InsNET: Accurate Basal and Bolus Insulin Dose Prediction for Closed Loop Diabetes Management

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Endocrinology

Background:

• Closed-loop diabetes management systems demonstrate superior glycemic control and patient compliance compared to open-loop systems.
• Deep learning models are increasingly utilized for developing critical components of artificial pancreas technology.
• Accurate insulin dose estimation is crucial for the efficacy of artificial pancreas systems.

Purpose of the Study:

• To propose a novel deep learning model, InsNET, for estimating basal and bolus insulin levels in type 1 diabetes patients.
• To enhance closed-loop diabetes management by incorporating physical activity data alongside traditional inputs.
• To improve the accuracy of insulin dose determination within artificial pancreas systems.

Main Methods:

• Developed InsNET, a deep learning model employing a Wide-Deep combination of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) layers.
• Integrated physical activity level as an input feature, differentiating it from prior models that used only continuous glucose monitoring (CGM), carbohydrate intake (CHO), and past insulin dosages.
• Validated the model using in-silico datasets, specifically the UVA/Padova and mGIPsim datasets.

Main Results:

• InsNET achieved a Mean Absolute Error (MAE) of 0.002 and Root Mean Squared Error (RMSE) of 0.007 on the UVA/Padova Dataset.
• The model demonstrated strong performance on the mGIPsim Dataset, with an MAE of 0.001 and RMSE of 0.003.
• These results indicate high accuracy in estimating insulin dosage requirements.

Conclusions:

• The proposed InsNET model offers a significant advancement in artificial pancreas technology for type 1 diabetes.
• Accurate basal and bolus insulin dose determination is achievable with deep learning, incorporating diverse physiological inputs.
• InsNET holds clinical relevance for improving insulin therapy management and patient outcomes in closed-loop systems.