Machine Learning Classification for Blood Glucose Performances Using Insulin Sensitivity and Respiratory Scores in Diabetic ICU Patients

Abstract

Diabetes Mellitus (DM) patients with acute respiratory failure in the Intensive Care Unit (ICU) are susceptible to hyperglycaemia with adverse outcome of mortality. Clinically, Partial Pressure of Oxygen over a Fraction of Inspired Oxygen (P/F) scores is use as an indicator for acute respiratory failure and studies have shown that Insulin Sensitivity (S_I) can be used as the glycaemic control biomarker for DM patients. Since the elevation of blood glucose in ICU patients is linked to the progression of the acute respiratory system, this preliminary study initiates the combination of S_I, P/F, and DM status as the main predictors for machine learning classification. This assessment was done to identify which classification models and predictors between insulin sensitivity (S_I), (P/F) scores, and diabetic status will give higher accuracy on Blood Glucose (BG) performance with 7 types of classifier models. In total, 5684 total inputs from 3 predictors extracted from 76 ICU patients were split into 80:20 ratio for training and test sets with five-fold cross-validations. BG performances using three predictors from training vs. test data show that the k-Nearest Neighbor and Neural Network classifiers showed that the highest accuracies achieved were 54.1% and 54.5%, respectively. The sensitivity and specificity evaluated for both model’s robustness demonstrated the possibility of using k-Nearest Neighbor and Neural Network for future BG performance prediction. Based on the model’s robustness increment result, 8% vs. 12% and 10% vs. 4% shows a possibility that S_I, P/F scores, and DM can be utilized together as an input to classify glycemic level using both classifier models with a larger dataset from respiratory failures patients.