Personalized ECG-based artificial intelligence models for early seizure prediction

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Seizure prediction is critically important, as it can help prevent serious injuries, improve quality of life, and potentially reduce the risk of SUDEP. Autonomic fluctuations are known to occur during the preictal and early ictal phases; however, the pattern and magnitude of these changes vary considerably across patients and seizure types. This highlights the need for developing patient-specific ECG-based seizure prediction models. In this study, ECG data from 30 patients who had at least six recorded seizures during presurgical VEEG evaluation were used to develop both machine learning and deep learning models. Heart rate variability (HRV) features were extracted from the ECG recordings and used to train classical ML classifiers, while a ResNet-50-based bimodal architecture was employed for deep learning model training. All models were evaluated using fivefold and leave-one-out cross-validation. Seizure prediction was tested at 5-, 10-, and 15-min intervals prior to seizure onset. Both HRV-based machine learning models and deep learning models successfully distinguished interictal from preictal ECG segments, with the bimodal deep learning framework outperforming traditional ML methods. In the patient-specific setting, the deep learning model demonstrated robust diagnostic capabilities, achieving a mean accuracy of 89.9% (SD = 8.2) alongside high precision (ranging from 70.3% to 100%) and sensitivity (ranging from 69.6% to 100%). The highest performance was observed in Patient P13 (99% accuracy, 100% precision, and 98.7% sensitivity) and Patient P14 (98% accuracy, 99.5% precision, and 97.3% sensitivity), with several patients (P24, P25, P26, P30) achieving a perfect 100% across accuracy, precision, sensitivity, and specificity. The lowest performance was seen in Patient P22 (71% accuracy, 70.3% precision, 76.9% sensitivity, and 64.4% specificity). False positive rates (FPR) were effectively minimized in the patient-specific DL models, ranging from zero false positives per hour (achieved in six patients) to a maximum of 13.15 per h (Patient P22). Notably, prediction performance was higher in temporal lobe epilepsy (92% ± 7.6%) compared to extra-temporal lobe epilepsy (87.3% ± 8.4%). In the generalized patient-independent model, the prediction accuracy dropped to 52%, which could be attributed to inter-patient heterogeneity in autonomic signatures, seizure types, and baseline cardiac rhythms. These findings demonstrate that recent advances in ML and DL enable reliable prediction of preictal states using ECG alone. Importantly, our results indicate that training models on long-term ECG recordings containing multiple seizures from the same individual, that is, using a patient-specific training strategy, yields a more dependable solution for seizure prediction than generalized approaches.