EEG Seizure Detection under a Latency Budget: A Comparison of Constrained Feature-Selection Methods

Abstract

Wearable electroencephalography (EEG) seizure detectors must balance detection performance against the cost of extracting and processing large feature sets. This study formulates feature-subset selection as a binary constrained-optimization problem with an explicit latency budget. An imbalance-aware Extreme Gradient Boosting (XGBoost) classifier is optimized for seizure-class F1 on a processed CHB-MIT dataset, and four constraint-handling strategies are compared: penalty function, logarithmic barrier, Lagrangian dual, and constrained Bayesian optimization. All four produced feasible subsets with different performance–latency trade-offs. The penalty method selected 16 of 40 features, reduced estimated single-window latency from 16.02 to 6.54 ms, and increased validation seizure-class F1 from 0.7405 to 0.7921. After the subset was fixed, the final classifier was refitted without using the held-out test partition. Test performance was F1=0.7390, macro F1=0.8197, sensitivity = 0.7519, specificity = 0.8946, a rea under the receiver operating characteristic curve (ROC-AUC)=0.9221, and accuracy =85.59%. The results show how an explicit engineering budget can guide compact EEG feature selection; the reported latency is platform-specific and is not presented as a universal clinical deadline.