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

Area of Science:

• Computational neuroscience
• Pediatric neurology
• Machine learning in medicine

Background:

• Tuberous Sclerosis Complex (TSC) is a genetic disorder associated with epilepsy.
• Drug-resistant epilepsy (DRE) poses a significant challenge in managing pediatric TSC.
• Predictive tools for DRE risk in TSC are crucial for timely intervention.

Purpose of the Study:

• To develop and validate an interpretable machine learning (ML) algorithm for predicting DRE risk in children with TSC.
• To enhance the transparency and clinical utility of predictive models for DRE in pediatric TSC.

Main Methods:

• Retrospective collection of clinical data from 88 pediatric patients with TSC-related epilepsy.
• Application of 9 ML algorithms, including Random Forest (RF), to build predictive models.
• Utilizing SHapley Additive exPlanations (SHAP) for model interpretability and feature importance analysis.

Main Results:

• The RF model demonstrated superior performance with an AUC of 0.862 and specificity of 0.930.
• Identified key predictors of DRE: infantile epileptic spasms syndrome (IESS) history, multifocal EEG discharges, multiple cortical tubers, and polypharmacy (≥3 ASMs).
• Validated model performance through tenfold cross-validation and decision curve analysis (DCA), confirming clinical utility.

Conclusions:

• The developed RF-based prediction model aids in the early identification of children with TSC at high risk for DRE.
• Improved model interpretability via SHAP facilitates individualized treatment decisions for pediatric TSC patients.
• The tool supports clinicians in managing DRE risk in TSC, enhancing patient care.