Article Preview
Top1. Introduction
Diabetes, the eighth leading cause of death in the world (World Health Organization, 2014), is due to either the body not producing enough insulin or the body not using insulin properly, which causes blood glucose levels (BGLs) to rise higher than normal. This high BGL leads to common symptoms of diabetes such as urinating often, feeling very thirsty, increasing hunger, and causing many complications (e.g., heart disease, stroke, kidney failure, etc.).
According to the statistical data from the 2014 National Diabetes Statistics Report (Centers for Disease Control and Prevention, 2014), 29.1 million Americans (i.e., 9.3% of the U.S. population) had diabetes in 2012, and 25.8 million Americans (i.e., 8.3% of the U.S. population) had diabetes in 2010 (American Diabetes Association, 2014). This 12.79% increase of diabetes patients just within two years has drawn a significant attention from the medical domain. To maintain and control a normal BGL, diabetes patients can be treated by injecting insulin. The major effect of injecting insulin on diabetes patients is to lower their BGLs to normal, but it might cause an insulin reaction, often referred to as hypoglycemia, that is, a condition characterized by an abnormally low BGL. Hypoglycemia can cause serious consequences, such as fatigue, dizziness, confusion, headaches, blurred vision, seizures, unconsciousness, or even death, on the patients. In order to detect whether a patient is experiencing hypoglycemia, doctors typically check the patient’s BGL routinely. A big challenge is how to accurately and effectively determine the BGL monitoring threshold to prevent the patients from suffering hypoglycemic events. This is exactly the focus of this paper.
Currently, the existing approaches to identify and detect hypoglycemia based on the routine BGL check can be roughly divided into two categories: domain-knowledge-based and machine-learning-based. The former approach relies solely on medical experts’ knowledge. Based upon their knowledge and experiences, medical experts have identified a BGL monitoring threshold, i.e., 70 mg/dL (American Diabetes Association, 2014), which can be used to detect the occurrence of hypoglycemic events. When the diabetes patients’ BGLs become lower than 70 mg/dL, an alarm that signals the patients suffering the symptoms of hypoglycemia is triggered. Often this threshold may work well with a number of patients since it is determined by the medical experts based on their past experiences, observations, intuitions, and domain knowledge. However, it is not always accurate, as different patients have different health conditions; hence one value does not fit them all. In addition, the threshold is static, but the problem that this paper deals with is often dynamic in nature. Each patient’s health condition is constantly impacted by many unknown and uncontrollable factors and the surrounding environment, e.g., temperature, food, weather, etc. The domain-knowledge-based approach lacks a formal mathematical computation that dynamically learns the monitoring parameter to meet the needs of each patient’s health condition.