Optimization of blood glucose prediction with LSTM-XGBoost fusion and integration of statistical features for enhanced accuracy

Abstract

Technological developments, most notably Continuous Glucose Monitoring (CGM) devices, have made it possible to manage diabetes mellitus by accessing trustworthy data, a chronic disease requiring constant monitoring of Blood Glucose (BG) levels to keep them as close to normal as possible, aiming to prevent potentially serious complications. Technological advancements have fuelled research in artificial intelligence to develop accurate methods for predicting future BG levels. These initiatives, supported by numerous studies, aim to improve the quality of life of people with diabetes by anticipating and avoiding dangerous fluctuations in BG levels.

Diabetes, a long-term health condition, requires checking BG levels to maintain them within a safe range and reduce the risk of severe complications. Modern advancements in technology, such as CGM systems, have made it easier to obtain data. This has led to increased interest in intelligence research to create methods for predicting future BG levels. These efforts, backed by studies, strive to improve the well-being of individuals with diabetes by predicting and preventing fluctuations in blood sugar levels.

In this study, our approach focuses on predicting future BG levels in diabetic patients using only CGM data. This approach stands out from many existing methods that require details about patients’ daily activities, such as meals, insulin therapy, and emotional factors. To this end, we explored different approaches for predicting glucose levels in diabetic patients, particularly emphasizing the use of the Long Short-Term Memory (LSTM) model. We examined various scenarios, including LSTM alone, LSTM with the integration of temporal features, and the combination of LSTM with the XGBoost model, incorporating additional features as inputs for the LSTM model.

Experimental results show that the proposed method performs exceptionally well in terms of Root Mean Square Error (RMSE) across both short and long terms. Compared to previous works, the proposed method is considered as effective and accurate even when considering long horizon. Considering horizons of 15, 30, 45, and 60 min, the combination of LSTM and XGBoost, with the integration of additional features for the LSTM model as input, achieved the lowest average values of RMSE, namely respectively 7.97 mg/dl, 9.63 mg/dl, 10.72 mg/dl, and 10.93 mg/dl. This result was validated using CGM data of 12 patients. This approach distinguishes itself by its ability to provide accurate results compared to other methods, emphasizing its potential in improving the management of diabetes mellitus.