Multimodal Signal Edge-Optimised Deep Learning Model for Real-Time Auditory Attention Classification

This work proposes a deep-learning framework for the real-time auditory attention classification using multimodal physiological signals, integrating electroencephalogram (EEG) and photoplethysmogram (PPG) signals. The architecture combines a multi-scale one-dimensional (1D) convolutional backbone to capture features across varying temporal resolutions, followed by a channel-wise attention mechanism using a Squeeze-andExcite (SE) block to enhance feature selectivity. A Bi-directional Gated Recurrent Unit (Bi-GRU) is incorporated to model sequential dependencies within the multimodal time-series data. The trained model is quantised, optimised, and deployed on the STM32L496G-DISCO microcontroller using the X-CUBEAI toolchain. Deployment on board demonstrated a low memory footprint (RAM: 57.91 KiB, Flash: 31.47 KiB) and real-time inference capability with an average latency of 476.9 ms and accuracy of 79.75%. The proposed model achieved an overall classification accuracy of 93.39%, Sensitivity of 88.72%, Precision of 91.69%, and F 1 -score of 92.07%, and AUC values of 0.96 across all classes.