Prediction of Risk of Infant Seizures using Machine Learning (PRISM) - Seizures are common neurological concerns in newborns and are early indicators of brain injury with an incidence of up to 5 infants per 1000 live births and are associated with increased risk of epilepsy, cerebral palsy, developmental delays, and mortality. Though early intervention can improve neurodevelopmental outcomes, continuous electroencephalography (cEEG), the current gold standard in seizure detection, is time- consuming and requires expert interpretation that is often unavailable in many neonatal intensive care units (NICUs). Existing automated seizure identification models focus on seizure identification and not risk prediction. There is a critical gap and urgent need for automated, real-time seizure risk prediction models that are reliable and efficient, enabling early risk-informed interventions. To address this, we propose the Prediction of Risk of Infant Seizures using Machine-learning (PRISM) project, to develop and validate reliable and clinically interpretable seizure risk prediction models using short EEG (≤1 hour) and multimodal data integrating video and clinical features with short EEG that has not been previously tested. Our hypothesis is that short-duration EEGs (≤ 60 minutes) contain predictive biomarkers of seizure onset within the subsequent 10 to 23 hours, achieving clinically actionable discrimination (receiver operating characteristic area under the curve, ROCAUC ≥ 0.80). To ensure feasibility, our team conducted a study on 20 patients, including 10 without and 10 with EEG confirmed seizures. Five different classifiers were trained and tested, out of which decision tree using time and frequency features resulted as the best performing model, achieving an ROCAUC 0.78 [confidence interval, CI of 0.44-1.00], precision-recall (PR) AUC 0.71 [0.33-1.00], and Brier score 0.23 [0.00-0.53], indicating promising results with good discrimination and calibration, suggesting short EEG contains informative signals needed for seizure risk prediction. In PRISM project, we will develop seizure risk prediction model using EEG only and multimodal models integrating video, clinical, and EEG data. For both models, we will explore performance with 60, 45, 30, 15, 10, and 5-minute epochs to identify the shortest duration of data needed to reliably predict (ROCAUC threshold ≥ 0.70) seizure risk. Additionally, we will explore models with individual data categories alone (video alone and clinical variables alone) to complement the analysis in EEG-only model. In a subsequent step, the best-identified models will be developed into a modular seizure risk prediction tool that is reliable, clinically interpretable, and can be used at the bedside. The PRISM project will provide evidence for the usefulness and effectiveness of multimodal seizure prediction, guiding future endeavors toward real-time, integrated decision support systems, and future multicenter clinical validation. Furthermore, this model could serve as a foundation for future applications in older infants and pediatric populations, broadening its clinical impact. This project aligns well with NINDS priorities in neonatal brain injury, computational neuroscience, and translational biomarker development.
