Postoperative neonatal mortality prediction using superlearning

Abstract

Listen

Translate article

BackgroundThe variable risks associated with neonatal surgery present a challenge to accurate mortality prediction. We aimed to apply superlearning, an ensemble machine learning method, to the prediction of 30-day neonatal postoperative mortality. Materials and methodsWe included neonates in the 2012-2014 National Surgical Quality Improvement Program Pediatric. Patients treated in 2012-13 were used in model development (n = 6499), and patients treated in 2014 formed the validation sample (n = 3552). Our superlearner algorithm included 14 regression and machine learning algorithms and included all preoperative patient demographic and clinical characteristics, including indicator variables for surgical procedures. Performance was evaluated using mean squared error and measures of discrimination and calibration. ResultsThe superlearner out-performed all individual algorithms with regard to cross-validated mean squared error. It showed excellent discrimination, with an area under the receiver-operating characteristic curve of 0.91 in development and 0.87 in validation. The superlearner showed good calibration in development but not in validation (Cox calibration test P = 0.06 and P < 0.001, respectively). Performance was improved when the superlearner was fit using only variables strongly associated with mortality in bivariate analysis (area under the receiver-operating characteristic curve 0.89, calibration test P = 0.63 in validation). ConclusionsSuperlearning provided improved or equivalent performance compared with individual regression and machine learning algorithms for predicting neonatal surgical mortality. This method should be considered for prediction in large data sets whenever complex mechanisms make parametric modeling assumptions unrealistic.