ML and DL Model Comparison with Skin Tone and Dynamic Range Analysis for rPPG

Remote-photoplethysmography (rPPG) is a valuable noninvasive technique for estimating Heart Rate (HR) by analyzing videos or images to measure blood volume changes in the skin, making it suitable for re-mote health monitoring and predicting cardiac diseases. This study evaluates Machine Learning (ML) and Deep Learning (DL) models for HR estimation from rPPG signals across facial regions. Among the models, decision tree and random forest performed the best, delivering higher accuracy and lower error rates compared to DL models, particularly in capturing HR patterns. However, performance varied across facial regions, suggesting a need for region-specific tuning. The impact of skin tone diversity was also assessed and balancing the dataset improved accuracy and fairness across all skin tones. Additionally, the model's dynamic range performance showed consistent accuracy at lower HRs, with increased errors in mid-range HRs, while maintaining robustness at elevated HRs. These findings highlight the need for careful consideration of region-specific factors, skin tone diversity, and dynamic range performance in developing reliable HR estimation models.