{Joachim A Behar and Panayiotis A KyriacouCallum PettitJonathan MantKarthik BudidhaPhilip J AstonElisa Mejía-MejíaKevin KotzenPeter H Charlton}
摘要
光电容积脉搏波(Photoplethysmogram, PPG)信号广泛应用于脉搏血氧仪和智能手表中。在分析PPG信号的过程中,心搏检测是一项基础性步骤。尽管已有多种PPG心搏检测算法被提出,但目前尚不清楚哪种算法表现最优。研究目的:本研究旨在实现以下三个目标:(i)构建一个用于设计与测试PPG心搏检测算法的框架;(ii)评估不同应用场景下PPG心搏检测算法的性能表现;(iii)探究患者人口统计学特征与生理状态对算法性能的影响。研究方法:基于来自八个数据集的PPG与心电图(ECG)同步数据,对15种心搏检测算法进行了评估。采用F1分数作为性能评价指标,该指标综合了敏感性(sensitivity)与阳性预测值(positive predictive value)。主要研究结果:在无运动干扰的情况下,有8种算法表现良好,在医院数据及静息状态下采集的可穿戴设备数据中,F1分数均达到≥90%。然而,在运动状态下,算法性能显著下降,F1分数范围为55%–91%;在新生儿中,算法表现劣于成人,F1分数为84%–96%(新生儿)对比98%–99%(成人);在心房颤动(atrial fibrillation, AF)患者中,F1分数为92%–97%,显著低于正常窦性心律下的99%–100%。研究意义:两种PPG心搏检测算法——“MSPTD”与“qppg”表现最佳,且二者在性能特征上具有互补性。本研究结果可为选择合适的PPG心搏检测算法提供实证依据。所有相关算法、数据集及评估框架均公开免费获取,便于后续研究与应用。
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| photoplethysmography-ppg-beat-detection-on | WFD: Wavelet Foot Delineation | F1 score: 86.5 |
| photoplethysmography-ppg-beat-detection-on | ATM: Adaptive Threshold Method | F1 score: 71.1 |
| photoplethysmography-ppg-beat-detection-on | MSPTD: MultiScale Peak & Trough Detection | F1 score: 97.5 |
| photoplethysmography-ppg-beat-detection-on | Pulses | F1 score: 96.6 |
| photoplethysmography-ppg-beat-detection-on | qppg: Adapted Onset Detector | F1 score: 96.9 |
| photoplethysmography-ppg-beat-detection-on | AMPD: Automatic Multiscale Peak Detection | F1 score: 97.2 |
| photoplethysmography-ppg-beat-detection-on | PWD: Pulse Wave Delineator | F1 score: 92.9 |
| photoplethysmography-ppg-beat-detection-on | ERMA: EventRelated Moving Averages | F1 score: 93.6 |
| photoplethysmography-ppg-beat-detection-on | IMS: Incremental Merge Segmentation | F1 score: 93.6 |
| photoplethysmography-ppg-beat-detection-on | PDA: Peak Detection Algorithm | F1 score: 92.2 |
| photoplethysmography-ppg-beat-detection-on | ABD: Automatic Beat Detection | F1 score: 96.8 |
| photoplethysmography-ppg-beat-detection-on | HeartPy | F1 score: 95.6 |
| photoplethysmography-ppg-beat-detection-on | SPAR: Symmetric Projection Attractor Reconstruction | F1 score: 95.3 |
| photoplethysmography-ppg-beat-detection-on | SWT: Stationary Wavelet Transform | F1 score: 59.0 |
| photoplethysmography-ppg-beat-detection-on | COppg: Percentile Peak Detector | F1 score: 92.4 |
| photoplethysmography-ppg-heart-rate | PWD: Pulse Wave Delineator | MAPE: 8.4 |
| photoplethysmography-ppg-heart-rate | SWT: Stationary Wavelet Transform | MAPE: 51.0 |
| photoplethysmography-ppg-heart-rate | ABD: Automatic Beat Detection | MAPE: 3.3 |
| photoplethysmography-ppg-heart-rate | ERMA: EventRelated Moving Averages | MAPE: 7.9 |
| photoplethysmography-ppg-heart-rate | IMS: Incremental Merge Segmentation | MAPE: 7.7 |
| photoplethysmography-ppg-heart-rate | HeartPy | MAPE: 4.9 |
| photoplethysmography-ppg-heart-rate | COppg: Percentile Peak Detector | MAPE: 9.6 |
| photoplethysmography-ppg-heart-rate | PDA: Peak Detection Algorithm | MAPE: 7.8 |
| photoplethysmography-ppg-heart-rate | MSPTD: MultiScale Peak & Trough Detection | MAPE: 2.4 |
| photoplethysmography-ppg-heart-rate | Pulses | MAPE: 3.3 |
| photoplethysmography-ppg-heart-rate | WFD: Wavelet Foot Delineation | MAPE: 17.5 |
| photoplethysmography-ppg-heart-rate | ATM: Adaptive Threshold Method | MAPE: 38.5 |
| photoplethysmography-ppg-heart-rate | AMPD: Automatic Multiscale Peak Detection | MAPE: 2.9 |
| photoplethysmography-ppg-heart-rate | qppg: Adapted Onset Detector | MAPE: 3.5 |
| photoplethysmography-ppg-heart-rate | SPAR: Symmetric Projection Attractor Reconstruction | MAPE: 4.6 |