Platform for precise, personalised glucose forecasting through continuous glucose and physical activity monitoring and deep learning

Abstract

Emerging research has demonstrated the advantage of continuous glucose monitoring for use in artificial pancreas and diabetes management in general. Recent studies demonstrate that glucose level forecasting using deep learning can help avoid postprandial hyperglycemia (≥ 180 mg/dL) or hypoglycemia (≤70 mg/dL) from delayed or increased insulin dosing in artificial pancreas. In this paper, a novel hybrid deep learning framework with integration of content-based attention learning is presented, to effectively predict the glucose measurements with prediction horizons (PH) = 15, 30 and, 60 minutes for T1D and T2D patients based on past data. We also present a complete cloud-based system and mobile app used for collecting CGM sensor, physical activity data, CHO values and insulin measurements to perform glucose forecasts using the proposed model running on Cloud. This model was validated using clinical data of individual with Type 1 diabetes (OhioT1DM) and individual with Type 2 diabetes. The mean absolute relative difference (MARD) was 12.33±3.15, 7.14±1.76% for PH=60 and, 30 min respectively on OhioT1DM clinical Dataset. The root mean squared error (RMSE) was 29.41±5.92 mg/dL and 17.19±3.22 mg/dL and the mean absolute error (MAE) was 21.96±4.67 mg/dL and 12.58±2.34 mg/dL for PH=60 and, 30 min respectively on the same clinical dataset. It was observed that inclusion of physical activity leads to improved glucose forecasting accuracy. Furthermore, all these results were obtained by training the model on only 8 days of clinical data of a single patient, followed by testing on clinical data on the following days. The results indicate that training on a single patient data may lead to better personalisation and better glucose forecasting results compared to existing works.