Comparing the predictive performance of diabetes complications using administrative health data and clinical data

When predicting adverse complications due to Type 2 Diabetes, often two different approaches are taken: predictions based on clinical data or those using administrative health data. No studies have assessed whether these two approaches reach comparable predictions. This study compares the predictive performance of these two data sources and examines the algorithmic fairness of the developed models. We developed XGBoost models to predict the two-year risk of nephropathy, tissue infection, and cardiovascular events in Type 2 Diabetes patients. The models using only clinical data achieved an average AUC of 0.78, while the models using administrative health data alone achieved 0.77. A hybrid model combining both data types resulted in an average AUC of 0.80, across complications. The models showed that laboratory data were key for predicting nephropathy, whereas comorbidity and diabetes age were most important for tissue infection. For cardiovascular events, age and a history of congestive heart failure was the most important predictors. Our analysis identified bias on the feature sex in all three outcomes: models tended to underestimate risk for females and overestimate it for males, indicating a need to address fairness in these applications. This study demonstrates the effectiveness of ML models using both data types for predicting diabetes complications. However, the presence of sex bias highlights the importance of improving model fairness for reliable clinical use.