Noninvasive Continuous Glucose Monitoring by Multisensor System: Improved Accuracy Using an Elastic Net Regression

Objective: An approach for noninvasive continuous glucose monitoring was assessed using a combination of the Solianis Multisensor system and a multivariate linear regression model identified with the least absolute shrinkage and selection operator (LASSO). The aim of this work is to improve accuracy exploiting the elastic net (EN) technique for model identification. Method: Data from 45 experimental sessions were considered, during which Multisensor data and reference blood glucose were acquired in parallel. Half of the experiments were used for model identification, and the remaining for model test. The model was identified with EN regression, a technique minimizing a cost function given by the sum of the residual sum of squares plus a regularization term given by the combination of an absolute norm and a squared norm over the coefficients of the linear model. Result: The EN model shows a 9.5% root mean square error reduction (from 63.1 to 57.1 mg/dl) and a 7.6% reduction of the mean absolute relative difference (from 38.1% to 35.2%) with respect to the LASSO. The achieved point accuracy [93% of points in zone A+B of the Clarke error grid (CEG)] is not yet comparable with that of minimally invasive devices, but glucose trends are estimated sufficiently well (around 85% of points in zone AR+BR of the continuous CEG). Conclusion: An improvement in the accuracy of estimated glucose profiles by the Multisensor system is achieved by EN regression even though it is not yet comparable with that of enzyme-based needle sensors. The Multisensor technology is potentially suitable to integrate sparse self-monitoring of blood glucose readings with glucose trend information, resulting in a useful solution as a potential adjunctive device in diabetes monitoring.