Contents I. Introduction
Diabetes mellitus is a common and serious autoimmune disease characterized by the inability of the pancreas to produce insulin, an essential hormone needed to convert food into energy, and thus to regulate blood glucose concentration. The management goal is to delay or prevent serious long-term diabetic complications, including blindness, kidney failure, strokes, heart attacks [1]. According to Diabetes Controls and Complication Trial [1] the risk of diabetes can be prevented by proper blood glucose monitoring and regulation. Hence, in order to control blood glucose evolution, self-monitoring of blood glucose with use of glucometers (from capillary blood obtained through finger prick) or continuous glucose monitoring devices is essential. However, blood glucose levels abnormally fluctuate, and too little insulin injection results in chronic high blood glucose levels, too much can cause hypoglycemia. For both patients and doctors diabetes management can pose a rather complicated task: patients are required to track their blood glucose levels and daily activities, and doctors should make appropriate therapeutic adjustments based on the monitoring data. A clinically important task in Diabetes treatment is prevention of nocturnal hypoglycemic events, as blood glucose level falls below 70 mg/dl [2]. Naturally, would be beneficial if the problem of blood glucose regulation could be treated proactively, i.e. the alerts would be given not at the moments of blood glucose excursions, but beforehand in a predictive way based on previous blood glucose measurements. This task can be addressed as a learning-to-rank problem. However, this task is ill-posed, so application of regularization methods is required.
