Machine learning models for predicting treatment response in infantile epilepsies

Epilepsy stands as one of the prevalent and significant neurological disorders, representing a critical healthcare challenge. Recently, machine learning techniques have emerged as versatile tools across various healthcare domains, encompassing diagnostics, treatment assessment, and prognosis. We compared 11 machine learning model to find the best ML model to predict drug treatment outcomes for our cohort, which we previously evaluated using classical statistical methods. Methods: In our study, we evaluated patients who presented to the pediatric neurology department of our university hospital with seizures at the age of 1 to 24 months and were diagnosed with epilepsy. We utilized 11 different machine learning techniques namely Decision Tree, Bagging, K-Nearest Neighbour, Linear Discriminant Analysis, Logistic Regression, Neural Networks, Deep Neural Networks, Support Vector Machine. Besides, we compared these techniques using various performance metrics to identify anti-seizure medicine response. We also utilized the chi-square feature selection methods to enhance performance in machine learning algorithms. Results: Two hundred and twenty-nine patients (110 male and 119 female) who were diagnosed between the ages of 1-24 months were included in the study. Support Vector Machine algorithm was found to be effective in drug resistant epilepsy detection, with the highest aure under curve value (0.9934) and achieving a test accuracy of 97.06 %. Conclusion: This study can shed light on future studies by showing that the Support Vector Machine algorithm can effectively determine the drug resistant epilepsy. The pediatric neurologist and experts should be referred to nonmedical treatment (epilepsy surgery, ketogenic diet) at the early stages and multidisciplinary approach should be provided.

Journal

EPILEPSY & BEHAVIOR

Year

2024

Volume

160

ISSN

1525-5050

eISSN

1525-5069

DOI

10.1016/j.yebeh.2024.110075

WoS UID

WOS:001335016900001