Smart Insulin Therapy Initiation In Insulin-Naive Subjects With Type 2 Diabetes Using Machine Learning

Objective: Insulin-naive subjects with type 2 diabetes (T2D) are usually initiated to basal insulin therapy, starting from a low initial insulin dose (IID) that is progressively adjusted to reach the optimal insulin dose (OID) and glucose target (OTG). This procedure is time consuming, especially for very insulin resistant subjects. To speed up the achievement of the OID, it would be useful to guess, before insulin titration, if subject-specific OID will be high or low to roughly estimate a smarter IID (e.g. IID=10U for low, 44U for high). Methods: We used 300 in-silico insulin-naive subjects of the T2D Padova simulator, who underwent a 52-week basal insulin titration trial, and for which demographics, fasting and post-prandial glucose, insulin and C-peptide concentrations were available. We first classified each subject’s OID as high or low. Then, we tested three bootstrapping LASSO logistic regression models to predict OID class using body weight, sex, age, fasting and postprandial glucose, insulin and C-peptide (Model 1); same covariates of Model1 apart from postprandial concentrations (Model 2); same covariates of Model 2 apart from hormones measurements (Model3). Models were trained and tested in 70%-30% ratio. Model performance was assessed based on the area under the Receiver Operating Characteristic curve (AUC-ROC). Results: Model2 well predicted OID, with performance comparable to Model1 (AUC-ROC=92.74% and 94.22%, respectively),despite Model1 would require postprandial measurements of glucose, insulin and C-peptide. Performance of Model3, which would have the advantage of not requiring any hormone measurements, is a bit worse but still acceptable (AUC-ROC=88.4%). Conclusion: Depending on the availability of fasting and/or post-prandial plasma glucose and hormone concentrations, one can use one of the three models to guess the IID to rapidly reach the OGT.