Contactless Physiological Parameter Assessment from Facial Videos Using Hybrid Deep Learning and Dynamic Optimization

This research introduces a non-invasive physiological assessment framework to evaluate metrics from facial video sequences via remote photoplethysmography (rPPG) and a deep learning approach. The system inspects facial videos at the frame level to detect minute blood volume fluctuations linked to physiological alterations, allowing determination of heart rate, blood pressure, blood oxygen levels, and body temperature without sensors. Facial areas are identified and spatially adjusted using MTCNN or Dlib to guarantee uniform signal extraction across the video frames. Pre-trained ResNet and EfficientNet models are then utilized to obtain spatial representations. To describe temporal physiological changes, the obtained features are processed with a spike-inspired temporal encoding approach, and a sparse attention-based temporal fusion module selectively merges temporal features while incorporating information related to blood pressure to enhance the accuracy of physiological parameter estimations. Model optimization employs the adaptive gradient modulation optimizer (AGMO), which dynamically promotes better convergence, learning stability, and generalization. Evaluation under various lighting conditions reveals that the proposed framework achieves highly precise physiological assessments, yielding a mean squared error (MSE) of roughly 0.01, surpassing existing methodologies. The results highlight the potential of integrating frame-level rPPG analysis, hybrid deep feature extraction, temporal attention, and adaptive optimization for dependable, real-time, and non-invasive physiological monitoring.