Classifying seizures in practice: Testing the 2025 ILAE framework among clinicians in South Africa

The 2025 International League Against Epilepsy (ILAE) seizure classification introduced a basic version for nonspecialists, but its usability is untested. We evaluated whether a multidisciplinary frontline workforce in South Africa could apply it, how consistently, and whether performance improved after self-directed exposure.

In a prospective, single-arm, paired pre-post study, 50 clinicians in neurology, internal medicine, emergency medicine, and nursing each classified the same 25 vignettes (37 seizure types) before and after self-directed reading of the 2025 practical guide, serving as their own controls and generating 3700 classifications. Free-text responses were scored against a two-epileptologist consensus reference at two nested levels: the main seizure class (Layer 1) and the class plus basic seizure type (Layer 2). Mixed-effects models with crossed random intercepts for participant and seizure-type unit estimated change; agreement used Fleiss kappa and Gwet AC1.

Layer 1 accuracy rose from 39.2% to 54.4% and Layer 2 from 14.9% to 34.1% (odds ratio [OR] = 2.57, 95% confidence interval [CI] = 2.18-3.03 and OR = 2.76, 95% CI = 2.40-3.18; both p < .001). All four disciplines improved, and Layer 2 gains were greater outside neurology (ratio of cumulative ORs = 1.44, 95% CI = 1.09-1.91; p = .011), although neurologists remained the most correct and nurses the least. Experience, sector, and first language showed no independent association. Agreement improved only modestly; participants became more often correct without becoming more alike. Two thirds of classifications still missed the seizure type, and generalized onset seizures misclassified as focal, the error most likely to prompt inappropriate treatment, rose from 13.9% to 18.9%.

This is the first evaluation of the 2025 ILAE classification in a sub-Saharan African context, and the first anywhere to include the multidisciplinary workforce that classifies most seizures. Unsupported exposure produced measurable improvement, showing the framework is learnable; yet accuracy and consistency remained low, and the most consequential error became more common. These findings support deliberate training, dissemination built for nonspecialists, and translation into the languages in which seizures are described.