Aims: Continuous monitoring of photoplethysmographic (PPG) signals enables the estimation of various physiological parameters such as heart rate (HR), pulse rate variability (PRV), and blood pressure (BP). However, low-quality PPG segments significantly compromise the reliability of these measurements. This study aims to develop a robust and efficient signal quality assessment (SQA) algorithm using machine learning to classify PPG signal segments as either high or low quality, thereby enhancing the accuracy of downstream physiological measurements.
Methods: A novel dataset comprising 7,402 manually labeled 10-second PPG segments was collected from 50 participants using the Polar Verity Sense optical heart rate sensor (sampling frequency: 55 Hz). Signals were preprocessed using a Butterworth bandpass filter (0.5–8 Hz), segmented, and annotated by experts based on a binary grading scheme. A comprehensive feature extraction process yielded 48 features spanning time-domain, frequency-domain, fiducial point-based, and correlation-based metrics. Several machine learning models were trained and evaluated, including Decision Trees, Support Vector Machines (SVM), Neural Networks, and Ensemble Boosted Trees. Feature selection was performed using the Minimum Redundancy Maximum Relevance (MRMR) criterion, reducing the feature set from 48 to 16 without degrading classification performance.
Results: The Ensemble Boosted Tree model using the GentleBoost method achieved the highest accuracy among all models, with 98.1% during validation. After MRMR-based feature reduction, the model retained strong performance with 97.6% validation accuracy and 97.0% testing accuracy using only 16 features. Key features derived from pulse-pulse interval and template-matching analyses, contributed significantly to classification performance.
Conclusions: The proposed machine learning approach offers a robust, accurate, and computationally efficient method for PPG signal quality assessment. The use of multi-domain features, combined with ensemble learning and feature optimization, enables reliable classification of signal quality. Future work should focus on establishing standardized, large-scale benchmark datasets to improve comparability and generalizability across SQA research.