Glucose Prediction using Wide-Deep LSTM Network for Accurate Insulin Dosing in Artificial Pancreas.

Abstract

Closed-loop diabetes management has been shown to indicate improved glycaemic control and better compliance over open loop diabetes management. Currently, commercially available diabetes management devices rely on continuous glucose monitoring (CGM) sensors which monitor glucose levels from the interstitial fluid (ISF). As there exists a physiological delay between the blood glucose levels compared to the ISF glucose levels, it is crucial to predict or forecast glucose levels, in order to prevent hyperglycaemic events due to delayed insulin dosing. Accuracy of the forecast also needs to be optimum such that overdosing on insulin does not lead to hypoglycaemia. In this paper, we describe a novel Long Short Term Memory (LSTM) network which follows a wide and deep approach for different features to deliver an accurate glucose prediction output. It achieved a Mean Absolute Relative Difference (MARD) of 2.61 and Root Mean Squared Error (MSE) of 5.04. Clinical relevance- This work is relevant for closed-loop diabetes management devices, which are currently being used to manage Type 1 Diabetes (T1D).