Remote Heart Rate Estimation Using Deep Learning-Based RPPG Models: A Comparative Analysis and LLM-Augmented Insights

Contactless heart rate (HR) monitoring using remote photoplethysmography (rPPG) has emerged as a transformative application in computer vision and biomedical engineering. This research investigates the effectiveness of deep learning-based rPPG models for HR estimation from facial videos. Specifically, we evaluate state-of-the-art models including MTTS-CAN, DeepPhys, PhysNet, and TSCAN on diverse skin tones and environmental conditions. Our findings demonstrate that deep models, when trained on spatial-temporal signals from facial regions, outperform traditional signal processing baselines. Moreover, we integrate Large Language Models (LLMs) to analyze heart rate trends and provide contextual medical insights, highlighting the promise of AI-assisted telemedicine. Experimental results show that MTTS-CAN consistently achieves high accuracy (94.45% at 0.5m) and robustness across lighting and distance variations. This work lays the foundation for future development of AI-driven physiological monitoring systems suitable for health diagnostics, emotion recognition, and deepfake detection.