Machine Learning-Based Indices Assessing Different Aspects of Beta-Cell Function in Pregnancy

Accurate assessment of pancreatic beta-cell function parameters, such as glucose sensitivity (G-Sens), rate sensitivity (R-Sens), and potentiation factor ratio (PFR), relies on mathematical modelling coupled with C-peptide measurement, both not always accessible in clinical settings. Machine learning may provide surrogate markers of model-and-C-peptide-based parameters. Aim of the study was to leverage machine learning to build predictive equations of G-Sens, R-Sens, and PFR in pregnant women, without the need of modeling and C-peptide. To this aim, predictive approaches were implemented (multivariate polynomial regressions), under different scenarios of data availability. We found that G-Sens prediction showed good performance (R-adj(2) = 0.45, p<0.0001 in test set), whereas results were unsatisfactory for R-Sens. PFR prediction showed moderate performance (R-adj(2) = 0.33, p < 0.01 in test set). In conclusion, machine learning is appropriate for G-Sens and PFR prediction, while R-Sens prediction appears hardly feasible.