Partial least squares (PLS) model for prediction of definitive and intermediate treatment outcomes in diabetes ketoacidosis (DKA) patients

Abstract

The objective of this paper is to share our experience in creating multivariate predictive models (PLS) for definitive and intermediate treatment outcomes in DKA patients. The research design involved a retrospective cohort of 16 DKA patients who were admitted during a period of 11 months. The clinical dataset was categorized into admission (disturbance variables that cannot be altered), treatment (manipulated variables that can be altered) and treatment outcome variables. PLS models with manipulated variables was found to be clinically preferable than the model with both manipulated and disturbance variables. 10 predictor variables were found to be important in PLS model with manipulated variables, out of which total intravenous insulin before conversion to subcutaneous insulin was found to be the most important predictor. © 2012 IFAC.