Objective: Hyperglycemia in neonatal intensive care units is associated with mortality and morbidity. This trial aims to use machine learning methods to provide a prediction of outcomes in hyperglycemic neonates, based on model-based metabolic (glycemic control) data as a non-invasive marker. Method: Glycemic control data from 44 patients (4499 hours) under the STAR-NICU or STAR-GRYPHON model-based glycemic controllers from Christchurch Women’s Hospital were used. Predictive models were built using attributes from hourly, patient-specific, model-based insulin sensitivity. Among these patients, 12 contracted sepsis, 8 suffered from intraventricular hemorrhage (IVH), and 8 died. The methods used were classification trees and K-nearest neighbors. The efficacy of the models was assessed evaluating sensitivity, specificity and accuracy. Result: Mean insulin sensitivity was different among different sub-groups: 7.51×10-4, 5.47×10-4, 2.42×10-4, and 7.50×10-4L/mU/min for patients who were septic, had IVH>grade 1, non-survivors, and survivors respectively. Variability assessed as IQR range was also different between groups, with 1.00×10-4, 4.99×10-5, 4.22×10-5, and 9.09×10-5L/mU/min respectively. It was possible to predict mortality with 85% sensitivity after the first 15 hours, and (later proven) sepsis with sensitivity of 80% within 20 hours. Conclusion: A clinically validated model-based insulin sensitivity measure and its variability, may provide information about patient condition and possible outcome, despite modeling limitations. This study emphasized the potential of machine learning to provide information on degrading patient condition and worsened outcome, as an alert to provide more intensive care.