Abstract 15947: Machine Learning to Predict Rapid Progression of Carotid Atherosclerosis in Patients With Impaired Glucose Tolerance in the ACT NOW Study

Abstract

Impaired glucose tolerance (IGT) is associated with increased cardiovascular disease. Identifying IGT patients with rapid atherosclerosis progression will help in early risk-stratification. Machine learning (ML) involves algorithms that can learn complex relationships from data to make accurate predictions. Unlike traditional low-dimensional statistical approaches, ML can handle high dimensional datasets to make accurate predictions. Aim: To test the performance of ML in identifying IGT patients who will develop rapid carotid plaque progression. Methods: In the ACT NOW study (IGT patients were randomized to receive either pioglitazone or placebo), 382 subjects had carotid intima-media thickness (IMT) measured at baseline and at 15-18 months. Patients with IMT change in the top 10 percentile (n=39) were classified as rapid progressors (RP) and the rest (n=343) considered non-rapid progressors (NRP). ML using Naïve Bayes model was performed using 115 clinical/laboratory variables to develop a model to fit the true value of the class label (RP, NRP). Experiments were conducted with 10-fold cross validation in training/validation sets. Class prediction was also performed using traditional logistic regression model using 5 variables found different between RP and NRP by univariate analyses. Model performance was assessed using area under receiver operating characteristic curve. Results: Mean±SD IMT change was 0.011±0.026 mm overall; 0.058±0.018 for RP and 0.006±0.02 for NRP. Area under the curve (AUC) for ML (0.855) was higher versus traditional logistic regression model (0.73) (Fig.). ML correctly classified 78% of instances with κ statistic of 0.32. Conclusions: ML using naïve Bayes method provided good predictive ability in identifying IGT patients who had the most rapid plaque progression. ML may improve clinical prediction over traditional low-dimensional biostatistical approaches.