PID vs. MPC Algorithms in Automated Insulin Delivery

Introduction to Automated Insulin Delivery

An automated insulin delivery (AID) system has three working parts: a continuous glucose monitor (CGM), an insulin pump, and a control algorithm, the software that turns each glucose reading into a decision about how much insulin the pump delivers [1]. These systems were developed for people with type 1 diabetes, and their indications are widening; in February 2025, for example, the FDA cleared Tandem's Control-IQ+ for adults with type 2 diabetes [9]. Researchers have built several kinds of controller, including proportional-integral-derivative (PID) control, model predictive control (MPC) and fuzzy logic [1]. Both PID and MPC run in systems on the market as of October 2026. The MiniMed 780G uses PID [3], while Tandem and Insulet use predictive designs [2][6].

PID Algorithm

A PID controller is a feedback loop. It compares the latest glucose reading with a target (the setpoint) and adjusts insulin delivery by looking at the gap from three angles [1]:

• Proportional: how far glucose is from target right now.
• Integral: the accumulated gap between measured and target glucose over time. This term works away small, persistent errors that the proportional term alone would leave behind.
• Derivative: how fast glucose is rising or falling, which gives the controller an early read on where it is heading.

The MiniMed 780G is the PID system most people will meet. A 2021 review lists both it and the earlier 670G as using "PID with insulin feedback" [1], and a 2026 study still describes the 780G's algorithm as PID [3].

MPC Algorithm

MPC starts from a mathematical model of how glucose responds to insulin. The algorithm uses that model to forecast glucose over a set window, called the prediction horizon, and chooses the insulin delivery that keeps the forecast closest to target [1]. Engineers who favor MPC point to two practical strengths. Limits, such as a maximum delivery rate or a cap on insulin on board (insulin still active from earlier doses), can be written directly into the calculation, and the framework makes it relatively easy to include the expected effects of meals and exercise [4].

Tandem and Insulet both use predictive control. Tandem's Control-IQ, which a 2026 study describes as an MPC algorithm that forecasts glucose up to half an hour ahead [3], was succeeded in 2025 by Control-IQ+. Tandem says Control-IQ+ adjusts insulin every 5 minutes based on predicted glucose values, and ships it on both of its pumps, the t:slim X2 and Tandem Mobi [2]. Insulet tested a personalized MPC algorithm in research versions of its Omnipod hybrid closed-loop system [5]. The Omnipod 5 that followed runs its algorithm on the Pod itself and adjusts insulin every 5 minutes based on current and projected glucose values [6].

Comparison of PID and MPC

• Effectiveness: The evidence doesn't crown a winner. A 2013 meta-analysis of published studies, which looked mostly at the breakfast response, found the two approaches had performed about the same, though its author noted that differences between studies limited that conclusion [8]. The clearest head-to-head test we found came three years later: a randomized crossover trial in 20 adults with type 1 diabetes, where a personalized MPC kept glucose in the target range 74.4% of the time during supervised 27.5-hour sessions, against 63.7% for PID, with no significant difference in time spent low [7]. Those were research algorithms, not the systems sold today. A 2026 pilot study in 20 adolescents compared two commercial systems and found the opposite edge. With breakfast announced, the MiniMed 780G (PID) and the t:slim X2 with Control-IQ (MPC) performed comparably; with breakfast unannounced, the PID system had a smaller two-hour rise and a lower peak [3].
• Technology: The old objection that MPC needs too much computing power for a wearable device no longer holds. The Omnipod 5 runs its algorithm on the Pod, a patch pump worn on the body, so the phone app doesn't need to stay nearby for automated delivery to continue [6]. Control-IQ+ runs on Tandem's t:slim X2 and Mobi pumps [2]. What matters more is implementation: neither MPC nor PID is a single algorithm, and two engineering teams can build either one very differently [4].
• Pitfalls and Warnings: Both approaches face the same physical limits. Insulin delivered under the skin takes time to absorb, which is why after-meal glucose rises remain a challenge for closed-loop systems, and every algorithm has to work around the limits of CGM accuracy [1]. MPC's model can build in the expected effect of a meal [4], but the 2026 adolescent study is a reminder that a predictive design doesn't automatically respond faster when a meal goes unannounced [3].

Conclusion

For someone comparing systems, the control method comes bundled with the pump; nobody picks PID or MPC on its own. The studies we found point in different directions: MPC came out ahead in a 2016 research trial [7], while the PID-based MiniMed 780G handled unannounced breakfasts better in a small 2026 study [3]. In their own trials, both the 780G and Control-IQ raised time in range over the therapy they were compared with [1]. Which system fits is a decision to make with your care team. Our article on Automated Insulin Delivery (AID) Algorithms covers the other control approaches.