A pump that delivers what it was programmed to deliver, vs. a system that watches the body and decides what to deliver next.
This is fundamentally a control-systems distinction applied to drug dosing, and it maps directly onto the classic open-loop vs. closed-loop control taxonomy. An open-loop drug delivery system — which describes most infusion pumps in clinical use — delivers a fixed, pre-programmed dosing schedule set by a clinician, with no sensor feedback loop adjusting that schedule based on the patient's real-time physiological state. A closed-loop system senses an actual physiological signal — continuously measured glucose, for instance — and an algorithm adjusts dosing in real time based on that measurement, the way an artificial pancreas / automated insulin delivery (AID) system continuously modulates insulin infusion based on continuous glucose monitor (CGM) readings rather than a fixed basal rate alone.
An open-loop system's control logic is simple by design: a programmed infusion rate (or a rate curve, in more sophisticated pumps) executes on schedule, and the only feedback in the system is the clinician periodically checking on the patient and manually reprogramming the pump if the clinical picture changes. There's no sensor measuring the physiological effect of the drug feeding back into the pump's dosing decision in real time. A closed-loop system adds three things an open-loop system doesn't have: a sensor continuously measuring a relevant physiological variable (a CGM measuring interstitial glucose), a control algorithm (commonly a form of model-predictive control or a PID-family controller in AID systems) that translates that measurement into a dosing decision, and an actuator (the insulin pump) that executes the algorithm's decision — all operating with minimal or no clinician intervention between measurement and dose adjustment.
Open-loop delivery is not simply a worse, older version of closed-loop — for many drugs and clinical contexts, a fixed schedule is exactly appropriate, safer, and sufficient, particularly where the therapeutic target doesn't fluctuate rapidly enough to need real-time adjustment, or where a continuously reliable real-time sensor for the relevant variable doesn't exist. Closed-loop systems earn their added complexity specifically where the physiological variable being managed (blood glucose) genuinely fluctuates unpredictably and rapidly enough that a fixed schedule would systematically over- or under-dose the patient across the day. But that responsiveness comes at a real engineering cost: the entire system's safety is now only as good as the sensor's accuracy, and a sensor fault, dropout, or calibration drift can propagate directly into an inappropriate dosing decision — which is exactly why AID systems build in extensive sensor-fault detection, dosing limits, and fail-safe fallback-to-manual behavior as core safety architecture, not an afterthought.
The entire value proposition of closed-loop delivery — dosing that tracks the patient's real physiological state instead of a static assumption — depends completely on the sensor providing an accurate, timely, low-latency measurement of that state. A continuous glucose monitor measuring interstitial rather than blood glucose introduces a genuine physiological lag the control algorithm has to account for, and any sensor noise, dropout, or drift directly corrupts the input the entire dosing decision is built on. This is why closed-loop system design spends as much engineering effort on sensor validation, fault detection, and algorithmic robustness to noisy input as it does on the control algorithm itself — a perfect algorithm fed a bad sensor reading still produces a bad dose.
Closed-loop is more capable for the specific problem it's built for — dosing a drug against a rapidly fluctuating physiological variable — but it isn't a strict upgrade for every drug delivery scenario. It requires a reliable real-time sensor for the target variable, adds real failure modes (sensor error propagating into dosing error) that open-loop systems simply don't have, and adds substantial algorithmic and regulatory complexity. Most drugs delivered by infusion pump today are appropriately open-loop precisely because no such sensor-to-dose feedback relationship is clinically necessary or technically available for that drug.
Explains why most infusion pumps operate open-loop — delivering a fixed, pre-programmed dosing schedule regardless of the patient's real-time physiological state — while closed-loop systems, such as artificial pancreas / automated insulin delivery systems, continuously sense a physiological signal like glucose and adjust dosing algorithmically in real time. Covers why closed-loop's added responsiveness comes with real sensor-dependency risk that open-loop systems don't have.
It's easy to frame closed-loop delivery as simply the more modern, more sophisticated option that open-loop systems are working their way toward. That framing misses that the two architectures solve different problems. Open-loop is the right architecture whenever a fixed dosing schedule adequately serves the therapeutic goal, and it avoids an entire category of sensor-dependent failure modes closed-loop systems have to design around. Closed-loop earns its complexity specifically for drugs managing a physiological variable, like glucose, that fluctuates unpredictably enough that a static schedule would systematically mis-dose the patient.
Open-loop infusion pumps are regulated primarily around dosing accuracy, occlusion/free-flow detection, and user-interface safety (preventing programming errors) — the control logic itself is simple and doesn't need to account for sensor-driven feedback dynamics. Closed-loop / automated insulin delivery systems add substantially more regulatory scrutiny: the control algorithm (often model-predictive control, sometimes with machine-learning components) has to be validated across a wide range of physiological scenarios, the CGM sensor's accuracy and failure modes have to be characterized, and the system needs demonstrated safety behavior for sensor dropout, signal noise, and communication failure between sensor and pump — FDA's interoperable automated insulin dosing (iAID) framework specifically addresses these multi-component, closed-loop system requirements.
The decision to build open-loop or closed-loop should follow directly from whether the target physiological variable fluctuates rapidly enough to need real-time correction, and whether a sufficiently accurate, low-latency sensor for that variable actually exists. Building unnecessary closed-loop complexity into a system that doesn't need it adds sensor-dependency risk and regulatory burden without a corresponding clinical benefit. Conversely, sticking with open-loop for a variable that genuinely needs real-time correction — as fixed-schedule insulin dosing historically did for many patients — leaves clinically meaningful improvement in glycemic control on the table, which is precisely the gap automated insulin delivery systems were engineered to close.
Hybrid closed-loop describes real-world automated insulin delivery systems that automate basal insulin dosing in a closed loop but still require the user to manually announce meals and administer bolus doses for carbohydrate intake — it's not a fully autonomous closed loop for all dosing decisions. It sits functionally between pure open-loop and a hypothetical fully autonomous closed loop, and it's the dominant real-world implementation of AID technology today because fully automating meal-related dosing without any user input remains a harder control problem.
Well-designed AID systems include sensor-fault detection logic that can recognize signal loss, unrealistic rate-of-change values, or persistent sensor error, and respond by reverting to a safe fallback mode — typically defaulting to a pre-set basal insulin rate similar to open-loop delivery, and alerting the user, rather than continuing to act on unreliable sensor data. This fallback-to-safe-default behavior is a required part of the system's safety architecture, not an optional feature.
In principle yes, and it's an active area of research — closed-loop delivery has been explored for propofol sedation depth (using processed EEG as the feedback signal) and for certain vasoactive drugs in critical care, among others. The limiting factor in each case is the same one that makes insulin AID viable: a sufficiently accurate, low-latency, continuously available real-time sensor for the target physiological variable has to exist before closed-loop control is a realistic option.
Try our Biomedical Engineering
More calculators, simulators, and guides for this discipline.