This study utilizes a Mueller matrix-based system to extract accurate glucose levels from human fingertips, addressing challenges in skin complexity. Integration of domain knowledge and data science aims to enhance prediction accuracy using a Random Forest model. The primary goal is to improve glucose level predictions by selecting effective features based on the Pearson product-moment correlation coefficient (PPMCC). The interpolation compensates for delayed glucose concentration. This study integrates domain knowledge and data science, combining a Mueller matrix-based system and a random forest model. It is noted that 16 effective features were identified from 27 test points collected from a healthy volunteer in the laboratory. These features were divided into training and prediction sets in a ratio of 8:2. As a result, the regression coefficient, R2, was 0.8907 and the mean absolute relative difference (MARD) was 6.8%, respectively. This significantly improves prediction accuracy, demonstrating the model's robustness and reliability in accurately forecasting outcomes based on the identified features. In addition, in the Institutional Review Board (IRB) tests at NCKU's hospital, all data passed the same preprocessing and model. The measurement results from an individual diabetic patient demonstrate high accuracy for blood glucose concentrations below 150 mg/dL, with acceptable deviation at higher levels and no severe error zones. Over a three-month period, data from the participating diabetic patient showed a MARD of 4.44% with the R2 of 0.836, and the other patient recorded a MARD of 7.79% with the R2 of 0.855. The study shows the proposed approach accurately extracts glucose levels. Integrating domain knowledge, data science, and effective strategies significantly improves prediction accuracy.
