Abstract & Overview
This paper presents a novel architecture that integrates optimized Transformer models with bidirectional classifiers to enhance the accuracy and robustness of psychophysiological metric analysis. By leveraging the self-attention mechanisms of Transformers to capture long-range temporal dependencies within physiological signals, such as heart rate variability, electrodermal activity, and pupil response patterns, the proposed model addresses the inherent noise and non-stationarity of biometric data. The architecture employs a bidirectional classification head to process signal sequences in both forward and backward temporal directions, ensuring a comprehensive contextual representation that significantly improves the detection of affective states and stress levels. Experimental evaluations from this paper demonstrate that this hybrid approach outperforms traditional deep learning benchmarks, offering a computationally efficient and highly sensitive framework for real-time monitoring of human psychophysiological health in diverse, non-controlled environments.