A Novel Pediatric Mortality Risk Prediction Score Based on Nonlinear Feature Transformations

The PRISM III score is widely used in pediatric intensive care units (PICU) and has been extensively validated in various settings. While simple enough to allow for fast manual evaluation, PRISM III dichotomizes predictive variables to form prediction scores, which may lead to critical failures. In this paper, we seek to develop a risk prediction score that preserves clinical knowledge embedded in features and structure of PRISM III while addressing limitations caused by variable dichotomization. The novel method transforms predictive variables using nonlinear logistic functions that allow for a fine differentiation between critical and normal values of the variables. Optimal parameters of the logistic functions are inferred for a given patient population via cyclic block coordinate descent, and may be readily re-learned as patient population and standards of care evolve. We tested the proposed technique on brain trauma patients admitted to the PICU of the Dell Children’s Medical Center of Central Texas between 2007 and 2012. The prediction power of the score is evaluated using area under ROC curve (AUC), Youden’s index J, and precision-recall balance in a leave-one-out cross-validation study. The results This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. DGE-1110007. N.M. Arzeno Department of Electrical and Computer Engineering, The University of Texas at Austin, Austin, TX, 78712 USA E-mail: [email protected] K.A. Lawson and Sarah V. Duzinski Trauma Services, Dell Children’s Medical Center of Central Texas 4900 Mueller Blvd Austin, TX 78723 H. Vikalo Department of Electrical and Computer Engineering, The University of Texas at Austin 1 University Station C0803, Austin, TX 78712 Tel.: +1-512-232-7922 E-mail: [email protected] 2 Natalia M. Arzeno et al. demonstrate that the new score significantly outperforms PRISM III in terms of all three criteria.