Machine Learning Methods for Continuous Glucose Monitoring Data and their Potential in Cardiovascular Risk Assessment

The thesis asks whether continuous glucose monitoring can help gauge cardiovascular risk in diabetes, beyond HbA1c, which reflects average glucose over three months. Diabetes involves high blood glucose and higher cardiovascular risk, and risk models might improve if they used richer glucose patterns. CGM records daily glucose fluctuations; machine learning can uncover patterns in such data.

Three studies make up the dissertation. One reviews existing evidence linking CGM-derived measures to cardiovascular disease. Another tests whether a foundation model trained on large CGM datasets can capture clinically useful patterns and identify individuals who might need further assessment for heart failure. The third asks whether the racial makeup of training data influenced machine-learning performance when estimating later blood sugar values among Non-Hispanic Black and White patients with type 1 diabetes.