Identifying Prediabetes in Canadian Populations Using Machine Learning.

Abstract

Prediabetes is a critical health condition characterized by elevated blood glucose levels that fall below the threshold for Type 2 diabetes (T2D) diagnosis. Accurate identification of prediabetes is essential to forestall the progression to T2D among at-risk individuals. This study aims to pinpoint the most effective machine learning (ML) model for prediabetes prediction and to elucidate the key biological variables critical for distinguishing individuals with prediabetes. Utilizing data from the Canadian Primary Care Sentinel Surveillance Network (CPCSSN), our analysis included 6,414 participants identified as either nondiabetic or prediabetic. A rigorous selection process led to the identification of ten variables for the study, informed by literature review, data completeness, and the evaluation of collinearity. Our comparative analysis of seven ML models revealed that the Deep Neural Network (DNN), enhanced with early stop regularization, outshined others by achieving a recall rate of 60%. This model's performance underscores its potential in effectively identifying prediabetic individuals, showcasing the strategic integration of ML in healthcare. While the model reflects a significant advancement in prediabetes prediction, it also opens avenues for further research to refine prediction accuracy, possibly by integrating novel biological markers or exploring alternative modeling techniques. The results of our work represent a pivotal step forward in the early detection of prediabetes, contributing significantly to preventive healthcare measures and the broader fight against the global epidemic of Type 2 diabetes.