Prognosis of Epileptic Seizure Event Onsets Using Random Survival Forests

Abstract

Abstract This article introduces a machine learning approach, based on a nonparametric, decision-tree-based random survival forest (RSF) model, for a continuous prognosis of epileptic seizure events using electroencephalogram (EEG) data. While earlier seizure prediction methods forecast seizure occurrences only at a specified future time, the RSF model allows estimation of the probability of seizure onset, in terms of a hazard function, over the entire prediction horizon. These estimates are crucial for developing individualized quantitative risk measures and management plans for epilepsy patients. Additionally, RSF can identify the key risk factors by capturing the interdependencies among the features extracted from EEG data. The performance of RSF was evaluated for prognosing seizure onsets of the rat and mice specimens in an 80 small animals cohort at the Texas A&M Department of Neuroscience and Experimental Therapeutics. The results suggest that RSF outperforms other contemporary survival models, including the popular Cox proportional hazard, with 87.5% lower integrated Brier Score (IBS) errors, and 17.5% higher concordance index (C-index). Further, a continuous seizure prediction sensitivity of 83% and specificity of 87% were obtained even over a 5-min prediction horizon (the average time between successive seizure onsets was 5 min long). These results suggest that the RSF model can be used to effectively quantify the likelihood of seizure onsets over time to the patients and caregivers, promoting informed decision making.