Machine Learning for Early Detection of Nonresponders to Lifestyle Intervention for Prediabetes

Introduction: Lifestyle intervention is the first-line treatment for prediabetes. More intensive intervention usually follows for those patients that fail to improve glycemia. We explored the possibility to early detect non-responders to lifestyle intervention from a single baseline evaluation. Methods: The sample was composed of 738 pre-diabetic (fasting glucose 100-125 mg/dl) patients (median age 54 IQR 45-62, 57% female, median BMI 30.4 IQR 27.5, 34.2) , that were prescribed a lifestyle intervention, and were re-evaluated within 12 months. Lifestyle intervention consisted of a hypocaloric diet (based on resting energy expenditure measured by indirect calorimetry, with a Mediterranean pattern, tailored on patient habits) and WHO physical activity guidelines. The outcome was normalization of fasting glucose at follow-up (0: ‡100 mg/dl, 1: <100 mg/dl) , with an historical failure rate of 44%. Baseline evaluation consisted of anthropometry, indirect calorimetry, abdominal fat ultrasound, laboratory exams, structured medical and nutritional interview, and psychological questionnaires, with a total of 114 predictors. Statistical and machine learning models were screened to select the best classifier for the outcome. Results: The best model was a random forest model (a machine learning model) , with a validation-set accuracy of 72%, ROC AOC of 0.64, sensitivity of 55%, and specificity of 86%. Discussion: The best developed model is able to identify more than half non-responders on the basis of a single baseline evaluation, while including few responders as false-positives. The model accuracy resulted higher than the historical rate of failure, so integrating the model suggestion in the clinical decision process should allow to more promptly treat future patients and improve the overall success rate. Machine learning models show potential in leveraging high-dimensional multi-domain routinely-collected clinical data to aid in clinical decision.