結合機器學習與多光源光體積血容積之非侵入式血糖偵測;Non-Invasive Blood Glucose Detection by means of Multi-Source PPG and Machine Learning

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Electrical Engineering > Electronic Thesis & Dissertation > Item 987654321/86706

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/86706

Title:	結合機器學習與多光源光體積血容積之非侵入式血糖偵測;Non-Invasive Blood Glucose Detection by means of Multi-Source PPG and Machine Learning
Authors:	高潔聲;Kao, Chieh-Sheng
Contributors:	電機工程學系
Keywords:	非侵入式;血糖偵測;多光源;光體積血容積;Non-invasive;glucose;PPG
Date:	2021-08-16
Issue Date:	2021-12-07 13:08:35 (UTC+8)
Publisher:	國立中央大學
Abstract:	近年來，為了使糖尿病患者免於日常扎針之苦，研究人員致力於發展光學式血糖量測方法。本研究結合多光源光體積血容積(Photoplethysmography，PPG)訊號與機器學習進行非侵入式的血糖預測。我們使用血紅素較易吸收的綠光(波長530nm)與葡萄糖較易吸收的紅外光(波長1550nm)自製雙光源H-bridge電路進行受測者指尖光體積血容積PPG訊號量測。PPG訊號由STM32F429 MCU搭配AFE4490類比前端晶片，透過藍芽回傳至電腦，再由python解碼。同時，本研究也透過ADS1299晶片收取受測者心電圖(Electrocardiography，ECG)、Edan iM50 patient monitor收取受測者血壓等生理訊號。 18位受測者在餐前以及餐後三十分鐘分別進行上述實驗，也使用市售OneTouch Ultra- Plus-Flex血糖機量測血糖，以便和預測的血糖值做誤差計算。PPG、ECG、血壓等生理訊號將提取出的19項特徵，經由Beer-Lambert law轉換為相互線性關係後，由partial F-test進行特徵選用。接著18位受測者輪流做為測試資料輸入多元線性回歸模型進行18 fold cross validation。本研究使用臨床上認可的Clarke Error Grid以及平均絕對偏差(mean absolute relative difference, mARD)當作血糖預測準確度的標準，獲得平均0.62的R2以及9.45%誤差。 ;Traditional finger-pick method for blood glucose monitoring requires invasively inserting a needle under the skin to get a blood sample. In order to relieve the pain of daily finger-pricking of diabetes patients, reserachers have dedicated themselves in developing optical glucose meters over the past decades. In our study, we combine the Photoplethysmography(PPG) signals of multi-wavelength LEDs and machine learning to develop a non-invasive blood glucose monitoring and prediction system. Two LED light sources with wavelengths of 530 nm and 1550 nm , which are sensititive to hemoglobin and glucose, respectively, were conducted by an H-bridge circuit to measure the fingertip PPG signal of subjects. PPG signal is captured with an STM32F429 MCU and AFE4490 analog front end, then transferred back to a computer via Bluetooth and decoded with Python. Other physiological signals such as Electrocardiography(ECG) with ADS1299 ADC and blood pressure with Edan iM50 patient monitor are measured as well. 18 volunteered subjects are requested to perform the above measurements while fasting and 30 minutes breakfast, and have their blood glucose measured by a OneTouch Ultra- Plus-Flex glucose meter to get a reference ground truth of the predictions. Features are transformed into linearity according to the Beer-Lambert law then a series of partial F-tests are applied on 19 features extracted from PPG, ECG and blood pressure signals to select features with high impact and reliability. An 18 fold cross validation is then performed with on a multiple regression model. This study applies the clinically accepted Clarke Error Grid and the mean absolute relative difference(mARD) as standards of the prediction accuracies, and has a mean R2 of 0.62 and mARD of 9.45%.
Appears in Collections:	[Graduate Institute of Electrical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	131	View/Open

社群 sharing

Loading...