The causes and mitigations of suboptimal accuracy are complex and require further study. We address here the question of whether the use of two or more sensors yields better accuracy than one.
The concept of redundancy is not new. It has long been used in settings in which failure could be catastrophic. For example, many computer systems on National Aeronautics and Space Administration (NASA) spacecrafts are redundant and employ the concept of voting. Voting algorithms process multiple data streams and reject discrepant or outlying data. In this manner, the impact of a malfunctioning unit can be minimized. The occurrence of severe hypoglycemia in a person with diabetes is analogous to failure of the flight control system of a NASA spacecraft; both are life-threatening.
The use of two or more sensors in a closed-loop system is attractive for several reasons. From a practical standpoint, it provides a reserve device for cases of sensor or telemetry failure. In addition, depending on the specific sensor, when a new sensor is inserted, it must be in place for at least 2 to 10 hours before calibration can proceed and sensed glucose readings are provided. If the only sensor in place fails, one must wait 2 to 10 hours for the new sensor to stabilize.
The concurrent use of two sensors in an individual also increases the chance of at least one being highly accurate. We routinely use two sensors concurrently in our closed-loop studies and choose the more accurate one to provide input to the controller. From recent closed-loop data, we prepared two figures. These figures illustrate sensor data in different individuals with type 1 diabetes, each of whom wore two sensors. In both figures, the sensor is calibrated only once, at the beginning of the study. represents a situation in which the two sensors track each other very closely over the 9-hour period. The situation in is quite different. In this case, sensor 2 (the higher tracing) tracked blood glucose quite well and had a mean ARD of 11.7%, whereas sensor 1 functioned poorly with a mean ARD of 24.8%. In a closed-loop system, this degree of inaccuracy would affect the rate of insulin delivery substantially. If the less accurate sensor were used in this case, it would have led to inadequate delivery of insulin with subsequent hyperglycemia. Fortunately, sensor 2 was used for the vast majority of this study to control the insulin delivery rate, and good glycemic control was achieved.
Figure 1. Two subcutaneous sensors in a person with type 1 diabetes undergoing closed-loop control. Blue symbols indicate data from sensor 1, and red symbols indicate data from sensor 2. Note that the two sensors track each other very well; the tracings are nearly (more ...)
Figure 2. Two subcutaneous sensors in a person with type 1 diabetes undergoing closed-loop control. Blue symbols indicate data from sensor 1, and red symbols indicate data from sensor 2. Note that the two sensors track each other poorly. Sensor 1 registered low (more ...)
There are several potential ways of using data from more than one sensor, and this topic has not been well studied. Using the average of two signals is not always the optimal method because if one sensor is performing very poorly, the average is also inaccurate to some extent. Another option is to compare the two sensor signals and avoid using sensor data when the two signals are discrepant beyond a specified criterion. This method was examined by Schmidtke and colleagues16
several years ago. This technique significantly improved the number of glucose readings in regions A and B of the Clarke error grid from 92.4 to 98.8%. Of course, in a closed-loop setting, the choice to avoid using either sensor deprives the controller of afferent input.
In order to address possible advantages of sensor redundancy, we examined the last year of data from our human closed-loop study. Given this study has not yet been completed and data have not yet been submitted for peer review, our findings are presented in general form. During each study, subjects with type 1 diabetes wore two sensors for 28 hours. Early in the course of the study, the more accurate of the two sensors was selected for the control of insulin and glucagon delivery. The accuracy of both sensors was followed for the duration of the study and was compared to venous glucose performed every 10 minutes on the highly accurate HemoCue® analyzer. Each sensor was calibrated at the start of the study and again 4 hours later. The “selected” sensor had a mean ARD of 14%, which was significantly lower when compared to the unselected sensor, with a mean ARD of 19%. In approximately 65% of subjects, the mean ARD of the two sensors were within four mean ARD percentage points of each other and close to 75% were within seven points of each other. Thus, in approximately 25% of cases, there was a large discrepancy in accuracy that exceeded seven mean ARD percentage points. The accuracy difference in this group of individuals would have led to substantial differences in the amounts of insulin and glucagon given by the algorithm, depending on which sensor was used. These findings support the use of two sensors in settings where accuracy is critical. We also found that in over 80% of cases, the sensor that was selected early continued to be the more accurate of the two throughout the remainder of the 28-hour study. In the small number of cases in which this was not the case, the overall difference in accuracy between selected and unselected sensors was less than 1.5 mean ARD percentage points.
Although wearing three or more sensors is impractical with current sensor technology, sensor arrays with multiple sensing units may be available in the future. The advantage of three or more sensors lies in the ability to “vote out” data from one or more sensors when the reading is discrepant from the others. This technique is based on the fact that sensor signals that are quite similar to others in the array are usually almost always more accurate than outliers. In animals, we tested such a technique using a statistical technique termed the Z score with median absolute deviation (ZMAD), which is based on median statistics. The subcutaneous sensor arrays contained four sensing units, each with its own platinum indicating electrode and its own telemetric channel. The Ag/AgCl reference electrode was shared among the four sensing units. A Z score for each sensing unit was calculated every several minutes. This score compared how much the sensed glucose from one unit deviated from the median sensed glucose of all units. If the Z score rose above a predetermined threshold, then data from that unit were considered an outlier and were therefore automatically excluded. This technique can be employed in real time. During long-term implantations, sensor array accuracy using the ZMAD technique was significantly and substantially better than using an average of the sensed glucose values.17
The use of redundant sensors addresses sensor drift, but does not address sensor delay. Delay, which is influenced by physiologic elements, inherent sensor properties, and data filtering, is expected to affect all sensors similarly. In a similar fashion, if all sensors in an array are similarly affected by calibration error, such as nonlinearity or miscalculation of background current, redundancy would be unlikely to improve accuracy.