Integrating Remote Photoplethysmography and Machine Learning on Multimodal Dataset for Noninvasive Heart Rate...

Area of Science:

• Biomedical Engineering
• Health Informatics
• Machine Learning

Background:

• Non-contact heart monitoring is vital for telemedicine and fitness tracking.
• Remote photoplethysmography (rPPG) offers contactless heart rate estimation but suffers from environmental and skin-related inaccuracies.
• Existing rPPG methods lack robustness due to sensitivity to lighting and skin pigmentation.

Purpose of the Study:

• To develop a more accurate and robust non-contact heart rate estimation method.
• To investigate the efficacy of a multimodal approach combining facial video and physical attributes.
• To compare machine learning models for multimodal heart rate estimation.

Main Methods:

• Collected a local dataset of 60 individuals, including 1-minute facial videos and physical attributes (age, gender, weight, height).
• Derived Body Mass Index (BMI) from collected weight and height data.
• Implemented and compared Support Vector Regression (SVR) and Random Forest (RF) models using the multimodal dataset.

Main Results:

• The multimodal approach significantly enhanced heart rate estimation performance.
• The Random Forest model achieved superior results, outperforming state-of-the-art rPPG methods.
• Key performance metrics for the RF model included MAE of 3.057 bpm, RMSE of 10.532 bpm, and MAPE of 4.2%.

Conclusions:

• Combining facial video with physical attributes provides a more accurate and reliable non-contact heart rate measurement.
• The developed multimodal approach shows significant potential for real-time applications in telemedicine and mass health screening.
• Machine learning, particularly Random Forest, is effective in processing multimodal data for enhanced physiological monitoring.