Sensor abnormality detection based on global prediction model for type I diabetes

Instead of self-monitoring of blood glucose which is in general only taken four times per day, continuous glucose monitoring (CGM) can provide real-time measurements of glucose levels more frequently. However, for successful clinical use, it is important to make sure that the sensor is running normally. Unfortunately, sensor abnormality has not been well analyzed and detected online although it is a very popular problem in real case and may result in unreliable CGM measurements. In the present work, a sensor abnormality detection method is developed based on prediction errors to identify sensor problems online. And the feasibility of a global prediction model for sensor abnormality detection is addressed by comparing the detection results based on prediction errors using subject-dependent prediction model and those using a global model. Different from the conventional glucose monitoring which in fact works as a low-level monitoring tool and focuses on direct realtime display of CGM readings, the proposed method is deemed to be a super-level monitoring tool which focuses on detecting the undesirable sensor abnormality by analyzing the underlying time-wise glucose correlations. The feasibility of the proposed method to serve as a completely new glucose monitoring engine is successfully assessed using clinical data.