非侵襲血糖値モニタリングのための近赤外イメージングと畳み込みニューラルネットワーク

抄録

Monitoring blood glucose levels is essential for patients with diabetes. In the previous study, blood glucose measurement has developed using near-infrared (NIR) imaging, we have found that the accuracy obtained by machine learning based model was reduced by adjustable positioning of the experimental setup and preprocessing techniques of the image. Therefore, in this study, we conducted in-vivo measurements by capturing images of light transmittance from a tunable pulse laser passing through the purlicue (the area between the index finger and thumb). For stable image capturing, an adjustable setup was designed to accommodate various hand sizes, 18 subjects underwent an Oral Glucose Tolerance Test (OGTT), providing input for predicting blood glucose levels. A Convolutional Neural Network (CNN) machine learning model was employed to predict blood glucose levels from absorbance difference images. The performance results to RMSE of 19.28 mg/dl and 0.50R². The results demonstrate the potential of using NIR imaging and image standardization preprocessing for non-invasive glucose monitoring, representing a significant leap forward in diabetes care management.