研究目的
To overcome individual discrepancies in non-invasive glucose measurement by designing a learning model that customizes parameters for each individual using multi-wavelength absorbance and machine learning algorithms.
研究成果
The proposed learning model effectively reduces individual discrepancies in non-invasive glucose measurement, achieving a correlation coefficient of 0.8 after 60 calibrations. It offers low cost and good robustness but requires personalized training. Future work should involve larger clinical trials and hardware improvements.
研究不足
The study is preliminary with only six subjects, limiting generalizability. The model requires individual calibration (60 invasive measurements for high accuracy) and cannot be directly transplanted between individuals without retraining. Hardware and environmental factors may affect accuracy.
1：Experimental Design and Method Selection:
The study involves designing a non-invasive glucose measurement prototype using multi-wavelength light absorbance based on the Beer-Lambert law. Methods include Independent Component Analysis (ICA) for feature extraction and Random Forest (RF) for glucose estimation.
2：Sample Selection and Data Sources:
Data were collected from six adult subjects (five diabetic, one healthy) over 15 days, with measurements taken five times daily using the prototype and a reference invasive glucose meter.
3：List of Experimental Equipment and Materials:
Prototype device (Earlight) with LEDs at 490, 660, 730, 850, and 930 nm, silicon photodiode, MAX471 current sensing IC, STM32f103 microcontroller, and SANNUO invasive glucose meter.
4：Experimental Procedures and Operational Workflow:
Absorbance data were collected from the earlobe, processed through ICA to extract independent components, and used in RF for glucose estimation. Training involved 80% of data from one subject, with testing on the remaining 20%.
5：0%. Data Analysis Methods:
5. Data Analysis Methods: Accuracy was evaluated using correlation coefficients and Clarke Error Grid Analysis (EGA). Convergence was tested by varying training dataset sizes.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
