改良ComBat仿射校正在东亚5中心×3设备矩阵下的偏差控制能力实证:基于锚点回打分的PPG-SDPTG跨中心跨设备可解释校正

Written by

in

Language

Evidence level: B

一句话:把MRI radiomics里已经成熟的”ComBat仿射校正在锚点回打分下能恢复多少已知偏移”这一方法学,移植到东亚5中心×指夹/腕表/耳夹3种PPG设备的SDPTG(b/a比)矩阵上,用仿真+实测双轨量化”改良ComBat”在中医脉象现代化数据基础设施中能压住多少设备间/中心间系统偏差。

1. 背景与本池定位

池子③承接#100 SDPTG切点在不同BMI/年龄段人群中的本地化校正(2026-10-10 YY第28轮),其结论能否在多中心真正落地,被一个”工程问题”卡住:东亚5个临床中心若使用3种不同PPG设备(指夹/腕表/耳夹),b/a比切点是否在跨中心跨设备后仍然可比?

多中心生物信号研究的标准做法是ComBat仿射校正(location/scale empirical Bayes)。ComBat最早用于基因芯片批次效应(Johnson 2007),后被MRI radiomics大量验证(PMC9695920, PMC10403369, PMC11950493),近年还有AutoComBat等泛化方法(PMC9325761)。其核心假设是批次效应可表示为”加性位置偏移+乘性尺度偏移”,并以经验贝叶斯方式从参考样本中估计参数。

对PPG-SDPTG来说,设备-中心效应正是典型的”加性+乘性”复合偏移:
– 加性偏移:不同LED波长(绿光/红光/红外)与光电探测器基线漂移 → SDPTG基线抬升/下降
– 乘性偏移:不同采样率/按压深度/皮肤耦合 → SDPTG波幅缩放
– 耦合偏移:b波、c波相对a波、d波、e波的位置漂移

这意味着ComBat仿射模型在结构上是适配的。但要把它从”已发表MRI证据”翻译到”PPG-SDPTG多中心多设备”,需要:
1. 在每个中心每种设备上保留锚点样本(anchor / phantom / reference subject)
2. 用锚点样本估计设备间/中心间的location-scale参数
3. 用回打分方法量化恢复率(锚点真实值 vs 校正前后估计值的偏差)
4. 检验未用锚点(unsupervised ComBat)与用锚点(reference-based ComBat)的恢复率差
5. 在BMI×年龄分层切点的背景下,验证校正后切点是否仍能保留分层特异性

2. 改良ComBat的”改良”在哪里

经典ComBat假设”批次参数对所有特征一致”。在PPG-SDPTG场景下,b/a比、c/a比、d/a比、e/a比对设备-中心偏移的敏感度不同——b波最容易受按压深度影响(乘性敏感),c/d波容易受皮肤温度与采样率影响(加性敏感)。因此需要:

  • 特征分层先验:按”对设备敏感度”对SDPTG 5个比值分组,组内共用location-scale参数
  • 协变量保护:把BMI、年龄、性别作为”生物学协变量”嵌入ComBat的preserve参数,避免校正同时抹掉真正的生理梯度
  • 锚点回打分作为校准器:在每个中心每种设备预录入10-20名”参考志愿者”,其b/a比事先在金标准指尖光电传感设备(已与PPG心电同步)上测量,作为”真值”——校正后用b/a恢复率量化偏差控制

3. 实验设计要点

参见同池 08-ops/experiments/pool3-modified-combat-affine-calibration-east-asian-ppg-multi-center-multi-device-anchor-rescore.md (P3-EXP-040)。三阶段:

  • 阶段1 仿真注入已知偏移 → 恢复率量化
  • 起点:单一中心(n=600)金标准PPG数据集 → 注入15种”仿真批次”(3设备 × 5中心组合)
  • 对比:无校正 / 未锚点ComBat / 锚点ComBat / 改良ComBat(特征分层+协变量保护)
  • 终点指标:b/a比RMSE恢复率、BMI×年龄分层AUC保留率

  • 阶段2 锚点回打分实测

  • 5中心各20名参考志愿者,金标准与3设备平行采集
  • 终点指标:跨设备b/a比偏差(均值差、SD差、ICC)、BMI分层切点漂移量

  • 阶段3 中医证候标签子样本验证

  • 在校正后数据上复测P3-EXP-039阶段3的痰湿证vs对照ROC,检验AUC是否保留
  • 终点指标:校正前后AUC差异ΔAUC < 0.03为可接受

4. 证据链与证据边界

已有证据(E1-E2层):
– ComBat仿射校正在多中心MRI radiomics的有效性(PMC9695920, PMC10403369, PMC11950493)
– AutoComBat对非正态/多模态特征的鲁棒性(PMC9325761)
– 韩国300人SDPTG b/a比年龄参考曲线(PMC9819370)
– PPG血管老化评估(41598-022-15240-4)
– 韩国老年女性PPG指夹vs腕表差异(sensors-21-07782, 2021)
– SDPTG与血压/心率/抑郁的关联(PMC6882620)
– 腕-指双通道PPG数据集(PMC11790827)

证据边界(E2-E3层,本课题性质):
– 上述文献中没有任何一项在”东亚5中心×3设备×中医证候标签”复合场景下做ComBat仿射校正
– 锚点回打分在MRI是金标准做法,在PPG-SDPTG尚无标准方案
– 设备间b/a比差异的”位置-尺度”分解是否成立,需阶段1仿真验证
– 中医证候标签是离散分类,ComBat的preserve参数仅对连续协变量设计,对离散证候标签需额外检验

未涉及的层面:
– 不同手指(拇指/食指/中指)皮肤厚度的系统差异 → 留给P3-EXP-041
– 跨种族外推(东亚vs白人vs非裔b/a比参考区间) → 已在#111立项
– 设备固件版本漂移 → 留给#108-2

5. 风险与失败模式

  • 过度校正:把BMI梯度的真实差异抹平 → 通过preserve参数+协变量保护降低
  • 锚点代表性不足:20名/中心若不能代表当地BMI/年龄谱,参数估计有偏 → 阶段2必须按BMI×年龄分层配额招募
  • 设备固件升级:阶段2/3期间若设备固件升级,参数需重估计 → 引入固件版本作为批次协变量
  • 证候标签漂移:阶段3痰湿证辨证标准若与阶段1不同,ΔAUC无可比性 → 沿用P3-EXP-039同一辨证手册

6. 关联深挖(本轮新增P2课题)

# 课题 方向 优先级
112 改良ComBat锚点回打分方案的固件版本漂移鲁棒性 不同固件版本的参数迁移性验证 P2
113 跨指(拇指/食指/中指)皮肤厚度对b/a比加性偏移的校准公式 指间系统偏差的解剖学修正 P2
114 ComBat对中医证候离散标签的”边缘效应”评估 校正后证候-切点联合分布的边缘一致性 P2
115 改良ComBat在长程(≥24h)可穿戴PPG数据上的动态批次识别 漂移作为时间协变量的扩展 P2

7. 相关阅读(本池内)

  • 100 science-yixue/sdptg-ba-ratio-bmi-age-stratified-cutoff-calibration.md (P3-EXP-039, 上游)

  • 26 science-yixue/tcm-pulse-hri-inter-operator-reliability-training-standardization.md (操作员层)

  • 29 science-yixue/pulse-pressure-normalization-hri-algorithm.md (压力归一化层)

  • 37 跨设备G_c(P)迁移学习 (域适应,与ComBat的差异在”模型层 vs 特征层”)

8. 参考文献(PubMed 可查,PMID 已逐条回查)

  • Jaramillo-Jimenez A, et al. ComBat models for harmonization of resting-state EEG features in multisite studies. Clin Neurophysiol. 2024. PMID 39369552 — 多中心研究中用 ComBat(位置-尺度仿射校正)校正站点效应的直接方法学证据,对应本文”改良 ComBat”的基线算法。
  • Chen W, et al. Multicenter Evaluation of Machine-Learning Continuous Pulse Rate Algorithm on Wrist-Worn Device. Digit Biomark. 2024. PMID 39670276 — 腕戴设备连续脉率算法的多中心评估,与本文”东亚 5 中心 × 多设备”矩阵的评估结构一致。
  • Nuuttila OP, et al. Validity of the Wrist-Worn Polar Vantage V2 to Measure Heart Rate and Heart Rate Variability at Rest. Sensors (Basel). 2021. PMID 35009680 — 腕戴 PPG 设备静息状态下测量心率与心率变异性的效度验证,是设备间偏差必须独立校准的直接依据。

9. 一句话总结

把MRI radiomics的”ComBat+锚点回打分”翻译到PPG-SDPTG多中心多设备场景,是把#100 BMI×年龄分层切点从”单中心可用”升级为”东亚5中心可推”的关键工程步骤。改良点在于”特征分层+协变量保护+证候子样本回测”,证据层级B,34个月三阶段落地。


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。

Empirical Validation of Improved ComBat Affine Correction for Bias Control in a 5-Center × 3-Device Matrix: PPG-SDPTG Cross-Center and Cross-Device Interpretable Correction Base…

Evidence level: B

One-sentence summary: Transplant the established methodological framework from MRI radiomics—”how much known bias can ComBat affine correction recover under anchor rescore”—to the SDPTG (b/a ratio) matrix across 3 PPG devices (finger clip/wristband/ear clip) in a 5-center East Asian setting, using simulation and real-world measurement tracks to quantify how much “Improved ComBat” suppresses system-level bias between devices and centers within the modernized TCM pulse data infrastructure.

1. Background and Pool Positioning

Pool #3 follows the localized calibration of SDPTG cutoff points across different BMI and age groups in #100 (2026-10-10 YY Round 28). The question remains whether these conclusions can truly be implemented in a multi-center setting, currently blocked by an “engineering problem”: If 5 clinical centers in East Asia utilize 3 different PPG devices (finger clip/wristband/ear clip), will the b/a ratio cutoff remain comparable across centers and devices?

Standard practice for multi-center biosignal research is ComBat affine correction (location/scale empirical Bayes). Originally developed for batch effects in gene chips (Johnson 2007), ComBat has been extensively validated in MRI radiomics (PMC9695920, PMC10403369, PMC11950493), with recent generalizations like AutoComBat emerging (PMC9325761). Its core assumption is that batch effects can be represented as “additive location shifts + multiplicative scale shifts,” estimated from reference samples via empirical Bayes.

For PPG-SDPTG, device-center effects represent a typical composite “additive + multiplicative” shift:
– Additive shift: Different LED wavelengths (green/red/infrared) and photodetector baseline drifts → SDPTG baseline elevation/depression
– Multiplicative shift: Different sampling rates, pressure depth, and skin coupling → SDPTG amplitude scaling
– Coupling shift: Positional drifts of b-wave, c-wave relative to a-wave, d-wave, and e-wave

This implies the ComBat affine model is structurally adaptable. However, translating it from “published MRI evidence” to “PPG-SDPTG multi-center multi-device” requires:
1. Retaining anchor samples (anchor / phantom / reference subject) at each center and device
2. Estimating location-scale parameters between devices/centers using anchor samples
3. Quantifying recovery rates via rescore methods (deviation between anchor true values and pre/post-correction estimates)
4. Comparing recovery rates between unsupervised ComBat (no anchors) and reference-based ComBat (with anchors)
5. Verifying whether corrected cutoffs retain stratified specificity under BMI × age stratification

2. Where the “Improvement” Lies in Improved ComBat

Classic ComBat assumes “batch parameters are consistent across all features.” In the PPG-SDPTG scenario, b/a ratio, c/a ratio, d/a ratio, and e/a ratio exhibit different sensitivities to device-center shifts—the b-wave is most sensitive to pressure depth (multiplicative), while c/d-waves are prone to skin temperature and sampling rate influences (additive). Therefore, the following adjustments are necessary:

  • Feature-stratified priors: Group the 5 SDPTG ratios based on “device sensitivity,” sharing location-scale parameters within groups
  • Covariate protection: Embed BMI, age, and gender as “biological covariates” into ComBat’s preserve parameters to avoid erasing genuine physiological gradients during correction
  • Anchor rescore as a calibrator: Pre-record 10–20 “reference volunteers” per center/device; their b/a ratios are measured against a gold-standard fingertip photoplethysmographic sensor (synchronized with PPG ECG) as “truth”—post-correction recovery rates quantify bias control

3. Experimental Design Highlights

Refer to the same pool 08-ops/experiments/pool3-modified-combat-affine-calibration-east-asian-ppg-multi-center-multi-device-anchor-rescore.md (P3-EXP-040). Three stages:

  • Stage 1: Simulated known offset injection → recovery rate quantification
  • Starting point: Single-center (n=600) gold-standard PPG dataset → inject 15 “simulated batches” (3 devices × 5 center combinations)
  • Comparison: No correction / Unsupervised ComBat / Anchor-based ComBat / Improved ComBat (feature stratification + covariate protection)
  • Endpoint metrics: b/a ratio RMSE recovery rate, BMI × age stratified AUC retention rate

  • Stage 2: Anchor rescore real-world measurement

  • 20 reference volunteers per center, parallel acquisition via gold standard and 3 devices
  • Endpoint metrics: Cross-device b/a ratio deviation (mean difference, SD difference, ICC), BMI-stratified cutoff drift magnitude

  • Stage 3: TCM syndrome label subsample validation

  • Re-test P3-EXP-039 Stage 3 Phlegm-Dampness vs. Control ROC on corrected data, verifying AUC retention
  • Endpoint metric: Acceptable if pre-post correction ΔAUC < 0.03

4. Evidence Chain and Evidence Boundaries

Existing Evidence (E1-E2 level):
– ComBat affine correction efficacy in multi-center MRI radiomics (PMC9695920, PMC10403369, PMC11950493)
– AutoComBat robustness for non-normal/multimodal features (PMC9325761)
– Korean 300-person SDPTG b/a ratio age reference curve (PMC9819370)
– PPG vascular aging assessment (41598-022-15240-4)
– Korean elderly female finger clip vs. wristband PPG differences (sensors-21-07782, 2021)
– SDPTG associations with blood pressure/heart rate/depression (PMC6882620)
– Wrist-finger dual-channel PPG dataset (PMC11790827)

Evidence Boundaries (E2-E3 level, inherent to this topic):
– None of the above literature conducts ComBat affine correction under the composite scenario of “5 East Asian centers × 3 devices × TCM syndrome labels”
– Anchor rescore is gold-standard in MRI but lacks a standard protocol for PPG-SDPTG
– Whether the “location-scale” decomposition of inter-device b/a ratio differences holds requires validation via Stage 1 simulation
– TCM syndrome labels are discrete classifications; ComBat’s preserve parameter is designed for continuous covariates, requiring additional testing for discrete syndrome labels

Unaddressed Aspects:
– Systematic skin thickness differences across fingers (thumb/index/middle) → reserved for P3-EXP-041
– Cross-race extrapolation (East Asian vs. White vs. African American b/a ratio reference ranges) → already initiated in #111
– Device firmware version drift → reserved for #108-2

5. Risks and Failure Modes

  • Over-correction: Erasing genuine BMI gradient differences → mitigated via preserve parameters + covariate protection
  • Insufficient anchor representation: If 20 volunteers per center do not represent local BMI/age spectra, parameter estimates become biased → Stage 2 must recruit with BMI × age stratified quotas
  • Device firmware upgrades: If firmware updates occur during Stages 2/3, parameters require re-estimation → introduce firmware version as a batch covariate
  • Syndrome label drift: If P3-EXP-039 Phlegm-Dampness diagnosis criteria differ from Stage 1, ΔAUC becomes incomparable → adopt the same diagnosis manual as P3-EXP-039

6. Related Deep Dives (New P2 Topics in This Round)

# Topic Direction Priority
112 Improved ComBat anchor rescore firmware version drift robustness Parameter migration validation across different firmware versions P2
113 Cross-finger skin thickness calibration formula for additive b/a ratio shifts Anatomical correction for inter-digit systematic bias P2
114 ComBat “edge effects” on discrete TCM syndrome labels Edge consistency of post-correction syndrome-cutoff joint distributions P2
115 Improved ComBat dynamic batch identification on long-term (≥24h) wearable PPG data Extension treating drift as a time covariate P2

7. Related Reading (Within This Pool)

  • 100 science-yixue/sdptg-ba-ratio-bmi-age-stratified-cutoff-calibration.md (P3-EXP-039, upstream)

  • 26 science-yixue/tcm-pulse-hri-inter-operator-reliability-training-standardization.md (operator level)

  • 29 science-yixue/pulse-pressure-normalization-hri-algorithm.md (pressure normalization level)

  • 37 Cross-device G_c(P) transfer learning (domain adaptation; difference lies in “model level vs. feature level”)

8. References (PubMed, each PMID checked against the PubMed record)

  • Jaramillo-Jimenez A, et al. ComBat models for harmonization of resting-state EEG features in multisite studies. Clin Neurophysiol. 2024. PMID 39369552 — direct methodological evidence for using ComBat (location-scale affine correction) to correct site effects in multi-center studies, matching the baseline algorithm of this article.
  • Chen W, et al. Multicenter Evaluation of Machine-Learning Continuous Pulse Rate Algorithm on Wrist-Worn Device. Digit Biomark. 2024. PMID 39670276 — a multi-center evaluation of a continuous pulse rate algorithm on wrist-worn devices, matching the “East Asian 5 centers × multiple devices” evaluation structure of this article.
  • Nuuttila OP, et al. Validity of the Wrist-Worn Polar Vantage V2 to Measure Heart Rate and Heart Rate Variability at Rest. Sensors (Basel). 2021. PMID 35009680 — validity validation of a wrist-worn PPG device for heart rate and HRV at rest, the direct basis for requiring independent calibration of inter-device bias.

9. One-Sentence Summary

Translating the “ComBat + anchor rescore” framework from MRI radiomics to the PPG-SDPTG multi-center multi-device scenario constitutes a critical engineering step to upgrade #100’s BMI × age stratified cutoffs from “single-center usable” to “East Asian 5-center extrapolatable.” Improvements focus on “feature stratification + covariate protection + syndrome subsample re-testing,” with evidence level B and a 34-month three-stage implementation timeline.


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 *