Towards Automatic Classification of Common Therapy Errors for Diabetes Therapy Support

摘要

Today, one in eleven adults is suffering from diabetes mellitus. Too high or too low blood glucose levels are dangerous. Incorrectly treated, it will lead to severe complications such as strokes, blindness, and ultimately, death. Hence, the correct medication with insulin from diabetes patients is a crucial parameter in therapy. Patients get educated regularly by diabetes experts in training sessions. These sessions contain data review by the experts to identify errors in the patients’ dosage behavior. However, the review is time-consuming, since the manual error identification for a wrong dosage is nontrivial. In this paper we investigate the automatic classification of insulin dosage into three categories, representing correctly ap- plied therapy and the most common therapy faults. We provide the experts with a pre-classified overview of the data, where the common errors are visually highlighted. This saves time in the consultation hour, enabling the expert to spend more time investigating the patients’ individual problems. In our evaluation, we compare multiple classification methods based on dynamic time warping against a convolutional neural network. The results show that the convolutional neural network can achieve accuracy levels that are promising, although further improvements are required.