Continuous glucose monitors (CGM) can help diabetes patients monitor their blood glucose concentration (BGC) with higher frequency and less manual effort compared to finger sticks. It receives information from a subcutaneous glucose sensor wirelessly. The sensor and CGM are also critical components of an artificial pancreas (AP) which relies on real time glucose concentration information as an input to calculate appropriate insulin infusion rates for patients with type 1 diabetes (T1D). The accuracy of glucose concentration provided by CGM become critical for glucose level management with AP systems. For most of available CGM sensors in the market, the glucose reading is calculated using the current signal is generated by the multilayer reaction in interstitial fluid. This current signal, considered as raw signal, need to be calculated by a predesigned filter to convert it to glucose concentration. However, the current signal may disturbed by local glucose fluctuations, the filter may not be accurate enough to convert the row signal to glucose concentration and the sensor signal may affected by interference from other devices or noise when communicating the data to the AP . As a result, outliers, missing data and large signal noise will make the CGM signal unreliable and cause the AP control systems to suggest incorrect insulin infusion flow rates that will drive patients to hypoglycemia or hyperglycemia.
In order to detect errors in CGM sensor signal, and replace detected erroneous signal with model based signal, two techniques, robust Kalman filter (RKF) and locally weighted partial least squares (LW-PLS), are introduced to build a hybrid CGM signal fault detection and redundancy analysis system. This innovative hybrid system leverages the advantages of automatic measurement error elimination with RKF and database supported prediction with LW-PLS. The system is designed with varying thresholds to distinguish between signal faults and real dynamic changes so that the number of erroneous fault detections can be reduced. This system is tested with clinical data from T1D patients. Data from 10 clinical experiments were analyzed, each experiment lasting for 56 hours. More than 500 CGM sensor errors were added to raw CGM data for detection and analysis. The results indicate that the proposed system can successfully detect most of the erroneous CGM signals and substitute with values computed by functional redundancy generated with the hybrid system.