Non invasive blood glucose estimation using green light photoplethysmography and machine learning

Area of Science:

• Biomedical Engineering
• Medical Devices
• Health Informatics

Background:

• Non-communicable diseases, particularly diabetes, are a leading global cause of mortality.
• Effective diabetes management relies on accurate blood glucose monitoring.
• Current glucose monitoring methods are often invasive, causing patient anxiety and discomfort.

Purpose of the Study:

• To investigate the potential of green light Photoplethysmography (PPG) signals for non-invasive blood glucose monitoring.
• To develop and validate a custom PPG acquisition setup for reliable data collection.
• To assess the accuracy and feasibility of PPG-based glucose monitoring as an alternative to invasive methods.

Main Methods:

• Developed a custom green light PPG acquisition setup for data collection from 80 subjects.
• Collected simultaneous reference capillary blood glucose readings using a lancing device.
• Applied signal processing, feature extraction, correlation analysis, and feature engineering.
• Trained and validated regression models, including Random Forest Regression (RFR), to predict glucose levels.
• Evaluated model performance using R², Mean Absolute Error (MAE), and bias analysis.

Main Results:

• The Random Forest Regression (RFR) model achieved the highest performance with an R² value of 0.92 and MAE of 4.8 mg/dL.
• Predicted glucose levels showed minimal bias across different glucose ranges.
• Mean differences were -5.11 ± 0.80 mg/dL (<90 mg/dL), -3.68 ± 1.24 mg/dL (90-120 mg/dL), and -5.61 ± 0.75 mg/dL (≥120 mg/dL).

Conclusions:

• Green light PPG signals effectively reflect blood glucose levels.
• The developed PPG-based method shows significant potential for accurate, non-invasive glucose monitoring.
• This technology could offer a more comfortable and less anxiety-inducing alternative for diabetes management.