Machine Learning for Epilepsy Diagnosis: Challenges in Low-Resource Settings

Diagnosing epilepsy is especially challenging in low- and middle-income countries (LMICs), where access to trained clinicians and specialised equipment like EEGs is limited. Our latest study, published by Oxford Epilepsy Research Group, examines how machine learning (ML) models can aid in identifying active convulsive epilepsy in these settings. While ML offers a promising solution, our research highlights a critical issue: models often perform poorly when applied outside the regions they were trained in.

Using data from five sub-Saharan African sites, we evaluated "region-naïve" models, which are trained in one location and deployed in another. These models showed excellent performance within their original regions but struggled when tested on unfamiliar settings. For instance, a model trained in Kenya performed well locally but saw significant drops in accuracy when applied to data from South Africa or Ghana. This variability underscores the importance of adapting ML models to the cultural and clinical contexts where they will be used.

Our findings have significant implications for the development of diagnostic tools for epilepsy in LMICs. A one-size-fits-all approach is unlikely to work. Instead, we recommend incorporating diverse datasets during model training and validating models with local data before deployment. By doing so, we can improve the reliability of ML tools and help bridge the diagnostic gap for millions of people living with epilepsy in resource-limited settings.

Read the publication here.