2 - Machine Learning–Based Prediction of Seizure Risk in Neonates Using One-Hour EEG

Publication Number: 1001.2

Vivek V. Shukla, University of Alabama at Birmingham, Birmingham, AL, United States; Avinash Singh, Children's of Alabama, Birmingham, AL, United States; Manasa Kalidindi, University of Alabama School of Medicine, Birmingham, AL, United States; Tao Chen, UAB, Spanish Fort, AL, United States; Donald T. King, Children's of Alabama EEG, Birmingham, AL, United States; Stephen A. Walker, Children's of Alabama, Birmingham, AL, United States; Salman Rashid, University of Alabama School of Medicine, Birmingham, AL, United States; Namasivayam Ambalavanan, University of Alabama School of Medicine, Birmingham, AL, United States; Rachel June Smith, University of Alabama at Birmingham, Birmingham, AL, United States; Arie Nakhmani, University of Alabama at Birmingham, Birmingham, AL, United States

Background: Neonatal seizures are often subclinical and associated with adverse neurodevelopmental outcomes. Continuous EEG monitoring is the diagnostic standard but is resource-intensive and is often unavailable in many units.

Objective: To evaluate the feasibility and predictive performance of machine learning (ML) models using a one-hour EEG data for early seizure prediction in neonates (≥ 35 weeks' gestation).

Design/Methods: One-hour EEG segment corresponding to the second hour of recording was extracted, excluding the setup hour to avoid poor-quality signals. Nineteen channels, including EEG, ECG, EMG, and SpO₂, were consistently available across all participants. Modeling approaches included (1) a baseline model using mean and standard deviation, and (2) an enhanced model incorporating time- and frequency-domain features. Following z-score normalization and LASSO feature selection (|coef| ≥ 0.5), five models-Decision Tree, Random Forest, Gradient Boosting, AdaBoost, and Logistic Regression-were trained using patient-level splits. Model discrimination (area under the receiver operating characteristic curve, AUC, and precision-recall AUC, PR AUC) and calibration (Brier score) were evaluated with 95% bootstrap confidence intervals

Results: A total of 20 neonates were included, including 10 with EEG-confirmed seizures and 10 without. The mean ± standard deviation gestational age was 38.4 ± 1.2 weeks and birth weight was 3,106 ± 466 grams. Models using enhanced time- and frequency-domain features outperformed baseline models. The Decision Tree model performed best, achieving an AUC of 0.78 [0.44-1.00], PR AUC of 0.71 [0.33-1.00], and a Brier score of 0.23 [0.00-0.53], indicating strong discrimination and good calibration (Table 1). Feature analysis identified spectral entropy and relative alpha power as the most important variables, both of which were lower in infants with seizures. (Figure 1).

Conclusion(s): These findings demonstrate the feasibility of ML-based seizure prediction using one-hour EEG data and support efforts to develop short-duration, multimodal, and interpretable seizure risk models reducing dependence on continuous EEG and enabling early, automated risk detection in neonatal neurocritical care.

Table 1: Model Performance Comparisons
AUC= Receiver Operating Characteristic Area Under the Curve; PR AUC= Precision-Recall AUC

Figure 1: A. Relative alpha power (8-12 Hz) by channel and B. Spectral entropy by channel.