Extended Data Fig. 10: Optimizing the time delay in the PLS model for glucose prediction.
From: Subcutaneous depth-selective spectral imaging with mμSORS enables noninvasive glucose monitoring
a, The VPG data of a typical subject with diabetes during the 5-h OGTT in the preliminary BESH. Measured VPGs (dots) were fitted with polynomials and the backtracked VPG with a certain time lag (stars) were used as reference in PLS models. b, RMSE between the model predictions and the reference VPGs, varying with the time lag from −25 to 0 min. The optimized time lag was at −16 min. The models were trained and tested with spectra from offset 3 in the preliminary BESH. Model predictions were generated using leave-one-subject-out cross-validation scheme (Fig. 2h). c, Counterpart of b for offset 2 and 3 as well as the concatenation (offset 2–3) of these two offsets in the expanded BESHs of 230 subjects. The optimal time lag was at −13 min for both two offsets and the concatenation. Model predictions were generated using subject-wise tenfold cross-validation scheme (Fig. 3d). d, Counterpart of the black curve in c in the training set comprised of 200 subjects in the expanded BESHs (Fig. 4a). The optimal time lag was at −13 min.
