MiniMed 780G System Outperforms Other Automated Insulin Systems Due to Algorithm Design,Not Bias: Response to Inaccurate Allegations

Dear Editor,

With interest we have read the article entitled “Real-World Evidence of Omnipod 5 Automated Insulin Delivery System Use in 69,902 People with Type 1 Diabetes.” ¹ The authors showed that not all automated insulin delivery systems perform similarly, and that the MiniMed 780G system (MM780G) ² outperforms Omnipod 5 system (OP5), ¹ Control-IQ system (CIQ), ³ and CamAPS-FX system (CFX) ⁴ (Fig. 1). However, the authors surprisingly and incorrectly attributed the superior performance of MM780G to selection bias, rather than differences in algorithm design.

FIG. 1.

Times in ranges from real-world users of OP5, CIQ, CFX, and MM780G, as cited in the OP5 real-world article. ¹ Values are represented as means/*, for this study values are represented as median. As the mean TIR was consistently lower than the median TIR throughout the entire article, the median TIR is expected to be an overestimation/^†, no break down made for values <54 mg/dL (values not reported in original article)/^‡, no break down made for values >250 mg/dL (values not reported in original article)/¥, number derived from calculation/^§, extracted from the 1-month results/. Two articles have detailed breakdowns of system settings. The right panel displays users who have used settings resulting in the best outcomes. The right panel shows those users utilizing settings that have provided the best outcomes. For OP5 these are users with a glucose target of 110 mg/dL. For MM780G system these are users with consistent use of a glucose target of 100 mg/dL and an active insulin time of 2 h. CFX, CamAPS-FX system; CIQ, Control-IQ system; MM780G, MiniMed 780G system; OP5, Omnipod 5 system; TIR, time in range.

A selection bias is improbable. Real-world data on MM780G comes from CareLink Personal database (CP). MM780G users are trained by certified product trainers or Medtronic personnel, and creating a CP account is part of this training. It is estimated that >95% of MM780G users have such an account. In addition, >90% of EMEA (Europe, Middle East, and Africa) users have consented to their data being used in research. Pump and Continuouns Glucose Monitoring (CGM) data are collected through scheduled uploads: automatically (nightly), manually, or at clinic device download stations (as preferred). In our analysis, >98% of EMEA uploads were automated. Indeed, real-world studies such as these lack visibility into the data of the few users without accounts/consent. Nonetheless, there is no indication of an overrepresentation of a specific phenotype among those missing data. The unlikelihood of selection bias is further supported by the consistency in outcomes from in silico to real-world studies, including the latest publication on >100,000 users.

The MM780G superior performance is due to algorithm design differences. Unlike OP5, MM780G allows glucose targets (GTs) as low as 100 mg/dL, whereas OP5's lowest is 110 mg/dL. MM780G can also deliver auto-correction boluses beyond the maximum bolus amount, as frequently as every 5 min. In contrast, CIQ aims for a broad glucose range (nonactivity range: 112.5–160 mg/dL) rather than a GT, and its auto-correction boluses differ substantially. CIQ auto-correction boluses can only be given once per hour, may start late (when glucose is predicted to exceed 180 mg/dL in 30 min vs. when glucose levels surpass 120 mg/dL for MM780G), and may be lower in insulin (giving 60% of the insulin needed to reach a target of 110 mg/dL, compared with potential 100% for a target of 120 mg/dL). For more on MM780G's algorithm and features, see Grosman et al. ⁵

The reported time in range (TIR) of MM780G is actually a conservative approximation of the algorithm's true potential. Authors included all users with ≥10 days of sensor glucose data (which is considered the minimum). Furthermore, they used all data from eligible users; this also encompassed data from times when MM780G was not in automation, which for few users constituted even the majority of measurements. In contrast, the real-world article of OP5 included only users with ≥90 days of CGM data, aiming to reflect typical device use and minimize transition effects. They also required ≥75% of days with CGM data to have ≥220 readings per day. This hints that, in fact, the real-world evidence for OP5 may suffer from selection bias and offer an optimistic estimate for TIR.