Feature Extraction for Estimating Acute Blood Glucose Level Variation From Multiwavelength Facial Images

Researchers have explored optical methods for measuring blood glucose levels in a minimally invasive manner; however, the stress and behavioral limitations of wearing a device have been challenging. To address this, a study was conducted on multiwavelength facial imaging for remotely measuring vital signs related to hemodynamics. The development of technologies for remote blood glucose estimation using multiwavelength facial imaging should lead to early detection and prevention of diabetes and improve the quality of life of diabetics. We extracted spatial features related to acute blood glucose variation from visible, near-infrared, and infrared facial images and evaluated their accuracy in estimating blood glucose levels for remote measurements. The infrared images captured features on the nose and the visible images showed features under the eyes and in the contour area. However, the estimation accuracy was low because these features might have been related to physiological states, light, and shadows rather than blood glucose levels. The near-infrared (NIR) images captured facial features that might have been related to blood glucose variation in facial capillaries, and the average root-mean-square error (RMSE) of blood glucose estimation was 15.44 mg/dL for the two subjects tested. Furthermore, we focused on the near-infrared band, referred to as the “living-body window” because of its high permeability to the living body and conducted a study using facial images acquired in the wavelength band of 1050–1650 nm (NIR-II) in addition to the 700–1000-nm band (NIR-I) used in the aforementioned experiment. The NIR-I and NIR-II images exhibited features related to blood glucose levels in blood vessels at different depths, including the entire face and orbital and lateral nasal areas through which the major arteries pass. The accuracy of blood glucose estimation was greater with NIR-I, with an average RMSE of 7.25 mg/dL for 12 subjects. It is suggested that increasing the number of data to improve the generalization performance of the model and searching for more specific wavelength bands suitable for blood glucose level monitoring will improve the accuracy of this method.