Real-Time Detection of Infusion Site Failures in a Closed-Loop Artificial Pancreas

Abstract

Background: As evidence emerges that artificial pancreas systems improve clinical outcomes for patients with type 1 diabetes, the burden of this disease will hopefully begin to be alleviated for many patients and caregivers. However, reliance on automated insulin delivery potentially means patients will be slower to act when devices stop functioning appropriately. One such scenario involves an insulin infusion site failure, where the insulin that is recorded as delivered fails to affect the patient's glucose as expected. Alerting patients to these events in real time would potentially reduce hyperglycemia and ketosis associated with infusion site failures.

Methods: An infusion site failure detection algorithm was deployed in a randomized crossover study with artificial pancreas and sensor-augmented pump arms in an outpatient setting. Each arm lasted two weeks. Nineteen participants wore infusion sets for up to 7 days. Clinicians contacted patients to confirm infusion site failures detected by the algorithm and instructed on set replacement if failure was confirmed.

Results: In real time and under zone model predictive control, the infusion site failure detection algorithm achieved a sensitivity of 88.0% (n = 25) while issuing only 0.22 false positives per day, compared with a sensitivity of 73.3% (n = 15) and 0.27 false positives per day in the SAP arm (as indicated by retrospective analysis). No association between intervention strategy and duration of infusion sets was observed ( P = .58).

Conclusions: As patient burden is reduced by each generation of advanced diabetes technology, fault detection algorithms will help ensure that patients are alerted when they need to manually intervene. Clinical Trial Identifier: www.clinicaltrials.gov,NCT02773875.

Keywords: artificial pancreas; fault detection; infusion site failure; model predictive control; safety; type 1 diabetes.

Conflict of interest statement

Declaration of Conflicting Interests: The author(s) declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: BAB has received research support from Medtronic, Dexcom, Insulet, Roche, Tandem, and Bigfoot Biomedical and is on advisory boards for Sanofi, Novo-Nordisk, and Becton Dickenson, and was a consultant for Dexcom. GPF is a consult for Abbott Diabetes Care and receives grant funding from Medtronic, Dexcom, Animas, Tandem, Bigfoot and Insulet. TTL has received research funding from Medtronic and Tandem and is currently employed by Insulet. DMM is on the advisory board for Insulet, is a consultant for Abbott Diabetes Care and receives research funding from Medtronic, Roche, and Dexcom. RG receives royalty payments on intellectual property related to the MPC algorithm employed in this study. FJD and ED have patents on the underlying MPC algorithms used in the study, and are currently receiving royalty payments on these patents. BWB has served as a consultant for Becton, Dickinson and Company. All other authors report no COI as it pertains to this work.

Figures

Figure 1.

Actual SF in an artificial pancreas. The shaded time period represents the 4.5-hour time frame after the infusion set failed and before the patient intervenes. The patient must intervene by replacing the infusion set to resume insulin delivery.

Figure 2.

Duration and characteristics of each infusion set insertion under (A) Zone-MPC + SF detection and (B) SAP. The two bars for each patient number represent the two 7-day observation periods under each intervention. Colors designate the status of the infusion set insertion upon removal. Censored sets are those that were removed at the end of each 7-day period even though they had not failed.

Figure 3.

Kaplan-Meier estimates for infusion set survival in a 7-day period. Shaded regions indicate the 95% confidence intervals for the curves (P = 0.58).

Figure 4.

Distribution of the hyperglycemia outcome: the number of minutes the patient experiences hyperglycemia (>250 mg/dL) in the 4 hours preceding an SF (SAP vs SAP + Retrospective Detection:P = 0.085; SAP + Retrospective SF Detection vs Zone-MPC + Real-Time SF Detection P = 0.060). Asterisks mark outliers in the data.

Figure 5.

Case study of false negatives. (A) One patient experienced a much higher incidence of SFs, potentially leading this patient to be more diligent. Only one out of seven SFs was missed for this patient. (B) An example of one patient where the SF dynamics were much slower than those observed in the training set (and those detected as TP).