artificially inject heterogeneous noise and missingness (MCAR/MAR mechanisms, 5%–30% gradient) by ethnicity; train both models separately; report per-ethnic AUC, calibration metrics, and NRI; plot the “noise injection amount → AUC decay” gradient curve to locate the crossover point, and use permutation tests to confirm differences are non-random.

The engineering implications are direct: the priority for tridosha determination products should flip. First, flatten data quality—uniform signal quality gating, ethnically balanced acquisition, and unified, validated artifact correction protocols—then discuss the ruler. Comparing two rulers with uneven data quality is essentially doing “fairness makeup” on dirty data: regardless of which ruler is chosen, the ethnic group with the poorest data quality will suffer simultaneously.

Evidence and Limitations

Verified: Race adjustment can improve specific ethnic AUC by correcting “key feature data quality disparities” (PMID 39136986); missingness correlates with protected attributes/outcomes and amplifies inter-group differences even at low rates (PMID 40585432); algorithms trained solely on complete cases may exhibit systematic bias against marginalized populations in target populations (PMID 35495871); manipulating missingness alters per-ethnic prediction performance and inter-group differences (PMID 38638298); skin pigmentation significantly affects PPG data quality, with missing data and outliers disproportionately concentrated in darker-skinned populations (PMID 40968160, 39388258); esoteric-tradition hypoxemia rates in Black patients are nearly three times those in White patients (PMID 33326721); physical device design, non-representative data, and conflation of social determinants are the three major bias sources (PMID 35901314); HRV is highly sensitive to premature beats/artifacts, and correction methods introduce population-dependent bias (PMID 31314618, 22654764); imputation accuracy under MCAR is significantly superior to MAR/NMAR (PMID 40896298).

Partially Verified: Within-group standardization can eliminate location drift (supported by definition and existing decomposition frameworks), but whether it is surpassed by absolute thresholds in “high-noise ethnic groups” lacks direct research, belonging to methodological extrapolation; empirical evidence exists regarding data quality disparities’ impact on model fairness (PMID 39136986, 40585432, 38638298), but this has not been directly tested in the tridosha HRV scenario.

To be Verified: In tridosha stratification, when heterogeneous noise/missingness (MCAR and MAR) is injected by ethnicity, whether both models suffer synchronous damage in that group’s AUC, what the decay gradient shape is, and if a crossover point exists; whether within-group standardization has larger AUC variance in high-missing-rate ethnic groups; whether imputation strategy selection independently contributes to inter-group performance gaps.

Related Reading

Evidence Level: B
References: PMID 39136986, PMID 40585432, PMID 35495871, PMID 38638298, PMID 40968160, PMID 39388258, PMID 33326721, PMID 35901314, PMID 31314618, PMID 22654764, PMID 40896298


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 *