Advancements in Non-invasive AI-Powered Glucose Monitoring: Leveraging Multispectral Imaging Across Diverse Wavelengths

The pursuit of non-invasive glucose monitoring has leveraged multispectral imaging technology. Our study focuses on improving GlucoCheck, a non-invasive AI-powered glucose monitor. Through processing a dataset of 3600 images, we derived four subsets for unique wavelengths and trained regression models. Performance evaluations, employing Mean Absolute Error (MAE), Clarke Error Grids (CEG), and Bland-Altman Plots (BAP), revealed promising outcomes, with a median MAE of 2.06 mg/dl, CEG Zone A% of 99.17%, and percentage of BAP outliers as 3.66%. While no clear correlation was found between wavelength and statistical accuracy, a significant relationship emerged between wavelength and clinical accuracy through agreement with BAP outliers. These nuanced findings highlight the potential of multispectral imaging and machine learning in advancing accurate glucose estimation.