Machine Learning aided Epilepsy Treatment

Prof. Jimeng Sun, funded by UCB, Apr 2015 - Present

Approximately 2.2 million of people, about 1% of population, in the U.S. are suffered from epilepsy. However, there is no golden rule for clinicians to make a decision which medicine, among over two dozen of anti-epileptic drugs (AEDs), should be given to patient regardless of whether the patient is newly diagnosed with epilepsy or has failed to be seizure-free with the previous AED.

First of all, classification, predictive modeling to be more accurate, of highly refractory, drug resistant, epilepsy patients from general epilepsy patients population using deep learning, especially Recurrent Neural Network is applied. Raw dataset of claim records for about 30 million patients are processed using big data analytics techniques on scalable systems such as Spark and Hadoop to construct qualified study cohort by using expert defined inclusion/exclusion criteria. Embedding for multi-level medical codes are learned and used as input to Recurrent Neural Network which utilize temporal information of medical history for each patient. Traditional classification algorithms such as Logistic Regression, SVM and Random Forest are also evaluated for performance comparison.

We also study a couple of possible approach toward clinical decision support system. First, we are developing a predictive model of drug stability of each AED for epilepsy patients using supervised learning algorithms. Moreover, we are developing a methodology of personalized treatment recommendation for epilepsy patients by using reinforcement learning algorithm.


. Big Data Analytics of Medical Claims Data for Early Prediction of Drug Resistant Epilepsy. In Neurology, submitted, 2017.