Identifying Prediabetes in Canadian Populations Using Machine Learning.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Public Health

Background:

• Prediabetes signifies elevated blood glucose levels, preceding Type 2 diabetes (T2D).
• Early identification of prediabetes is crucial for preventing T2D progression in at-risk populations.
• Machine learning (ML) offers potential for improving prediabetes prediction accuracy.

Purpose of the Study:

• To identify the most effective ML model for prediabetes prediction.
• To determine key biological variables for distinguishing prediabetic individuals.
• To leverage ML for enhanced early detection of prediabetes.

Main Methods:

• Analysis of 6,414 participants from the Canadian Primary Care Sentinel Surveillance Network (CPCSSN).
• Selection of ten key variables based on literature, data completeness, and collinearity.
• Comparative evaluation of seven ML models, including a Deep Neural Network (DNN) with early stop regularization.

Main Results:

• The Deep Neural Network (DNN) model achieved the highest recall rate of 60% for prediabetes prediction.
• The DNN model demonstrated superior performance compared to six other ML models evaluated.
• Key biological variables were identified as critical for distinguishing prediabetic individuals.

Conclusions:

• The DNN model represents a significant advancement in the early detection of prediabetes.
• Integrating ML models like DNN can enhance preventive healthcare strategies against T2D.
• Further research is warranted to refine prediction accuracy by incorporating novel biomarkers or alternative ML techniques.