Comparison of machine learning models for seizure prediction in hospitalized patients

• Authors: Aaron F. Struck, University of Wisconsin; Andres Rodriguez Ruiz, Emory University; Gamaledin Osman, Henry Ford Hospital; Emily J. Gilmore, Yale University; Hiba Haider, Emory University; Monica Dhakar, Emory University; Matthew Schrettner, University of South Carolina Greenville; Jong W. Lee, Harvard Medical School; Nicolas Gaspard, Yale University; Lawrence J. Hirsch, Yale University; M. Brandon Westover, Harvard Medical School
• Journal Title: Annals of Clinical and Translational Neurology
• Volume: Volume 6, Number 7
• Publisher: Wiley Open Access: Creative Commons Attribution Non-Commercial No Derivatives | 2019-07-01, Pages 1239-1247
• Type of Work: Article | Final Publisher PDF
• Publisher DOI: http://doi.org/10.1002/acn3.50817
• Subjects: Biology, Neuroscience
• Keywords: Science & Technology; Life Sciences & Biomedicine; Clinical Neurology; Neurosciences; Neurosciences & Neurology; CONTINUOUS EEG; SUBARACHNOID HEMORRHAGE; ELECTROGRAPHIC SEIZURES; NONCONVULSIVE SEIZURES; PERIODIC DISCHARGES; PHENYTOIN; ELECTROENCEPHALOGRAPHY; PROBABILITY; FAILURE; SCORE

Abstract

Objective: To compare machine learning methods for predicting inpatient seizures risk and determine the feasibility of 1-h screening EEG to identify low-risk patients (<5% seizures risk in 48 h). Methods: The Critical Care EEG Monitoring Research Consortium (CCEMRC) multicenter database contains 7716 continuous EEGs (cEEG). Neural networks (NN), elastic net logistic regression (EN), and sparse linear integer model (RiskSLIM) were trained to predict seizures. RiskSLIM was used previously to generate 2HELPS2B model of seizure predictions. Data were divided into training (60% for model fitting) and evaluation (40% for model evaluation) cohorts. Performance was measured using area under the receiver operating curve (AUC), mean risk calibration (CAL), and negative predictive value (NPV). A secondary analysis was performed using Monte Carlo simulation (MCS) to normalize all EEG recordings to 48 h and use only the first hour of EEG as a “screening EEG” to generate predictions. Results: RiskSLIM recreated the 2HELPS2B model. All models had comparable AUC: evaluation cohort (NN: 0.85, EN: 0.84, 2HELPS2B: 0.83) and MCS (NN: 0.82, EN; 0.82, 2HELPS2B: 0.81) and NPV (absence of seizures in the group that the models predicted to be low risk): evaluation cohort (NN: 97%, EN: 97%, 2HELPS2B: 97%) and MCS (NN: 97%, EN: 99%, 2HELPS2B: 97%). 2HELPS2B model was able to identify the largest proportion of low-risk patients. Interpretation: For seizure risk stratification of hospitalized patients, the RiskSLIM generated 2HELPS2B model compares favorably to the complex NN and EN generated models. 2HELPS2B is able to accurately and quickly identify low-risk patients with only a 1-h screening EEG.