BUMS: A Novel Balanced Multi-Model Machine Learning System for Real-Time Blood Glucose Prediction and Abnormal Glucose Events Detection

Diabetes, a chronic condition with a growing global prevalence, exerts lasting effects on individuals' health and well-being, necessitating continuous control and monitor of blood glucose for stable levels. Meeting this fact, recent years have seen an increasing adoption of machine learning algorithms to accurately predict blood glucose values. In this work we present a novel multi-model approach, BUMS (Balanced Multi-model Scheme), designed to accurately predict blood glucose levels in real time. The primary goal of this system is to mitigate the risks associated with critical blood glucose events, such as hypoglycemia and hyperglycemia, which significantly impact individuals living with diabetes. BUMS combines three distinct algorithms: Long Short-Term Memory (LSTM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost), all leveraging continuous glucose monitoring data collected at 5-minute intervals. To ensure robustness and balance in the predictive models, we introduce a pre-Trained Balancer into the multi-model architecture. Our approach is validated using data from the publicly available DirectNet dataset, featuring continuous blood glucose measurements from 30 patients. The Balancer module is pre-Trained on data from 5 patients before being tested on data from the remaining 25 patients, employing Linear Regression as its foundation. We evaluate the performance of our system across various prediction horizons, ranging from 25 to 855 minutes, using 169 test cases. The results demonstrate an overall Root Mean Square Error (RMSE) of 4.8125, indicating the model's high predictive accuracy. Notably, among the 169 test cases, only one case was incorrectly identified, resulting in an accuracy rate of 96.29% in detecting hypoglycemic events.