Comprehensive Summary
This paper, presented by Fu et al., discusses the development and testing of a machine learning (ML) algorithm designed to determine the prognosis of drug-resistant epilepsy (DRE) in pediatric patients with Tuberous Sclerosis (TSC). Nine ML algorithms and patient-to-patient customizability, provided from SHapley Additive exPlanations (SHAP), were combined to create predictive models. These models were then tested using clinical data obtained retrospectively from a cohort of 88 pediatric patients with DRE related to TSC. Of the nine ML algorithms, the easily interpretable random forest (RF) model, based on routine structural imaging and clinical features, yielded the highest classifier (AUC) of 0.862 and specificity of 0.930. It was found that a history of infantile epileptic spasms syndrome (IESS), use of three or more antiseizure medications, multifocal epileptiform discharges, and having three or more cortical tubers were all notable predictors of TSC patients who developed DRE. Results of a decision curve analysis (DCA) corroborated the validity and successful calibration of the model, demonstrating its clinical potential in treatment planning using the RF model.
Outcomes and Implications
Tuberous Sclerosis Complex is a rare genetic disorder that often results in benign tumor growth within multiple organ systems and can manifest neuropsychiatric symptoms, of which epilepsy is the most common. Most TSC-related epilepsy is drug resistent, which often has a delayed diagnosis, making a predictive model important in early detection and planning the appropriate course of treatment. The RF model is well suited for clinical application as it boasts a simple structure, transparency from SHAP values, and uses clinically available tools, making it convenient for physicians to use. The limitations of the study include a small sample size from a single source, comorbidities, and variation in EEG interpretation; before the model is clinically implemented such limitations must be further investigated.