Prediction and Analysis of Blood Glucose Levels based on Tabnet

Abstract

Background: Blood glucose level prediction plays a significant role in the management of diabetes. Accurate prediction of blood glucose levels helps patients and doctors to make informed decisions regarding diet, exercise, and medication. The use of machine learning algorithms for blood glucose prediction has gained attention in recent years. Tabnet is one such algorithm that has shown promising results in various prediction tasks. Aim: The aim of this study is to evaluate the performance of Tabnet for blood glucose level prediction and compare it with other commonly used algorithms, including LR, DT, SVM, RF, and EN. Methods: A dataset of blood glucose levels of diabetic patients was used for this study. The dataset was preprocessed, and features were selected using correlation-based feature selection. Tabnet and other algorithms were trained on the dataset using 5-fold cross-validation. The performance of each algorithm was evaluated using root mean squared error (RMSE) and mean squared error (MSE). Results: The experimental results showed that Tabnet performed the best in terms of RMSE and MSE, with values of 0.5097 and 0.2523, respectively. The LR algorithm had an RMSE of 0.5126 and an MSE of 0.2629, while the DT algorithm had an RMSE of 0.7543 and an MSE of 0.5689. The SVM algorithm had an RMSE of 0.5165 and an MSE of 0.2663, while the RF algorithm had an RMSE of 0.5188 and an MSE of 0.2691. The EN algorithm had an RMSE of 0.5547 and an MSE of 0.3077. Conclusion: In this study, Tabnet was found to be the best algorithm for blood glucose level prediction compared to other commonly used algorithms. The results demonstrate the potential of Tabnet for predicting blood glucose levels in diabetic patients, which can assist in effective diabetes management.