Improving non-invasive glucose estimation with monthly calibrated photoplethysmography and implicit HbA1c

Most noninvasive blood glucose technologies, especially wearable photoplethysmography devices, require multiple calibrations and are often limited to narrow cohorts such as unmedicated or mild cases. We assess whether a single pretest once per month can meet clinical accuracy while broadening applicability through cohort-specific models.

We develop models for three groups: (i) individuals not using antidiabetic drugs, (ii) those using oral antidiabetic drugs only, and (iii) those using antidiabetic drugs in combination with other medications. Models are trained on cohort data with and without the monthly pretest and are then applied directly to personal testing without retraining. Inputs include dual-channel photoplethysmography signals and an inferred HbA1c (glycated hemoglobin) feature. Accuracy is summarized by mean absolute relative difference, clinical safety by the Parkes Error Grid, and improvements by a nonparametric rank-sum test.

Here, we show that the best models using a single monthly pretest achieve mean absolute relative differences of 9.59, 12.23, and 16.40% for groups (i), (ii), and (iii), respectively. In the most complex group (iii), prediction errors are significantly lower than our earlier work according to the rank-sum test. The single-pretest models produce no clinically unacceptable readings on the Parkes Error Grid, likely due to dual-channel input and the inferred glycated-hemoglobin feature.

A single monthly pretest enables accurate and clinically safe noninvasive glucose measurement across diverse patient groups. The approach operates without retraining or fine-tuning and can adapt to new users and devices through edge computing, supporting integration into current wearables for everyday diabetes management and public-health prevention.