COMPARATIVE ANALYSIS OF LSTM AND BILSTM MODELS FOR PATIENT BLOOD GLUCOSE PREDICTION

Abstract

Accurate prediction of blood glucose levels is critical for timely management of hyperglycemia and hypoglycemia in patients with Type 1 diabetes mellitus. Diabetes represents a growing global health challenge, with ineffective glycemic control leading to severe long-term complications including cardiovascular disease, neuropathy and kidney failure. Numerous studies have demonstrated the effectiveness of deep learning approaches for blood glucose prediction, as they can capture complex nonlinear dynamics and long-term temporal dependencies in physiological signals. In this study, we present a comparative analysis of Long Short-Term Memory (LSTM) and Bidirectional Long Short Term Memory (BiLSTM) architectures for blood glucose prediction 30 minutes ahead, using a personalized modeling approach with one model trained per patient. Both models were evaluated on the OhioT1DM dataset (2018 and 2020 releases), comprising continuous glucose monitoring (CGM) data from 12 patients with Type 1 diabetes, recorded at 5-minute intervals. The dataset includes physiological and behavioral signals such as CGM readings, insulin doses, meal information, and physical activity data. Both models take a 2-hour historical window (24 time steps, 7 features) as input. The LSTM uses two layers with 64 hidden units each, while the BiLSTM employs two bidirectional layers with an effective hidden size of 128 units per layer. Both models were trained for 50 epochs using the Adam optimizer (lr = 0.001), Mean Squared Error (MSE) loss, dropout (0.2), and gradient clipping. Model performance was assessed using RMSE (Root Mean Squared Error), Mean Absolute Error (MAE), Mean Absolute Relative Difference (MARD), and Time in Good Range (TG). The LSTM achieved a mean RMSE of 31.44 ± 24.16 mg/dL, MAE of 18.73 ± 9.23 mg/dL, MARD of 11.00 ± 3.25%, and TG of 78.17 ± 7.24%. The BiLSTM achieved a mean RMSE of 29.83 ± 17.42 mg/dL, MAE of 18.78 ± 7.44 mg/dL, MARD of 11.30 ± 3.35%, and TG of 76.21 ± 8.63%. Results indicate that neither architecture consistently outperforms the other across all patients, suggesting that individual glucose dynamics play a significant role in model performance. These findings highlight the potential of personalized deep learning approaches in clinical decision support systems for diabetes management.