Research on multi-parameter fusion non-invasive blood glucose detection method based on machine learning.

Area of Science:

• Biomedical Engineering
• Machine Learning Applications
• Medical Diagnostics

Background:

• Traditional blood glucose testing methods present challenges including pain and discontinuous data acquisition.
• There is a need for accurate, non-invasive alternatives for continuous glucose monitoring.

Purpose of the Study:

• To develop and validate a multi-parameter fusion, non-invasive blood glucose detection method.
• To leverage machine learning and photoplethysmography (PPG) signal analysis for improved glucose monitoring.

Main Methods:

• Utilized a signal validity check based on correlation operations for PPG data processing.
• Developed two non-invasive glucose detection models using bootstrap aggregation and random forests algorithms.
• Employed feature parameter analysis of PPG signals for comprehensive blood glucose prediction.

Main Results:

• The random forests algorithm-based model demonstrated superior accuracy in blood glucose prediction.
• Achieved a correlation coefficient of 0.972, a mean square error of 0.257, and a relative error below ±20%.
• The proposed method's relative error meets Chinese national standards for blood glucose prediction.

Conclusions:

• The developed non-invasive blood glucose testing method meets clinical accuracy requirements, as confirmed by Clarke Error Grid Analysis.
• The random forests model offers a promising, accurate, and non-invasive approach to blood glucose monitoring.
• This method addresses the limitations of traditional testing, offering better patient comfort and data continuity.