Development of non-invasive diabetes risk prediction models as decision support tools designed for application in the dental clinical environment

The objective was to develop a predictive model using medical-dental data from an integrated electronic health record (iEHR) to identify individuals with undiagnosed diabetes mellitus (DM) in dental settings. Retrospective data retrieved from Marshfield Clinic Health System’s data-warehouse was pre-processed prior to conducting analysis. A subset was extracted from the preprocessed dataset for external evaluation (N(validation)) of derived predictive models. Further, subsets of 30%–70%, 40%–60% and 50%–50% case-to-control ratios were created for training/testing. Feature selection was performed on all datasets. Four machine learning (ML) classifiers were evaluated: logistic regression (LR), multilayer perceptron (MLP), support vector machines (SVM) and random forests (RF). Model performance was evaluated on N(validation). We retrieved a total of 5319 cases and 36,224 controls. From the initial 116 medical and dental features, 107 were used after performing feature selection. RF applied to the 50%–50% case-control ratio outperformed other predictive models over N(validation) achieving a total accuracy (94.14%), sensitivity (0.941), specificity (0.943), F-measure (0.941), Mathews-correlation-coefficient (0.885) and area under the receiver operating curve (0.972). Future directions include incorporation of this predictive model into iEHR as a clinical decision support tool to screen and detect patients at risk for DM triggering follow-ups and referrals for integrated care delivery between dentists and physicians.