Prospective Detection of Hypoglycemia and Near-Hypoglycemia Inducing Pressure-Induced Sensor Attenuation Onset in Continuous Glucose Monitoring Time Series

BACKGROUND: Continuous glucose monitoring (CGM) sensors are vulnerable to pressure-induced sensor attenuations (PISAs). Pressure-induced sensor attenuations are characterized by a rapid but erroneous drop in blood glucose (BG) readings, and if these BG readings are hypoglycemic or near-hypoglycemic they can cause unwarranted actions to be taken (eg, triggering of hypoglycemia alarms, shut-off of an insulin pump). In this work, we propose a prospective algorithm (PA) to detect the onset of PISAs where the minimum sensor-reported BG value during the PISA is less than 85 mg/dL. METHODS: The PA is composed of a screening module and a probability estimation module (PEM). The PEM was trained on a data set with 846 PISAs (59 122 hours of data, 67 patients). The optimal version of the PA was selected by assessing performance of each version of the PA on a data set with 89 PISAs (8189 hours of data, 11 patients). Given the highly imbalanced nature of the problem (baseline prevalence of PISAs onset in the testing set is 1.2%), performance was primarily assessed using precision-recall (PR) curves. RESULTS: The final version of the PA was tested on a data set with 325 PISAs (26 890 hours of data, 33 patients), resulting in an area under the PR curve of 0.590 (95% confidence interval [0.509, 0.672]), indicating excellent performance against the baseline occurrence. CONCLUSIONS: The results indicate that the PA could feasibly be used to inform other decision-making processes, for example, canceling of hypoglycemia alarms that would otherwise be triggered.