Résumé
Background and aims: Detection of immunological risk (IR) for type 1 diabetes (T1D) requires testing for presence of islet autoantibodies (AB) . However, who and when to test is challenging. Our goal is to develop a (pre) screening technology to estimate the T1D IR from a simple continuous glucose monitoring (CGM) home test. Methods: Data from a NIH TrialNet ancillary study with 53 healthy relatives to T1D with (N=35) or without (N=18) AB, mean±SD age of 25.4±10.7, HbA1c of 5.1±0.3%, and BMI of 22.9±5.2 (kg/m2) were used. Subjects wear CGM for a week and consumed 3 caloric drinks (Boost) . Nine glycemia features were extracted from the 75 minutes post-Boost CGM traces and used for IR classification. K-fold cross-validation with eight different classification models (i.e., linear, nonlinear, advanced, etc.) was used to develop an IR classifier. Receiver operating characteristic (ROC) curves, sensitivity, and specificity were used to select the best performing classification model. Results: IR classifier based on a neural network (NN) model with upsampling and 10-fold cross-validation achieved the best performance between different classification models: an AUC-ROC of 0.93, a sensitivity of 89%, and a specificity of 75%. A random forest model-based classifier performed almost equally well. The Naive Bayes model achieved the best sensitivity of 100%, but had an AUC-ROC of 0.78, and specificity of 32%. On the other hand, the support vector machines models achieved the worst performance with the best being AUC-ROC of 0.63, a sensitivity of 65%, and a specificity of 54%. Conclusion: We have developed a new methodology, which combines a dedicated one-week CGM home test with a neural network-based classifier that reliably predicts the individual IR in terms of presence or absence of islet AB. The test is self-administered, does not require a visit to a hospital or a lab, and can be used as an alternative or in addition (e.g., for prescreening) to the standard test for presence of islet autoantibodies.