An interpretable machine learning approach for predicting drug-resistant epilepsy in children with tuberous sclerosis complex.

Background: This study developed and validated an interpretable machine learning (ML) algorithm for predicting the risk of drug-resistant epilepsy (DRE) in children with Tuberous sclerosis (TSC).

Methods: To estimate the risk of DRE in pediatric TSC patients, an interpretable ML model was developed and validated. Clinical data were retrospectively collected from 88 pediatric patients with TSC-related epilepsy. 9 ML algorithms were applied, such as random forest (RF), to construct predictive models. To improve interpretability, SHapley Additive exPlanations (SHAP) were employed, providing both global and individualized feature importance explanations.

Results: The RF model outperformed all other algorithms, yielding an AUC of 0.862 and a specificity of 0.930. Key predictors of DRE included a history of infantile epileptic spasms syndrome (IESS), multifocal discharges on EEG, three or more cortical tubers, and the use of three or more antiseizure medications (ASMs). The model was further evaluated using tenfold cross-validation and showed good calibration and clinical utility, as confirmed by decision curve analysis (DCA).

Conclusion: The RF-based prediction model provides a valuable tool for early identification of children with TSC at high risk for DRE, supporting individualized treatment decisions. The integration of SHAP improves model transparency and enhances clinical interpretability.