A supervised machine learning semantic segmentation approach for detecting artifacts in plethysmography signals from
Zhicheng Guo1, Cheng Ding2, Xiao Hu2,3
1Department of Computer Science, Duke University, United States of America.
Physiological Measurement
|November 18, 2021
Summary
This study introduces a new algorithm to precisely locate artifacts in photoplethysmography (PPG) signals from wearable devices. This advanced artifact detection improves the reliability of PPG for continuous heart condition screening.
Area of Science:
- Biomedical Engineering
- Signal Processing
- Machine Learning
Background:
- Wearable devices with photoplethysmography (PPG) sensors offer accessible, continuous monitoring for heart conditions.
- Artifacts frequently corrupt PPG signals, hindering accurate analysis and diagnosis.
- Current methods lack precision in identifying the exact location of these signal corruptions.
Purpose of the Study:
- To develop a supervised algorithm for precise artifact localization within PPG signals.
- To improve the reliability of PPG data for early diagnosis and screening of cardiac conditions.
Main Methods:
- Artifact detection framed as a 1D segmentation problem.
- Novel approach combining an active-contour-based loss with an adapted U-Net architecture.
- Algorithm trained and validated on multiple PPG datasets (PPG DaLiA, WESAD, TROIKA).
Main Results:
- The proposed method significantly outperformed baseline approaches across all evaluated datasets.
- Achieved high DICE scores: 0.8734 on PPG DaLiA, 0.9114 on WESAD, and 0.8050 on TROIKA.
- Demonstrated superior performance compared to the next best method by approximately 7 percentage points.
Conclusions:
- The developed algorithm precisely identifies artifact locations in PPG signals.
- This granular artifact information surpasses previous binary quality assessments.
- Enables more informed development of algorithms for cardiac arrhythmia detection.
