Optimising test intervals for individuals with type 2 diabetes: A machine learning approach

Background: Chronic disease monitoring programs often adopt a one-size-fits-all approach that does not consider variation in need, potentially leading to excessive or insufficient support for patients at different risk levels. Machine learning (ML) developments offer new opportunities for personalised medicine in clinical practice. Objective: To demonstrate the potential of ML to guide resource allocation and tailored disease management, this study aims to predict the optimal testing interval for monitoring blood glucose (HbA1c) for patients with Type 2 Diabetes (T2D). We examine fairness across income and education levels and evaluate the risk of false-positives and false-negatives. Data: Danish administrative registers are linked with national clinical databases. Our population consists of all T2D patients from 2015-2018, a sample of more than 57,000. Data contains patient-level clinical measures, healthcare utilisation, medicine, and socio-demographics. Methods: We classify HbA1c test intervals into four categories (3, 6, 9, and 12 months) using three classification algorithms: logistic regression, random forest, and extreme gradient boosting (XGBoost). Feature importance is assessed with SHAP model explanations on the best-performing model, which was XGBoost. A training set comprising 80% of the data is used to predict optimal test intervals, with 20% reserved for testing. Cross-validation is employed to enhance the model’s reliability and reduce overfitting. Model performance is evaluated using ROC-AUC, and optimal intervals are determined based on a “time-to-next-positive-test” concept, with different durations associated with specific intervals. Results: The model exhibits varying predictive accuracy, with AUC scores ranging from 0.53 to 0.89 across different test intervals. We find significant potential to free resources by prolonging the test interval for well-controlled patients. The fairness metric suggests models perform well in terms of equality. There is a sizeable risk of false negatives (predicting longer intervals than optimal), which requires attention. Conclusions: We demonstrate the potential to use ML in personalised diabetes management by assisting physicians in categorising patients by testing frequencies. Clinical validation on diverse patient populations is needed to assess the model’s performance in real-world settings.