SeLECTS-assisted identification based on structured EEG reports: Development and external validation of a multicenter...

Benign epilepsy with centrotemporal spikes (SeLECTS) is a common pediatric epilepsy characterized by high-amplitude central-temporal spikes on electroencephalogram (EEG). Early identification remains challenging owing to inter-reader variability and inter-center protocol differences. Existing models often rely on single-center raw EEG, limiting generalizability and clinical deployment. An explainable, cross-institutional tool built on structured EEG reports plus routine laboratory tests could improve screening and workflow. This study aimed to develop a machine learning model for SeLECTS assistance and validate its performance on an external dataset, comparing multiple algorithms and evaluating discrimination, calibration, and net clinical benefit, with interpretable outputs for quality control and patient stratification. We performed a multicenter retrospective study (internal n = 1067; external n = 94). Among 242 candidate variables, those with >20% missingness were removed; remaining missing data were handled by multiple imputation. Least Absolute Shrinkage and Selection Operator selected 13 key features, including age; central/temporal discharge localization; spikes or spike-wave complexes; background rhythm; facial/limb clonus; and laboratory indices (e.g., lactate dehydrogenase, white blood cell differential, 25-hydroxyvitamin D). Ten common classifiers were trained and tested. Area Under the Receiver Operating Characteristic Curve (AUROC) was the primary metric; accuracy, sensitivity, specificity, and F1 score were secondary. Calibration and decision curve analyses assessed probability agreement and clinical net benefit. SHapley Additive exPlanations provided global and patient-level interpretability. eXtreme Gradient Boosting performed best. Internal validation yielded AUROC 0.97, accuracy 0.91, sensitivity 0.83, and specificity 0.94. External validation showed AUROC 0.96, accuracy 0.88, sensitivity 0.74, and specificity 0.94. Compared with internal performance, external results declined only slightly (AUROC -0.01; accuracy -0.03; sensitivity -0.09; specificity unchanged). Calibration indicated close agreement between predicted and observed risks. Decision curve analysis showed consistent positive net benefit across low-to-moderate thresholds. SHapley Additive exPlanations highlighted temporal discharges, age, central discharges, and Lactate Dehydrogenase as leading contributors, consistent with known SeLECTS characteristics. Using 13 multicenter, routinely available features, the eXtreme Gradient Boosting model demonstrated high discrimination, good calibration, and meaningful net benefit in internal and external cohorts. Designed as a second-view aid after EEG reporting, the interpretable, lightweight model supports risk alerts and tiered review and is suitable for scalable deployment across centers.