Type-1 Diabetes is a chronic autoimmune disorder where the body’s white blood cells attack the pancreatic betacells. Neither its cause nor the means to prevent it are known. Patients require external insulin to manage their blood glucose levels and prevent conditions such as Hyperglycemia and Hypoglycemia. This raises the necessity for the knowledge of real-time Glucose Reading. This study aims to predict blood glucose levels using deep learning algorithms such as Long Short Term Memory (LSTM) and Convolutional Recurrent Neural Network (CRNN) on the OhioT1DM dataset. The former model uses only historical blood glucose level data, while the latter includes additional information such as insulin shots and meal carbs. The prediction horizon is set to 60 minutes, and the models are evaluated using Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). The obtained results demonstrate good accuracy for both models, with LSTM achieving RMSE = 4.208 ± 0.92 mg/dL and MAE = 3.22 ± 0.58 mg/dL, and CRNN achieving RMSE = 4.0846 ± 0.9 mg/dL and MAE = 3.0845 ± 0.52 mg/dL. These predictions could help optimize closed-loop insulin delivery systems for better management of Type-1 Diabetes.
Paper
Full text
Deep Learning based Time Series Modelling for Glucose level prediction of Type-1 diabetes
Semantic Scholar · Medicine · 2023
Abstract
Type-1 Diabetes is a chronic autoimmune disorder where the body’s white blood cells attack the pancreatic betacells. Neither its cause nor the means to prevent it are known. Patients require external insulin to manage their blood glucose levels and prevent conditions such as Hyperglycemia and Hypoglycemia. This raises the necessity for the knowledge of real-time Glucose Reading. This study aims to predict blood glucose levels using deep learning algorithms such as Long Short Term Memory (LSTM) and Convolutional Recurrent Neural Network (CRNN) on the OhioT1DM dataset. The former model uses only historical blood glucose level data, while the latter includes additional information such as insulin shots and meal carbs. The prediction horizon is set to 60 minutes, and the models are evaluated using Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). The obtained results demonstrate good accuracy for both models, with LSTM achieving RMSE = 4.208 ± 0.92 mg/dL and MAE = 3.22 ± 0.58 mg/dL, and CRNN achieving RMSE = 4.0846 ± 0.9 mg/dL and MAE = 3.0845 ± 0.52 mg/dL. These predictions could help optimize closed-loop insulin delivery systems for better management of Type-1 Diabetes.