Artificial intelligence-based non-invasive blood glucose measurement

Prediction of blood sugar levels is a critical task in the effective management of diabetes. The study focuses on leveraging the power of machine learning models such as CatBoost, the dataset created in this research was carefully selected for glucose estimation from PPG signals. It consists of data obtained from 50 people who came to Abant İzzet Baysal University Internal Medicine Department with diabetes complaints during the study period, with ethics committee approval. Information for each individual includes laboratory glucose measurements and approximately one minute of recorded finger PPG signals. Among the various machine learning models tested, CatBoost emerged as the best-performing model for predicting blood sugar levels. The CatBoost model demonstrated its efficiency and accuracy in glucose level predictions by achieving an impressive coefficient of determination (R²) metric of 0.7191 and a mean absolute error (MAE) of 25.21. Feature importance analysis highlighted the importance of certain features, such as median difference and kurtosis, in the prediction model built with CatBoost, underlining their important role in determining blood glucose levels. The inclusion of explainable AI techniques increased the interpretability and transparency of predictive models. Using SHAP values and the confusion matrix, the study facilitated a deeper understanding of the models' decision-making process by providing valuable insight into the factors influencing the predictions. In conclusion, this research highlights the potential of machine learning-based approaches in predicting blood glucose levels from PPG signals. Leveraging advanced models such as CatBoost and employing explainable artificial intelligence methods, this study paves the way for improved diabetes management through accurate, non-invasive, and data-driven predictive methodologies. KEYWORDS: Blood Sugar Prediction, Photoplethysmography, Machine Learning, SHAP, XAI