Preictal Period Optimization for Deep Learning-Based Epileptic Seizure Prediction

TL;DR

The paper addresses epileptic seizure prediction from scalp EEG and the challenge of defining the preictal period. It introduces a CNN-Transformer classifier to detect spatiotemporal preictal dynamics and a new CIOPR metric for comprehensive, continuous-input performance evaluation, enabling per-patient OPP selection. Applied to 19 pediatric CHB-MIT subjects with subject-specific LOOCV, the approach achieves high sensitivity (99.31%), specificity (95.34%), AUC (99.35%), and F1 (97.46%), with an average prediction horizon of 76.8 minutes before onset. CIOPR reveals substantial inter- and intra-patient variability in optimal preictal length and demonstrates the importance of customizing OPP for practical usability in clinical seizure prediction systems.

Abstract

Accurate prediction of epileptic seizures could prove critical for improving patient safety and quality of life in drug-resistant epilepsy. Although deep learning-based approaches have shown promising seizure prediction performance using scalp electroencephalogram (EEG) signals, substantial limitations still impede their clinical adoption. Furthermore, identifying the optimal preictal period (OPP) for labeling EEG segments remains a challenge. Here, we not only develop a competitive deep learning model for seizure prediction but, more importantly, leverage it to demonstrate a methodology to comprehensively evaluate the predictive performance in the seizure prediction task. For this, we introduce a CNN-Transformer deep learning model to detect preictal spatiotemporal dynamics, alongside a novel Continuous Input-Output Performance Ratio (CIOPR) metric to determine the OPP. We trained and evaluated our model on 19 pediatric patients of the open-access CHB-MIT dataset in a subject-specific manner. Using the OPP of each patient, preictal and interictal segments were correctly identified with an average sensitivity of 99.31%, specificity of 95.34%, AUC of 99.35%, and F1- score of 97.46%, while prediction time averaged 76.8 minutes before onset. Notably, our novel CIOPR metric allowed outlining the impact of different preictal period definitions on prediction time, accuracy, output stability, and transition time between interictal and preictal states in a comprehensive and quantitative way and highlighted the importance of considering both inter- and intra-patient variability in seizure prediction.

Figures (6)

• Figure 1: Architecture of the CNN-Transformer deep learning model. Inputs were of shape $1280\times 23$, for 5-second 23-channel EEG segments. Dimensions $T$, $C$, and $F$ correspond to time points, channels, and feature maps respectively. The shape of the data following each layer is displayed as $T\times C \times F$. For the transformer stage, $C$ and $F$ are aggregated into a single dimension and expressed as $T\times (C, F)$. Specific information on each layer can be found in the Supplementary Material, Table 3.
• Figure 2: Continuous output predictions based on 600 minutes of EEG data preceding seizure onset, along with the fitted sigmoidal curve. The black dotted vertical lines, derived from the sigmoid fit, denote the start and end points of the transition period between interictal and preictal states. The red dotted line, directly calculated from the output, designates the start of the seizure prediction convergence, where the classifier's output converges. Notations: $ND$ = Negative Duration (interictal predictions), $TP$ = Transition Period, and $SPC$ = Seizure Prediction Convergence.
• Figure 3: Four figures showing different levels of fitting between the classifier's output and the fitted sigmoid curve, sorted with decreasing Pearson correlation coefficient. The horizontal axis represents minutes before seizure onset (seizure onset at utmost right), and the vertical axis is the classification output. Captions show the case number, seizure ID, preictal class definition, and Pearson correlation coefficient.
• Figure 4: Output profiles of four patient-specific classifiers for case chb02 with 60, 45, 30, and 15 minutes preictal class definitions respectively. Input is 5.8 hours of continuous EEG recordings before the onset of seizure #1 of chb02. The red dotted line shows the seizure prediction convergence, and the black lines show the transition period boundaries. The horizontal axis represents minutes before seizure onset (seizure onset at utmost right), and the vertical axis is the classification output. Sub-captions show the preictal class definition, CIOPR value, and F1-score achieved.
• Figure 5: Output profiles of four patient-specific classifiers for case chb07 with 60, 45, 30, and 15 minutes preictal class definitions, respectively. Input is 7.0 hours of continuous EEG recordings before the onset of seizure #1 of chb07. The red dotted line shows the seizure prediction convergence, and the black lines show the transition period boundaries. The horizontal axis represents minutes before seizure onset (seizure onset at utmost right), and the vertical axis is the classification output. Sub-captions show the preictal class definition, CIOPR value, and F1-score achieved.