A novel approach for generating physiological interpretations through machine learning

Abstract

Predicting blood glucose trends and implementing suitable interventions are crucial for managing diabetes. Modern sensor technologies enable the collection of continuous glucose monitoring (CGM) data along with diet and activity records. However, machine learning (ML) techniques are often used for glucose level predictions without explicit physiological interpretation. This study introduces a method to extract physiological insights from ML-based glucose forecasts using constrained programming. A feed-forward neural network (FFNN) is trained for glucose prediction using CGM data, diet, and activity logs. Additionally, a physiological model of glucose dynamics is optimized in tandem with FFNN forecasts using sequential quadratic programming and individualized constraints. Comparisons between the constrained response and ML predictions show higher root mean square error (RMSE) in certain intervals for the constrained approach. Nevertheless, Clarke error grid (CEG) analysis indicates acceptable accuracy for the constrained method. This combined approach merges the generalization capabilities of ML with physiological insights through constrained optimization.