Closed-Loop PID Algorithm Shows Promise for Automatic Blood Pressure Control in Critical Care

Controlling blood pressure in critically ill patients is one of the most demanding tasks in modern medicine, and a team of researchers from Duke University and North Carolina State University believes that a well-established tool from controls engineering could do it better than humans. In a new proof-of-concept study published in Frontiers in Medical Technology, the investigators describe how a closed-loop, automatic control system built around a proportional-integral-derivative (PID) algorithm successfully regulated mean arterial pressure (MAP) across computer simulations, a physical benchtop model, and live anesthetized swine undergoing surgery. The work represents an early but significant step toward replacing the current clinical standard, in which human operators manually intervene between a patient monitor and an infusion pump, a process that introduces delay, potential error, and the risk of over- or undershooting therapeutic goals.

The clinical stakes are considerable. Guidelines for hypertensive emergencies vary widely depending on the underlying condition, with targets ranging from a 20 to 25 percent reduction in blood pressure within the first hour for malignant hypertension to a systolic pressure below 120 mmHg within 20 minutes for aortic dissection. Even in closely monitored trial settings, a minority of patients meet these goals. In the INTERACT2 trial of blood pressure control in hemorrhagic stroke, only 33.4 percent of patients in the intensive treatment group reached the target systolic blood pressure of less than 140 mmHg at one hour. Excessive or overly rapid reduction is also common and dangerous. In the STAT registry, hypotension requiring cessation of intravenous antihypertensive therapy occurred in 6 percent of patients with neurological emergencies and was associated with 3.5-fold mortality. Volatility is equally problematic, with 60 percent of patients in a United States registry of acute severe hypertension experiencing re-elevation after initial control and 4 percent developing iatrogenic hypotension requiring treatment.

The research team, led by Raghav Sriram, Bijan Abar, and Joshua Seth Broder, turned to controls engineering, a field that specializes in closed-loop feedback to achieve specified targets, such as thermostatic control, aircraft autopilot systems, and automatic insulin delivery. All three models in the study employed a PID algorithm to control MAP. The PID controller works by calculating an error signal, the difference between the measured MAP and the target set point, and then summing three components to determine the final output, which corresponds to the infusion rate of fluids or vasoactive drugs. The proportional (P) term is directly related to the size of the error, producing a larger output when the current MAP is far from the target. The integral (I) term reflects accumulated error over time, allowing the controller to increase output if the patient is less responsive to vasoactive drugs. The derivative (D) term is related to the speed at which the system is approaching or moving away from the target, helping to prevent overshoot or undershoot. Each term is weighted by a tunable constant, and the balance between speed, overshoot, and stability must be determined for each application.

In the computer model, the team simulated an automatically controlled drug infusion to regulate MAP dynamically, treating hypotension, hypertension, or maintaining normotension. The simulation modeled hemodynamic responses to a hypothetical drug infusion with pharmacodynamic properties including onset delay and decay. The system was tested in 2,000 simulations with each of five initial MAP values (50, 60, 75, 85, and 95 mmHg), totaling 10,000 simulations, and again with the imposition of hypertensive and hypotensive challenges. The results were striking: the computer model rapidly reached and maintained target MAP within 5 mmHg, with an on-target percentage (OTP) of 90 percent at 240 simulated seconds and 95 percent at 294 simulated seconds. The entire set of 10,000 regular simulations was completed in just 22.57 real minutes, demonstrating the efficiency of the platform for rapid iteration.

The benchtop model added physical components to test hardware and software integration before animal testing. A 100-milliliter saline bag represented the patient’s circulatory system, with its measured hydrostatic pressure serving as a proxy for MAP. Two flexible catheters were introduced to the saline bag, one to measure pressure through an arterial pressure monitoring catheter connected to a pressure transducer and medical monitor, and one to add or remove volume through a custom-built syringe pump connected to a microcontroller. The medical monitor transmitted pressure values once every 5 seconds to the microcontroller, which adjusted the syringe pump based on the PID algorithm’s output. The system was tested across five initial pressures (60, 85, 110, 135, and 160 mmHg) with a target of 110 mmHg, using P, PI, and PID controllers across 45 trials. The benchtop system reached and maintained target MAP with a mean rise time of 0.47 minutes, a mean settling time of 0.62 minutes, and a mean OTP of 91.25 percent, demonstrating that the algorithm could perform well with real equipment.

In the animal model, the team applied the previously tested control system to anesthetized swine undergoing non-survival surgical procedures for physician training purposes. The intervention drug was either norepinephrine to raise MAP or clevidipine to lower MAP, both administered intravenously. Norepinephrine is a catecholamine vasopressor commonly used for shock treatment, with a short mean half-life of approximately 2.4 minutes. Clevidipine is an intravenous dihydropyridine calcium channel blocker with rapid onset and offset, making both agents well suited for an automatic control system. In the first swine, automatic norepinephrine augmented MAP by more than 15 mmHg with a mean rise time of 2.25 minutes and a mean OTP of 56.9 percent. No trial maintained an equilibrium value of target MAP within 5 mmHg, so no settling time could be calculated. In the second swine, clevidipine decreased MAP by more than 15 mmHg with a mean rise time of 1.00 minute, a mean settling time of 2.19 minutes, and a mean OTP of 93.8 percent. The team also demonstrated stepped incremental targets to control the rate of pressure change, achieving a mean OTP of 85.85 percent in those trials.

The investigators acknowledge important limitations. The animal results were confounded by concurrent non-survival surgery, and the first swine likely had hemorrhagic shock, which would be best managed with blood transfusion rather than vasopressor support. The sequential, non-independent design of the animal trials meant that statistical inference and comparison between trials could not be made, and quantities were descriptive only. Only two swine specimens were used, and the team was unable to repeat experimental cycles under identical conditions. The study was not powered to determine which permutation of PID models and starting conditions was optimal, and the team did not systematically investigate the optimal P, I, and D terms for any of their models. Despite these constraints, the authors argue that the swine model demonstrated that the automatic control system was technically feasible under dynamic, clinically realistic conditions.

Beyond the technical performance, the researchers emphasize that the system entered no drug-specific information, suggesting that it may be robust enough to modulate blood pressure across a wide range of initial conditions and clinical scenarios. Varying the PID constants could allow even more accurate control for specific medical conditions, medications, and patient factors. The desired time to target blood pressure may differ substantially between hypertensive emergencies, where rapid correction is associated with adverse outcomes, and hypotensive emergencies, where immediate correction to a minimum perfusing pressure is essential. The team also notes that a PID controller allowing accurate, precise, and stable blood pressure regulation with customizable time-to-target would be an invaluable research tool for testing therapeutic interventions and establishing best targets, addressing a limitation of past research that has been constrained by the inability to reach and maintain blood pressure targets reliably.

Before such a system could be safely deployed in clinical practice, the authors outline extensive regulatory and engineering safety considerations. The proposed system would be defined as Software in a Medical Device (SiMD) under current United States Food and Drug Administration regulations, and manufacturers could choose to integrate closed-loop automation features fully or allow third-party add-on integration. Potential failure modes include sensor errors from dampened arterial signals, miscalibrated transducers, and pump malfunction, all of which already apply to existing human-guided blood pressure control. Cybersecurity is another critical concern, as a programmable medication infusion pump could be hacked remotely or locally, resulting in potentially fatal over- or undertreatment. The authors propose safeguards including redundancy, automatic reversion to a safe mode with fixed infusion rate limits, and clear alerts for non-nominal conditions. Human oversight would remain essential, with providers able to override the automated controller in cases of frank system failure or failure to achieve programmed blood pressure goals.

The most obvious initial clinical settings for deployment would be emergency departments, intensive care units, and operating theaters, substituting for existing human-dependent blood pressure control. Once properly validated, the range of possible use settings could expand to include prehospital and interhospital critical care transport, military installations, austere locations on Earth, and even space missions beyond low Earth orbit, where future crews may require on-board care with scant personnel. The intended operators would initially include the same critical care personnel currently managing blood pressure in acute care patients, with the user base potentially expanding to include paramedics and others as systems are validated. The authors point to closed-loop insulin delivery systems as a precedent, noting that these are associated with higher on-target percentages than open-loop systems and lower rates of hypoglycemic coma. Future studies can optimize the system and compare its performance with human operators in advance of potential human clinical trials, marking the next step in a progression that could ultimately transform how blood pressure is managed in the sickest patients.

Subject of Research: A Duke University-led team reports proof-of-concept evidence that a closed-loop PID control algorithm can automatically regulate mean arterial pressure across computer, benchtop, and anesthetized swine models, with future work needed to optimize the system and compare it with human operators before human trials.

Article Title: Automatic control of arterial blood pressure: computer, benchtop, and animal models

Article References: Sriram, R., Abar, B., Gullapalli, S., Hu, Y., Patel, A., & Broder, J. S. (2026). Automatic control of arterial blood pressure: computer, benchtop, and animal models. Frontiers in Medical Technology, 8, Article 1955568. https://doi.org/10.3389/fmedt.2026.1955568

Image Credits: AI Generated

DOI: 10.3389/fmedt.2026.1955568

Keywords: closed-loop control, PID algorithm, blood pressure, mean arterial pressure, critical care, norepinephrine, clevidipine, vasoactive drugs, medical devices, emergency medicine, biomedical engineering, proof of concept