ABSTRACT
Patients with type 1 diabetes manually regulate blood glucose concentration by adjusting insulin dosage in response to factors such as carbohydrate intake and exercise intensity. Automated near-term prediction of blood glucose concentration is essential to prevent hyper- and hypoglycaemic events in type 1 diabetes patients and to improve control of blood glucose levels by physicians and patients. The imperfect nature of patient monitoring introduces missing values into all variables that play important roles to predict blood glucose level, necessitating data imputation. In this paper, we investigated the importance of variables and explored various feature engineering methods to predict blood glucose level. Next, we extended our work by developing a new empirical imputation method and investigating the predictive accuracy achieved under different methods to impute missing data. Also, we examined the influence of past signal values on the prediction of blood glucose levels. We reported the relative performance of predictive models in different testing scenarios and different imputation methods. Finally, we found an optimal combination of data imputation methods and built an ensemble model for the reliable prediction of blood glucose levels on a 30-minute horizon.
