BiLSTM-TSBGP: A Univariate Blood Glucose Prediction Model for Type-1 Diabetes using Continuous Glucose Monitoring Data

Abstract

Effectively managing blood glucose levels is crucial for individuals with Type 1 diabetes, as extreme highs or lows can lead to severe health complications. Wearable technologies, such as continuous glucose monitoring (CGM) devices, have emerged as vital tools in diabetes management, enabling patients to monitor glucose levels and administer insulin proactively. These devices provide a wealth of time-series data, making them ideal for machine learning applications. In this study, we developed the BiLSTM-TSBGP (Bidirectional Long-Short Term Memory Univariate Blood Glucose Prediction) model and evaluated its performance on two datasets: the Ohio Type 1 Diabetes Mellitus (OhioT1DM) dataset and the UVA/Padova Type 1 Diabetes Metabolic Simulator dataset. The OhioT1DM dataset, comprising data from 12 subjects, provides continuous glucose monitoring and insulin infusion records at 5-minute intervals. On average, for the Ohio dataset, our model achieved a Root Mean Square Error (RMSE) of 12.89 mg/dL and Mean Absolute Error (MAE) of 10.65 mg/dL for a 30-minute prediction horizon (pH), and RMSE of 17.38 mg/dL and MAE of 12.30 mg/dL for a 60-minute pH. The UVA/Padova dataset, featuring simulated data for 10 adult profiles, yielded RMSE values of 10.71 mg/dL and 16.24 mg/dL for 30-minute and 60-minute pH, respectively, with corresponding MAE values of 7.55 mg/dL and 11.72 mg/dL. Our results show that the BiLSTM-TSBGP model outperformed other models in forecasting blood glucose levels and offering a reliable tool for managing Type 1 diabetes.