Preictal period optimization for deep learning-based epileptic seizure prediction

<i>Objective</i>. Accurate seizure prediction could prove critical for improving patient safety and quality of life in drug-resistant epilepsy. While deep learning-based approaches have shown promising performance using scalp electroencephalogram (EEG) signals, the incomplete understanding and variability of the preictal state imposes challenges in identifying the optimal preictal period (OPP) for labeling the EEG segments. This study introduces novel measures to capture model behavior under different preictal definitions and proposes a data-centric deep learning methodology to identify the OPP.<i>Approach</i>. We trained a competent subject-specific CNN-Transformer model to detect preictal EEG segments using the open-access CHB-MIT dataset. To capture the temporal dynamics of the model's predictions, we fitted a sigmoidal curve to the model outputs obtained from uninterrupted multi-hour EEG recordings prior to seizure onset. From this fitted curve, we derived key performance measures reflecting the timing of predictions, including classifier convergence, average error, output stability, and the transition between interictal and preictal states. These measures were then combined to compute the Continuous Input-Output Performance Ratio, a novel metric designed to comprehensively compare model behavior across different preictal definitions (60, 45, 30, and 15 min) and suggest the OPP for each patient.<i>Main results.</i>The CNN-Transformer model achieved state-of-the-art performance (area under the curve of 99.35% and<i>F</i>1-score of 97.46%) using minimally pre-processed EEG signals. The 60-minute preictal definition was associated with earlier seizure prediction, lower error in the preictal state, and reduced output fluctuations, leading to significantly higher CIOPR scores (<i>p</i>< 0.001). Conventional accuracy-related metrics (sensitivity, specificity, F1-score) were less sensitive to varying preictal definitions and often discordant with CIOPR findings. Cross- and intra-patient heterogeneities in the prediction times were also observed, complicating the establishment of a global preictal interval.<i>Significance</i>. The newly developed metrics demonstrate that varying the preictal period significantly impacts the timing of predictions in ways not captured by conventional accuracy-related metrics. Understanding this impact and the inter-seizure heterogeneities is essential for developing intelligent systems tailored to individual patient needs and for underlining practical limitations in detecting the preictal period in real-world clinical applications.