Development of a Non-Invasive Blood Glucose Detection System Based on Photoplethysmogram (PPG) Signals and Random Forest Regressor

Abstract

Frequent checking of blood glucose levels is important for managing diabetes well. However, traditional methods that require taking blood from the finger are uncomfortable, expensive, and can lead to patients not following the recommended routine. This work introduces a low-cost, non-invasive system for estimating blood glucose levels using near-infrared optical sensing. Methods: The system uses an MAX30100 sensor along with an Arduino microcontroller to collect Photoplethysmogram (PPG) signals. From these signals, the system identifies Heart Rate (BPM) and Oxygen Saturation (SpO2). To handle the complex relationship between optical absorption and glucose levels, a Random Forest Regression (RFR) model was created. This model uses expanded polynomial features, including logarithmic and quadratic terms. The model was trained and tested using a clinical dataset with 164 synchronized measurements. To improve accuracy, a strict process was used to remove any measurements that were higher than 200 mg/dL, which could cause sensor saturation. For real-time use, a dynamic Exponential Weighted Moving Average (EWMA) filter was added to reduce the random variations in the optical signals. The system was tested on a set of 33 blind test samples, and it showed a Mean Absolute Error (MAE) of 15.51 mg/dL and a Root Mean Squared Error (RMSE) of 20.07 mg/dL. The Coefficient of Determination $\left(\mathrm{R}^{2}\right)$ was 0.544. The system performed well in general, showing slightly better results on new data compared to the data it was trained on.