OBJECTIVES: The presence of complications, coexisting comorbidities, and the chronic, incurable nature of diabetes frequently lead to a decline in health-related quality of life (HRQoL). Therefore, this study aimed to gain insight into HRQoL among individuals with diabetes and assess potential enhancement in predictive accuracy by using machine learning (ML) algorithms.

METHODS: Employing a cross-sectional design, this study utilized data from the full-year consolidation of the 2018-2021 Medical Expenditure Panel Survey (MEPS). Individuals aged ≥ 18 years with self-reported diabetes were included. A comprehensive set of 49 factors spanning patient, provider, and health system levels were integrated as potential predictors to forecast HRQoL, measured through the Veteran's RAND 12-item (VR-12) physical component summary scores (PCS) and mental component summary scores (MCS). Individuals without reported VR-12 scores were excluded. Individuals were categorized into groups based on the top 25th percentile for PCS and the top 25th percentile for MCS, separately. The dataset was divided into a 70:30 split for model development and validation. Logistic regression and ML models, including k-nearest neighbors (KNN), support vector machine (SVM), extreme gradient boosting (XGB), and random forest (RF), were compared using receiver operating characteristic (ROC) curve visualization and area under the ROC curve (AUC).

RESULTS: Among 9,443 adults with self-reported diabetes, 53% of them were female. Seventy-one percent of individuals were white followed by 19% black. One-fifth of the individuals were Hispanic. When assessing PCS, RF (AUC=0.79) performed the best, followed by logistic regression (AUC=0.786) and XGB (AUC=0.778). When assessing MCS, XGB (AUC=0.678) performed the best, followed by RF (AUC=0.676) and logistic regression (AUC=0.652).

CONCLUSIONS: The study demonstrated that ML algorithms yielded only slight enhancements in predicting HRQoL in diabetes in MEPS compared to traditional logistic regression, with RF and XGB proving to be the most efficient algorithm to assess PCS and MCS, respectively.