Seamless integration for enhanced seizure prediction using HybridConvMobileNet on Typhoon HIL

Real-time seizure prediction is essential for enabling timely interventions that significantly improve patient outcomes. Therefore, in this present work, we have introduced HybridConvMobileNet, a novel hybrid model that integrates 1D convolutional neural networks (CNN) with the MobileNet model to achieve efficient and accurate seizure prediction. The proposed model uses 1D Short-Time Fourier Transform (STFT) coefficients from pre-processed EEG data as input features. In the proposed algorithm, the developed 1D CNN framework captures the critical spatial features from frequency-domain EEG data, while MobileNet network enhances computational efficiency and speed, making the model highly appropriate for real-time applications. The efficacy of the developed model is corroborated on the Children's Hospital Boston-Massachusetts Institute of Technology (CHB-MIT) and Siena benchmark datasets. On the CHB-MIT dataset, the model reached 99.70% accuracy, 99.31% sensitivity, and a 99.43% F1-score, while on the Siena dataset, it reached 99.67% accuracy, 99.08% sensitivity, and a 99.57% F1-score, outperforming eight existing methods across both datasets. Furthermore, real-time implementation on the Typhoon HIL emulator with embedded C2000 microcontrollers demonstrated a low mean detection latency of 0.1 to 1 second, underscoring its potential for clinical applications in seizure monitoring and control.