Non-invasive blood glucose sensing by machine learning of optic fiber-based speckle pattern variation

Abstract

. Significance: The ability to perform frequent non-invasive monitoring of glucose in the bloodstream is very applicable for diabetic patients. Aim: We experimentally verified a non-invasive multimode fiber-based technique for sensing glucose concentration in the bloodstream by extracting and analyzing the collected speckle patterns. Approach: The proposed sensor consists of a laser source, digital camera, computer, multimode fiber, and alternating current (AC) generated magnetic field source. The experiments were performed using a covered (with cladding and jacket) and uncovered (without cladding and jacket) multimode fiber touching the skin under a magnetic field and without it. The subject’s finger was placed on a fiber to detect the glucose concentration. The method tracks variations in the speckle patterns due to light interaction with the bloodstream affected by blood glucose. Results: The uncovered fiber placed above the finger under the AC magnetic field (150 G) at 140 Hz was found to have a lock-in amplification role, improving the glucose detection precision. The application of the machine learning algorithms in preprocessed speckle pattern data increase glucose measurement accuracy. Classification of the speckle patterns for uncovered fiber under the AC magnetic field allowed for detection of the blood glucose with high accuracy for all tested subjects compared with other tested configurations. Conclusions: The proposed technique was theoretically analyzed and experimentally validated in this work. The results were verified by the traditional finger-prick method, which was also used for classification as a conventional reference marker of blood glucose levels. The main goal of the proposed technique was to develop a non-invasive, low-cost blood glucose sensor for easy use by humans.

Noninvasive blood glucose sensing by secondary speckle pattern artificial intelligence analyses

“…The preprocessed series of image input data for the DNN is normalized to make it optimal for a convolution-based network. The ML data are modified in time series form, which is further modified for the case of applied AC magnetic fields to infer data at an applied frequency 17 . Each person was tested several times during each session to increase the variability in data.…”

“…In that situation, a speckle pattern will not typically be seen because the speckle patterns created by different wavelengths normally average each other out, which his known as self-interference. 16,17 In our proposed sensing mechanism, the secondary speckle patterns reflected from a human skin illuminated by coherent light are captured using a defocused camera for temporal analysis. Instead of changing randomly while the camera is defocused, a speckle pattern will move or vibrate in the transversal plane.…”

“…Each speckle image is correlated to the next image, and for each image, the correlation between the current and reference frames is calculated and averaged over time. 17 The temporal variation of the graph peaks correlation value represents the changes in glucose levels. The data under normal conditions are directly used to train the algorithm.…”

“…Speckles can be a point of reference to monitor changes in the scattered light’s phase where the low-coherence light with different wavelengths is considered for analysis. In that situation, a speckle pattern will not typically be seen because the speckle patterns created by different wavelengths normally average each other out, which his known as self-interference 16 , 17 …”

“…The multimode fiber (covered and uncovered) touching the subject’s finger under a magnetic field is applied to detect the glucose concentration. The uncovered fiber placed below the finger under an alternating current (AC) magnetic field (150 Gauss) at 140 Hz was found to have a lock-in amplification role, improving glucose sensing 17 . The current work further develops the speckle-based noninvasive blood glucose sensing combined with data processing by machine learning (ML) and deep neural network (DNN) analysis.…”