Machine Learning for Outcome Prediction in EEG-Monitored Children in the Intensive Care Unit
Abstract number :
1.241
Submission category :
4. Clinical Epilepsy / 4D. Prognosis
Year :
2018
Submission ID :
496267
Source :
www.aesnet.org
Presentation date :
12/1/2018 6:00:00 PM
Published date :
Nov 5, 2018, 18:00 PM
Authors :
Iván Sánchez Fernández, Boston Children’s Hospital, Harvard Medical School; Hospital Sant Joan de Déu, Universidad de Barcelona; Arnold Sansevere, Boston Children's Hospital, Harvard Medical School; Marina Gaínza-Lein, Bost
Rationale: Multiple studies evaluated predictors of poor outcome in critically ill children who undergo continuous electroencephalogram monitoring (cEEG) in the intensive care unit (ICU). These studies take an explanatory approach including building regression models based on prior knowledge and plausible biological etiologies. In contrast, little attention has been paid to purely predictive models. The aim of this study was to evaluate the performance of models predicting in-hospital mortality in critically ill children undergoing continuous electroencephalogram (cEEG) in the intensive care unit (ICU). Methods: We evaluated the performance of machine learning algorithms for predicting mortality in a database of critically ill children undergoing cEEG in the ICU. We randomly split the original dataset into 70% of patients for a training subset and 30% for a testing or validation subset. Results: Four-hundred-and-fourteen patients (54% male) were analyzed. Their median (p25-p75) age was 4.2 (0.8-11.3) years. The etiology was structural symptomatic in 52%. Electrographic seizures occurred in 25% of patients. Fifty-nine (14%) patients died in the hospital. All models automatically generated by machine learning techniques added information to the explanatory models based on prior medical knowledge (Table 1, Figure 1). Differences compared to the explanatory models reached statistical significance for the 3 best predictive models: forward stepwise selection, backward stepwise elimination, and backward and forward elimination and selection, which had the same predictive performance. There was a tendency towards add on performance for support vector machine with linear kernel and for LASSO regression (Figure 1, Table 2). The performance of random forest, and for support vector machine with polynomial kernel and with radial kernel did not add much to the explanatory model (Figure 1, Table 2). The variables considered in each model are presented in Supplementary Table 1. Conclusions: Using few variables and a relatively small number of patients, machine learning techniques added information to explanatory models for prediction of in-hospital mortality. Funding: Epilepsy Research Fund.