Predicting efficacy of antiseizure medication treatment with machine learning algorithms in North Indian population

Purpose

This study aimed to develop a classifier using supervised machine learning to effectively assess the impact of clinical, demographical, and biochemical factors in accurately predicting the antiseizure medications (ASMs) treatment response in people with epilepsy (PWE).

Methods

Data was collected from 786 PWE at the Outpatient Department of Neurology, Institute of Human Behavior and Allied Sciences (IHBAS), New Delhi, India from 2005 to 2015. Patients were followed up at the 2nd, 4th, 8th, and 12th month over the span of 1 year for the drugs being administered and their dosage, the serum drug levels, the frequency of seizure control, drug efficacy, the adverse drug reactions (ADRs), and their compliance to ASMs. Several features, including demographic details, medical history, and auxiliary examinations electroencephalogram (EEG) or Computed Tomography (CT) were chosen to discern between patients with distinct remission outcomes. Remission outcomes were categorized into ‘good responder (GR)’ and ‘poor responder (PR)’ based on the number of seizures experienced by the patients over the study duration. Our dataset was utilized to train seven classical machine learning algorithms i.e Extreme Gradient Boost (XGB), K-Nearest Neighbor (KNN), Support Vector Classifier (SVC), Decision Tree (DT), Random Forest (RF), Naïve Bayes (NB) and Logistic Regression (LR) to construct classification models.

Results

Our research findings indicate that 1) among the seven algorithms examined, XGB and SVC demonstrated superior predictive performances of ASM treatment outcomes with an accuracy of 0.66 each and ROC-AUC scores of 0.67 (XGB) and 0.66 (SVC) in distinguishing between PR and GR patients. 2) The most influential factor in discerning PR to GR patients is a family history of seizures (no), education (literate) and multitherapy with Chi-square (χ2) values of 12.1539, 8.7232 and 13.620 respectively and odds ratio (OR) of 2.2671, 0.4467, and 1.9453 each. 3). Furthermore, our surrogate analysis revealed that the null hypothesis for both XGB and SVC was rejected at a 100 % confidence level, underscoring the significance of their predictive performance. These findings underscore the robustness and reliability of XGB and SVC in our predictive modelling framework.

Significance

Utilizing XG Boost and SVC-based machine learning classifier, we successfully forecasted the likelihood of a patient's response to ASM treatment, categorizing them as either PR or GR, post-completion of standard epilepsy examinations. The classifier’s predictions were found to be statistically significant, suggesting their potential utility in improving treatment strategies, particularly in the personalized selection of ASM regimens for individual epilepsy patients.