Models for In Silico Simulation of Possible Failures of Continuous Glucose Monitoring Sensors

Objective: Type 1 diabetes simulators, such as US2008/067725, are widely used to test in silico the performance of closed-loop control algorithms for the artificial pancreas. More realistic/challenging simulations could be obtained with a proper description of the continuous glucose monitor (CGM) sensor malfunctionings often experienced in clinical practice. In this study, we propose models of the most common failure episodes that deteriorate CGM device performance. Method: We analyzed 24 CGM time series obtained using the Dexcom® SEVEN® PLUS sensor, each lasting approximately 8 h. Blood glucose (BG) references were also collected in parallel. Through the comparison of reference BG and CGM traces, we identified three possible failure types: disconnection, i.e., transmitter–receiver connectivity losses; invalid output, i.e., sensor self-diagnosed unreliable readings; and transient loss of sensitivity caused by pressure on the sensor, leading to sudden drops in sensor reading. Result: A two-state (working/faulty) Markov chain, with transition probabilities inferred from the data, proved to effectively describe disconnection and short-in-time invalid outputs. Transient losses of sensitivity were modeled as a square-wave Boolean disturbance, describing pressure on the sensor, subtracted to the sensor output after passing through a first-order low-pass system that models pressure-induced slow changes in the insertion site. Probability distributions of model parameters (pressure duration, system gain, and time constant) were estimated from the data. Conclusion: The proposed models proved to be able to effectively capture essential features of three classes of common failures of CGM devices. Thus they can be usefully integrated in type 1 diabetes simulators to allow a more realistic in silico testing of artificial pancreas control algorithms and of their robustness against occasional malfunctionings of CGM devices.