利用光體積變化描記訊號與深度學習技術於非侵入式雲端血糖連續監測系統

Abstract

本研究利用光體積變化描記（PPG）訊號與深度學習技術，開發非侵入式血糖監測系統。首先使用蒙地卡羅方法（Monte Carlo Method）模擬光子在皮膚內的傳播路徑，以確定光源（LED）與光偵測器（PD）之間的最佳距離。接著，我們開發了血糖量測原型機，利用Arduino結合心率感測晶片(MIKROE 2036)捕捉三種波長(950 nm, 660 nm, 525 nm)的PPG訊號，並進行數據預處理。在深度學習部分，我們使用多層感知器(MLP)、卷積神經網絡（CNN）、殘差神經網絡（ResNet）模型，基於50位受試者提供的250筆三分鐘訊號數據建立血糖預測模型。此外，該系統設計了專為雲端運算而優化的軟體架構，使血糖監控不僅限於單一設備，能夠在任何支援的設備上實時且連續的監控血糖水平。這種便捷和安全的血糖管理方案為糖尿病患者提供了更加有效的健康管理工具。

This study develops a non-invasive blood glucose monitoring system using photoplethysmography (PPG) signals and deep learning technology. Initially, the Monte Carlo Method was used to simulate the photon propagation paths in the skin to determine the optimal distance between the light source (LED) and the photodetector (PD). Subsequently, we developed a blood glucose measurement prototype, utilizing Arduino combined with the heart rate sensing chip (MIKROE 2036) to capture PPG signals at three wavelengths (950 nm, 660 nm, 525 nm) and perform data preprocessing. In the deep learning phase, we employed Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Residual Neural Network (ResNet) models to build blood glucose prediction models based on 250 three-minute signal data samples from 50 subjects. Additionally, the system was designed with a cloud computing-optimized software architecture, enabling real-time and continuous blood glucose monitoring on any supported device, beyond a single device. This convenient and secure blood glucose management solution provides diabetes patients with a more effective health management tool.