Modelling seizure-related predictors of epilepsy diagnostic gap in two urban informal settlements of Nairobi using machine learning.

Background: There is a wide gap in epilepsy diagnosis, particularly in low- and middle-income countries. We used machine learning models to identify seizure-related factors associated with the epilepsy diagnostic gap within the Nairobi Urban Health and Demographic Surveillance System (NUHDSS), Kenya, to inform effective community-level interventions. Methods: Data were drawn from a two-stage, population-based census. In Stage-I, 56,425 residents of NUHDSS were screened for possible convulsive and non-convulsive epilepsy using a standardized questionnaire. In Stage-II, individuals who screened positive were invited for clinical assessment and diagnostic confirmation by neurologists. We used latent class analysis to classify symptom patterns. Seven machine learning models were trained, with extreme gradient boost and random forest models achieving the highest area under the receiver operating characteristic curve (98 %). Results: A total of 528 individuals were diagnosed with epilepsy, among whom 80 % ( Conclusion: There is critically huge diagnostic gap for epilepsy in Nairobi's informal settlements. People with subtle, fewer or less obvious seizure types are more likely to be undiagnosed. These findings highlight the importance of seizure symptom characteristics in understanding patterns of underdiagnosis. Thus, approaches to reducing the diagnostic gap should take into consideration subtle and non-convulsive seizure presentations, such as training on symptom recognition and timely care-seeking.