Vasopressor Control Module Development and Integration into the Hardware-in-Loop Automated Testbed for Resuscitation Controller Evaluation

doi:10.21203/rs.3.rs-7819342/v1

Download PDF

Research Article

Vasopressor Control Module Development and Integration into the Hardware-in-Loop Automated Testbed for Resuscitation Controller Evaluation

https://doi.org/10.21203/rs.3.rs-7819342/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Introduction:

Patients in hemorrhagic shock who are unresponsive to fluid administration can urgently require vasopressor therapy. However, manual vasopressor titration in mass casualty or resource-limited settings can be arduous, increasing the risk of under- or over-treatment. Closed-loop vasopressor adaptive resuscitation controllers (V-ARCs) offer a potential solution, but their development is hindered by the lack of realistic and adaptable laboratory test platforms. To solve this problem, we developed a Vasopressor Control Module (VCM) for a hardware-in-loop automated testbed for resuscitation controllers (HATRC) to enable the systematic evaluation of V-ARCs.

Methods and Results

The VCM’s design was informed by vasopressor administration data captured in a hemorrhagic shock swine model, which revealed four key physiological variables that defined the hemodynamic response to vasopressor infusion: lag time, real response, overshoot, and pressure-time responsiveness. The incorporation of these variables enabled the VCM to replicate physiological variability, dose-dependent responsiveness, and noise conditions representative of the clinical setting. The platform was validated by preliminarily comparing multiple V-ARC designs under different testing conditions and successfully differentiating their performance.

Conclusions

This work establishes a flexible, physiologically grounded platform for vasopressor controller development, reducing dependence on animal testing while enabling rapid and robust controller evaluation. Future work will expand physiological modeling, incorporate additional hemodynamic variables, and support multiagent resuscitation.

Closed-loop controllers

vasopressor therapy

hardware-in-loop

automation

hemorrhagic shock

fluid resuscitation

product testing

Successful treatment of hemorrhagic shock depends on achieving a few key objectives: (i) rapid control of the source of the bleed, (ii) fluid volume replenishment, and (iii) adequate perfusion of critical organ systems (1, 2). The current standard of treatment relies on the administration of large volumes of fluids. However, some patients are non-volume responsive and exhibit a progressive physiological decline despite aggressive fluid infusion (3). For those severely hypotensive trauma patients, early adjunctive administration of vasopressors (4–6) may provide hemodynamic support when blood products are not available while also mitigating the potential for fluid overload.

The two vasopressors recommended for patients with hemorrhagic shock are norepinephrine (NE) and vasopressin (7); NE primarily works via vasoconstriction, inotropy, and chronotropy, while vasopressin works as a vasoconstrictor at the level of the blood vessels promoting salt and water retention and restoring blood flow to the kidneys (8). Vasopressors are powerful drugs that require extremely close management by medical providers who are skilled in administering them. Under resource constrained conditions and/or during mass-casualty (MASCAL) events, close attention from skilled staff may not be available, meaning that patients who require vasopressor therapy may not have access to the lifesaving care they need.

In battlefield or austere environments with limited monitoring capabilities, simply restoring blood pressure with vasopressors may obscure ongoing organ hypoperfusion, exacerbating tissue ischemia. Even in clinical trauma settings, where patients can be closely monitored, imprecise titration can lead to excessive vasopressor exposure causing deleterious effects on renal and cardiac function. So, although common, the use of vasopressors is controversial and it is correlated with poor outcomes and increased mortality in trauma patients if not correctly administered and monitored (9, 10).

To solve this problem, we propose the development of vasopressor adaptive resuscitation controllers (V-ARCs), a subtype of physiological closed-loop controllers (PCLC) that balance the infusion of vasopressors and resuscitation fluids to maintain a patient at a desired blood pressure. The development of V-ARCs will require iterative testing to ensure that they are both safe and effective, and while preclinical animal research is the gold standard for demonstrating safety and efficacy of a therapeutic intervention prior to use in humans, animal testing is expensive, time consuming, and ethically complex. By comparison, benchtop models for in vitro research are inexpensive, allow rapid prototyping, and have no ethical issues to navigate. Additionally, benchtop optimization of PCLC performance prior to deployment in an animal model ensures that the smallest number of animals is required to successfully demonstrate safety, efficacy, and/or proof-of-concept. Non-animal models for studying automated vasopressor administration include in silico models that use computer and mathematical analysis that simulates cardiovascular physiology (11–14). One of the most advanced embodiments of this technology is the use of digital twins, wherein a model is trained using high-fidelity clinical data to simulate an individual patient (15, 16). The trained model can then be supplied with treatment data to predict how the patient will respond. However, these models require extensive computing resources to run effectively, are highly dependent on both the quality and quantity of the training data, face generalizability challenges, and are extremely sensitive to initial conditions making it difficult to simulate specific trauma scenarios if adequate patient data matching those scenarios are lacking (17–19).

Here, we set out to develop an extension to an existing hardware-in-loop automated testbed for resuscitation controllers (HATRC) (20) that would allow it to simulate a patient’s response to vasoactive medications. Briefly, the HATRC system provides a bulk model for hemodynamic physiology that can simulate the mean arterial pressure (MAP) response to whole blood and/or crystalloid infusion (21, 22). The platform offers extensive customizability, such as specifying when a patient becomes volume non-responsive, while remaining highly repeatable for comparing resuscitation controller designs, something currently unavailable with typical in silico models (23). This extension, called the Vasopressor Control Module (VCM), aims to recreate the physiological response to vasopressor infusion in a benchtop model, thus obviating the need for extensive use of animals to develop and test V-ARCs.

The development and testing of the VCM module is presented first with the analysis of pressure data collected during vasopressor administration in a swine model of hemorrhagic shock, and a selection of key physiological response variables are introduced. Next, is a description of the VCM, its characterization and validation, along with a similar overview of a mechanical noise generator module (NGM). Lastly, proof-of-concept testing of two decision table-driven V-ARCs are presented and quantitatively analyzed using controller performance metrics to differentiate their performance.

Animal Data Analysis

Development of the VCM started with an analysis of swine data with the goal of characterizing the hemodynamic changes that occur from vasopressor administration during hemorrhagic shock (see methods). A pump-controlled hemorrhage was used to reach a MAP of 35 mmHg, followed by titration of NE until MAP was restored to at least 65 mmHg, then a single weaning step was attempted at the end to observe the hemodynamic effects of a reduction in vasopressor infusion rate.

The transient MAP signal during vasopressor administration was analyzed to extract key fiducial waveform markers as well as four derived metrics to characterize the pressure and time responsiveness to the vasopressor therapy. Key markers included the pressure at the moment the dosage change occurs (P_t0), the pressure when a noticeable response is detected (P_Start), a “point of inflection” where the first derivative of the pressure signal drops sufficiently below an observed maximum (POI), the pressure response vertex (P_V), and the relative stable pressure (P_Stab) following each dosage. A representative subject waveform is shown in Figure 1 with the points of interest marked and each marker is summarized in methods Table 2. From this initial broad array, we identified four essential physiological metrics that could be used to describe any given animal’s hemodynamic response to an increase in the rate of NE infusion: (i) lag time, (ii) real response, (iii) overshoot and (iv) pressure-time responsiveness (dMAP/dt). An illustrative diagram of the derived features can be seen in Figure 10 in the methods as well as detailed description of each metric in methods Table 3.

To aid in the design of VCM, two critical questions to answer were (i) whether any of these extracted variables were a function of vasopressor concentration or (ii) whether any variable was significantly altered by inter-subject variability. Statistical analyses across the entire dataset were performed for each metric to answer these questions. Overall, we found that vasopressor concentration only significantly impacted pressure-time responsiveness, while the other three variables were not significantly impacted by dosing (Figure 2A-D). Conversely, inter-subject variability had a larger effect on three of the variables, with only lag time not being significantly impacted (Figure 2E-H). This analysis allowed us to tailor the noise and variability of the test platform response and specify the conditions affecting it (e.g., subject variability or real-time dosage dependence).

In addition, we assessed the impact of increasing or decreasing vasopressor dosage on each of these response variables as the VCM can potentially be configured for testing vasopressor weaning off response (Figure 3). Lag Time was the only variable with a significant difference; the other three metrics were not significant, indicating there was no dependence on whether dosage was increasing or decreasing and enabling symmetric bi-directional response profiles to be used. It should be noted, however, that decreasing dosage steps were taken less frequently (only 6/7 subjects received a single dose reduction step and one dataset was excluded due to missing data), so only NE concentrations with balanced increasing and decreasing dosing directions were used in this analysis, excluding two additional subjects and resulting in an overall low sample size (n = 3). Thus, more data are needed to further refine this analysis.

Vasopressor Control Module

The VCM comprises a microcontroller, stepper motor, and needle valve which is plumbed into the fluid circulating portion of the HATRC loop. Flow is occluded as the valve is tightened, increasing the flow resistance and raising the pressure (see Methods subsection “Modifications of the Test Platform” for more information). The response variables were used when programming the microcontroller to achieve the desired response. To convert the administered dosage to a pressure delta, we first characterized the Real Response of the animal data using logarithmic regression as shown in Figure 4A. To improve the accuracy of the characterization, we transformed the data by taking the natural log of the dosages followed by fitting the data with a quadratic regression shown in Figure 4B. The resulting R² value of the quadratic fit on the transformed data increased to above 0.99. This function was used to predict the net pressure change induced by any given dosage administered by the V-ARC.

Since the VCM is operated via stepper motor, this offers a convenient characterization of the VCM by defining the Delta Pressure in terms of motor step position. This was achieved using a piecewise set of quadratic regression functions shown in Figure 5. A range of 1700 microsteps was found to cover a sufficient pressure window with a baseline position having no pressure effect and an observed maximum pressure delta of approximately 83 mmHg. Four pressure regions were defined including Low (0-3 mmHg), Mid (3-12 mmHg), High (12-35 mmHg), and Very High (35-83 mmHg). Each regression had R²values of 0.99 or greater except for the Low region, which was 0.8. This was due to there being no significant change in pressure until ~300 steps at which point the more non-linear response began. A more refined baseline position can be defined in future iterations to potentially improve accuracy in this low delta pressure region, though the overall impact would likely be small.

These functions describing the step—delta pressure relationship provided a means to precisely control the Real Response of simulated vasopressor dosing but are insufficient for controlling the dynamic dMAP/dt behavior. However, solving each function for Delta Pressure and differentiating with respect to step position enabled the prediction of the step rate change of pressure based on the current step position (i.e., dP/dStep). Based on a desired dMAP/dt, the real-time result of dP/dStep was used to iteratively solve for the required motor step rate (i.e., dStep/dt). The system was validated across a range of vasopressor doses (2.5, 5, and 7 mL/min) at two dMAP/dt rates (8.368 and 4.0 mmHg/min). The time-based pressure results are shown in Figure 6. Linear regressions of the pressure data had R² values all above 0.99 demonstrating the system’s ability to generate a linear pressure response with respect to time, and the average pressure responses were accurate within 94% and 97 % for 4.0 mmHg/min and 8.368 mmHg/min, respectively.

Noise Generator Module

The NGM was designed using a microcontroller, linear stepper motor, and a pinch valve. The stepper motor drove a lead screw which extended and contracted to actuate the valve, pinching the soft tubing of the fluid circulating portion of the HATRC loop to cause pressure perturbations (see Methods subsection “Modifications of the Test Platform” for more information). We characterized the NGM in a similar fashion to the VCM (Figure 7), however the different motor and valve configuration had a lower resolution of both actuation and pressure. We used piecewise quadratic regressions to define the Step—Delta Pressure relationship across three pressure regions, Low (0-2 mmHg), Mid (2-11 mmHg), and High (11-60 mmHg). The total step range was limited to 45 microsteps and the maximum delta pressure observed was 60 mmHg. All the regressions had R² values over 0.99. Using these functions, the module was given desired pressure deltas to perturb the system and actuated the motor to the necessary position.

Controller Proof-of-Concept

The pressure vs time plots for the two controllers evaluated are shown in Figure 8. Test runs were conducted with and without the inclusion of pressure noise as well as with and without the assistance of an adaptive resuscitation controller (ARC) running. Each scenario was run in triplicate, and the 3-run averages are shown in the figure. Overall, the effect of noise is visually evident and the aggressive decision table (AggTable) had a visually faster response time compared to the more conservative decision table (ConTable) configuration (Figure 8).

Four typical controller metrics were calculated to determine if statistically significant differences between specific controllers were identified (Figure 9). Median performance error (MDPE, Figure 9A) was most positive for AggTable, ARC-On, Noise-On at a value of 2.01% while the lowest MDPE was the ConTable, ARC-Off, Noise-Off at -3.28%. with a number of controller configurations being significantly different from other controller configurations. For rise time efficiency (Figure 9B), the ConTable was consistently slower vs. Agg table (~4.6 min vs ~3.1 min, lower is better). The AggTable also performed better for the controller effectiveness metric (78% vs ~65%, higher is better, Figure 9C). Lastly, we quantified a target overshoot metric (Figure 9D) which tended to be larger for the AggTable (2.79% vs 1.82% ConTable), ARC-On (3.17% vs 1.44% ARC-Off), and Noise-On (3.97% vs 0.63% Noise-Off). The VCM module was capable of differentiating controller performance, highlighting its utility as a development platform for controller troubleshooting.

Hemorrhagic shock remains a leading cause of trauma-related death, which can often be prevented with restoration of blood pressure via fluid resuscitation. In some instances, resuscitation by fluid alone may not be possible due to fluid non-responsiveness or resource constraints limiting access to blood products. Vasopressor adjuvant therapies during fluid resuscitation may help in these instances if their rates can be titrated effectively. This can require careful monitoring of blood pressure, something that may not be practical in trauma care where MASCAL situations can limit personnel availability. As such, we are focused on developing closed-loop vasopressor administration controllers for trauma care, which require large datasets and extensive troubleshooting to be safe and effective. We previously developed closed-loop controllers for both whole blood and crystalloid infusion using a hardware-in-loop test platform for iterative testing and optimization prior to successful animal testing. Here, we explored the development of an adaptation to the hardware-in-loop system enabling its use for vasopressor controller development.

Analyses of previous animal data revealed a set of variables that could be used to characterize the physiological pressure response to a range of NE doses. The VCM was then designed to mimic the characteristics of this response simulating a patient receiving NE. The magnitude and rate of changes in MAP were of particular interest, and we validated the VCM’s ability to reproduce these features in a way that also allows customization by the user. We also implemented a module that mechanically added noise to the pressure signal improving realism. Finally, we conducted proof-of-concept comparisons between two similar vasopressor controllers under non-noisy and noisy conditions as well as with an assistive fluid resuscitation controller. Evaluating the controllers using standard metrics confirmed the platform’s usability for identifying potentially nuanced differences in performance between novel V-ARCs.

In addition to the waveform-specific metrics, we identified whether they were significantly impacted by inter-subject variability, dosage, as well as dosing direction. The pressure response rate, dMAP/dt, was the only variable found to be significantly affected by both the dosage rate and inter-subject variability, while the Real Response and Overshoot were only statistically different with respect to inter-subject variability. Lag Time did not demonstrate a significant effect from either inter-subject variability or dosage, but it was the only variable that showed a significant difference with dosing direction. This highlights the complexity of physiologic modeling as well as the need for varied and sometimes tailored testing.

The primary use for the VCM module is the systematic evaluation of closed-loop vasopressor administration and resuscitation. Proof of concept controller testing highlighted this utility by identifying feature and controller configuration differences with regard to controller performance. The more aggressive DT logic was significantly different from all conservative DT logics for rise time efficiency and controller effectiveness, highlighting how the aggressive logic reached target pressure more quickly and held at target for longer. Similarly, the pairing of V-ARC with ARC consistently led to target overshoot but had less effect on other performance metrics. These differences can be used to further tune controller logic to more effectively pair V-ARC and ARC together or to create less aggressive DT logic that reaches target pressure quickly while reducing overshoot effects. These results were extracted from a single scenario setup; with more scenario development to include different hemorrhage rates, varying starting pressures, as well as alternate target pressures, more controller performance trends can be extracted in a high-throughput manner prior to more costly, resource intensive animal testing.

Although the system was able to achieve its primary aim in distinguishing between vasopressor controllers, there is still room for improvement. The datasets used for the initial responsiveness characterization were relatively limited. More robust datasets would permit a wider range of patient conditions to be modeled and improve the accuracy of the predicted pressure response. For example, the current dataset does not provide information on how a severely hypotensive patient might respond to a high initial dose of vasopressors or how the response might change with the presence of certain comorbidities. Thus, the need for animal testing is reduced, but not eliminated. There are also improvements that can be made to the VCM. The Overshoot variable was not modeled in this iteration of the module. As this was often a self-resolving feature that only lasted briefly, the more steady-state behavior is still reliable. However, the overshoot and median performance error metrics may score differently, particularly for controllers that utilize less frequent, but larger increases in NE dosage. Some features like lag time and dMAP/dt were held constant for the proof-of-concept testing but can be improved by adding a factor of inter-subject randomness to the value(s) including logic that considers the dosage- and direction-dependent deviations from the mean.

This work demonstrated the successful development and integration of a vasopressor control module into a hardware-in-loop resuscitation testbed. The platform was physiologically informed by animal datasets and proved to be effective at differentiating controller performance differences between two controller designs. In the broader context of trauma care, this work contributes to solving a pressing operational challenge—safe and effective vasopressor titration in resource-limited or MASCAL settings where continuous manual monitoring is impractical or impossible. The ability to rigorously test and refine autonomous control strategies in a controlled, high-fidelity environment supports the development of technologies that can extend critical care capabilities to austere and high-demand scenarios. Future work will incorporate additional hemodynamic parameters and explore multi-agent resuscitation strategies, as well as develop a robust set of simulated trauma scenarios for down-selecting an assortment of V-ARC designs prior to animal testing. Ultimately, this approach can lead towards safer, faster, and more effective deployment of closed-loop vasopressor control in both military and civilian trauma systems.

This section is structured as follows to describe the entire research study: description of animal data capture, preprocessing and data analysis, a short description of the existing test platform, VCM development and integration into the platform for vasopressor functionality, and, lastly, an overview of the proof-of-concept testing and the V-ARCs used.

Animal Data Capture and Analysis

This research was conducted in compliance with the Animal Welfare Act, implementing Animal Welfare regulations and the principles of “The Guide for the Care and Use for Laboratory Animals” (24). The protocol (A-24-003) received approval from the Institutional Animal Care and Use Committee (IACUC) at the United States Army Institute of Surgical Research where the research was conducted. The facility is fully accredited by AAALAC International. The experimental design utilized a refined swine model (Sus scrofa domestica) previously validated for similar applications (25). Ten intact female Yorkshire crossbred swine, approximately four months old and weighing around 40 kg, were utilized. Swine were selected due to their well-documented physiological resemblance to humans, particularly in cardiovascular function (26, 27).

Throughout the study, subjects were administered Buprenorphine SR (0.24 mg/kg) for analgesia. All subjects were maintained at a surgical plane of anesthesia and analgesia, beginning with intramuscular tiletamine-zolazepam (4–6 mg/kg) for tranquilization. Isoflurane (0–5%) was used for initial inhaled anesthesia before transitioning to total intravenous anesthesia (TIVA), employing ketamine (0–10 mg/kg) and midazolam (0–2 mg/kg). Anesthetic agents were titrated to effect, and animals remained supine for the duration of the experiment.

As described previously (25), following initial instrumentation, subjects underwent a surgical splenectomy followed by an initial controlled hemorrhage and resuscitation event using whole blood and ARC to recover animals to a target MAP of 65 mmHg. After a one-hour stabilization, a second controlled hemorrhage to a target MAP of 35 mmHg was performed guided by an automated decision table termed “AutoBleed” (29). Norepinephrine (NE), chosen based on the European trauma guidelines (7), was titrated to evaluate blood pressure responsiveness to vasopressor administration under hemorrhagic conditions. Concurrently, Lactated Ringer’s solution was infused at a fixed rate of 10 mL/min. Norepinephrine dosing followed a predefined escalation protocol (Table 1), with each dose held for a total of up to 20 mins or until the research team determined MAP had stabilized at that step, whichever happened sooner.

Table 1
Stepwise norepinephrine infusion rates and corresponding delivery volumes used for MAP titration
Norepinephrine Dose (mcg/min)	Infusion Rate (mL/min)
0.5	0.13
1	0.25
2	0.5
4	1
8	2
12	3
16	4
20	5

Stepwise increases in vasopressor infusion continued until a sufficient response to vasopressors was reached, typically MAP reaching at least 65 mmHg. Next, the dose was reduced one step to assess weaning off effects. After MAP stabilization at the weaning off step, experimental procedures were concluded. Subjects were humanely euthanized via intravenous administration of sodium pentobarbital (FatalPlus, Vortech; Dearborn, MI), in accordance with the American Veterinary Medical Association guidelines (30).

Preprocessing and Data Analysis

Vasopressor infusion datasets from each animal were processed to look at the effect of vasopressor dosing on MAP. As such, the primary variables extracted for each swine were MAP and total infused vasopressor, each with respect to time (sampled at 1/5 Hz). A moving average was applied to the MAP signal to smooth the overall data trends during vasopressor administration. Data were separated by vasopressor dosage step, generating individual data frames that showed how blood pressure responded to each dose of vasopressors. Within each of these, key fiducial points of interest were identified as follows and further described in Table 2: P_t0, P_Start, POI, P_V, and P_Stab.

Table 2
Summary of key fiducial markers from vasopressor datasets
Extracted Feature	Description
P_t0	Beginning of vasopressor delivery
P_Start	The moment where the data crosses a specific slope threshold, indicating the first real positive (or negative, for down steps) response to the dose
POI	The fifth consecutive data point wherein the instantaneous slope is 50% below the local dMWP/dt maximum
P_V	The vertex, or maximum pressure achieved after P_Start
P_Stab	The stabilization point after PV, indicated by when the slope change of 3 consecutive moving data windows are beneath 0.05. If condition is not met, the slope threshold is increased to 0.1.

Because vasopressor rates could increase or decrease between steps due to the weaning off assessment performed, logic for identifying these fiducial markers was adjusted to accommodate differences in upward and downward trends. Fiducial markers were used to calculate four additional derived metrics based on time and pressure differences between markers: Lag Time, Overshoot, Real Response, and pressure-time responsiveness. Each of these are summarized in Table 3 and diagrammed in Fig. 10 to better visualize their meaning.

Table 3
Summary of derived metrics characterizing vasopressor responsiveness
Metric	Description
Lag Time	An estimation of how long it took to see an actual pressure response from the time vasopressor was first administered at that dose. Calculated as time delta between P_t0 and P_Start.
Overshoot	How high (or low) the pressure reaches relative to the level it ultimately settles at for any given dose. Calculated as the pressure difference between P_V and P_Stab.
Real Response	The total change in pressure from the initial baseline to its stable value. Calculated as the pressure difference between P_t0 and P_Stab.
Pressure-Time Responsiveness (dMAP/dt)	The average slope between starting pressure and as the pressure approaches its vertex point. Calculated as the slope between P_t0 and POI.

Each of these markers and metrics were calculated for each vasopressor up or down step and summarized to quantify overall data trends related to vasopressor delivery during hemorrhagic shock resuscitation. In addition, combined plots of all vasopressor rates for each pig were generated, allowing for comparisons across different dosages within the same animal.

Modifications of the HATRC Platform

The hardware-in-loop testbed for resuscitation controllers was previously designed for evaluating fluid resuscitation controllers, comprised of computer-controlled pumps and a hydrostatic reservoir contoured to mimic different pressure-volume response trends corresponding to different fluid infusate types (21, 22). In this work, we incorporated vasopressor infusion functionality by mimicking data trends observed in swine during vasopressor therapy. The vasopressor control and noise generator module development are described below, along with their communication integration with the greater HATRC test platform.

Vasopressor Control Module

To simulate the use of vasopressors in our HATRC system we aimed to simulate a similar effect that vasopressors produce in the body, i.e., to raise pressure by increasing the flow resistance of the system. Unlike vasopressors, which achieve this effect primarily via systemic vasoconstriction, we implemented a device to variably choke the flow at a single point. The device comprised a NEMA 23 stepper motor coupled with a needle valve which was integrated into the circulating portion of the loop. As the stepper motor tightened the valve, flow was constricted, resulting in increased pressure that was effectively independent of the fluid volume while still permitting volume-responsive pressure changes as needed.

Characterization of the device was conducted by actuating the stepper motor through a designated range of 1700 total steps in one-step increments. A one-second pause was held at each position for the first 1600 steps and was extended to 5 seconds for the final 100 steps. This allowed the pressure response to stabilize at each position with the extended pause accommodating the greater pressure deltas per step that occurred as the flow became more occluded. The baseline position of the needle valve was defined where no measurable impact on system pressure was observed, and the upper limit was set to produce a maximum pressure delta of approximately 87 mmHg. The pressure vs step data were inverted then segmented into four regions which enabled a piecewise set of second-order polynomial regression functions to be defined as

$$\:Step\left({\Delta\:}P\right)=\left\{\begin{array}{c}f{\left({\Delta\:}P\right)}_{1},\:\:0\le\:\varDelta\:P<\varDelta\:{P}_{1}\\\:\dots\:,\:\:\dots\:\\\:f{\left({\Delta\:}P\right)}_{4},\:\:\varDelta\:{P}_{3}\le\:\varDelta\:P<\varDelta\:{P}_{max}\end{array}\right.$$

where $\:Step$ is the estimated stepper motor position required and $\:{\Delta\:}P$ is the desired pressure delta from baseline. The boundaries for each region were determined by optimizing the coefficients of determination of each regression function while minimizing the prediction error at the boundary points.

Since physiological pressure responses are not instantaneous, in addition to simulating the vasopressor-induced pressure changes, we also wanted to simulate the dynamics of the pressure over time by controlling the stepper motor actuation rate. Thus, the vertex forms of the quadratic $\:Step\left({\Delta\:}P\right)$ functions were solved for pressure to obtain $\:{\Delta\:}P\left(Step\right)$ functions, and the first derivatives were solved for each to obtain $\:\frac{dP}{dStep}$ functions for each region. This provided a set of position-dependent functions which could then be used in the following equation:

$$\:\frac{dStep}{dt}=\frac{\frac{dP}{dt}}{\frac{dP}{dStep}}$$

where $\:\frac{dStep}{dt}$ is the necessary actuation rate of the stepper motor to produce a desired pressure response of $\:\frac{dP}{dt}$ based on inputting the current step position into the appropriate $\:\frac{dP}{dStep}$ function.

Noise Generator Module

To generate noise, a NEMA 14 stepper motor-driven linear actuator was employed to introduce ± 7 mmHg pressure fluctuations by pinching the flow-loop tubing. The actuator was first positioned at a median “baseline” step index, then driven above or below this point to manipulate the pressure.

To determine the baseline and MAP step indices to pressure, the actuator was cycled through a range of 45 steps in single-step increments. Each step was held for five seconds for the first 30 steps, and 15 seconds for the final 15 steps. The sequence was then mirrored in reverse order. During this sweep, steady-state pressure was recorded at each step. The resulting pressure vs step relationship was split into three regions, each modeled by a second order polynomial, similar to the VCM. These piecewise regressions take a desired delta pressure (ΔP) as input and output the corresponding step index required to achieve that ΔP from the baseline. The home position step was defined as corresponding to a ΔP of ~ 9 mmHg from the absolute baseline, allowing both positive and negative pressure noise to be effected.

With the characterization complete and the home position defined, a generated sequence of random ΔPs was passed through the piecewise polynomial functions to compute the target step. The actuator was then driven from its home position to the target step, held for one second, when another ΔP value is passed. This method produces controlled, randomized perturbations into the system.

HATRC Platform and Communications

The primary HATRC system runs on a PC in MATLAB (MathWorks, Natick, MA, USA). The VCM and NGM’s microcontrollers were integrated into HATRC using specialized libraries and serial communication protocols. The system pressure is collected using a data acquisition system (PowerLab, ADinstruments, Sydney, Australia) and sent to the HATRC system which provides this as an input to the V-ARC control script. The V-ARC in use determines the vasopressor dosage and sends this as an output to the HATRC system. The desired dosage is then transmitted via serial communications to the VCM’s microcontroller where it is first converted to a predicted delta pressure and the stepper motor is sent to the corresponding position as determined by the characterization functions. All modules operate with a 5-second sampling period for compatibility with the HATRC system.

Vasopressor Controller Proof-of-Concept Testing

To better understand the testing procedure, each controller logic is first introduced, followed by how proof-of-concept testing was configured for the VCM module.

ARC Overview

For automated fluid administration, we utilized a previously developed adaptive resuscitation controller, or ARC which has been successfully validated in a large animal model of hemorrhagic shock (25). Briefly, ARC is driven by MAP as its primary input variable with a target MAP value set as the goal for directing fluid therapy. An initial bolus of 100 mL of fluid is administered over one min to estimate pressure-fluid responsiveness which is fitted with a linear regression model. With that responsiveness, ARC calculates a flow rate to allow for an increase in pressure at a rate of 2 mmHg/min for the assistive model used here. Fluid responsiveness is continuously re-calculated during resuscitation allowing for adjustment of fluid rates to maintain the pressure response rate. Flow rates are halted to 0 mL/min once target MAP is reached.

Decision Table Logic Overview

The Decision Table V-ARC (DT V-ARC) was a type of controller that operated using a predefined set of vasopressor dosages. An error signal, obtained by taking the difference between the current and target MAP, was used to determine whether to increase vasopressor dosage. A 30-second sampling time was used for both versions of the decision table, and the dosage was increased by one step on the table if a positive error value was received. Once the target pressure was reached or exceeded, the system held the current dosage. The conservative table was obtained by making dosage changes that were half of the magnitude of the aggressive table, doubling the total number of steps used.

Scenario Development

To assess the performance of these preliminary V-ARC controllers and how the system parameters may impact the performance, a set of four simple test scenarios were derived. The HATRC system was initialized at a starting MAP of 45 mmHg, and a target MAP of 65 mmHg was used for each 15-min run. Tests were conducted in triplicate for all permutations of with or without noise and either utilizing or not utilizing a secondary ARC alongside the V-ARC controller. Conventional performance metrics such as median performance error (MDPE), target overshoot, controller effectiveness, and rise time efficiency were calculated to characterize performance differences (23).

Statistical Methods

For the analyzed swine data, we evaluated if there were significant differences between each derived metric and their impact on vasopressor activity. Specifically, this was split into four separate analyses looking at whether vasopressor concentration and pig variability significantly impact lag time to pressure increase, pressure increase or real response, overshoot, and pressure response rate (Fig. 10). Each of these were evaluated by two-way ANOVA with data separated by vasopressor dosing and swine subject. They were analyzed separately for differences between increase and decrease NE dosing steps on each of the 4 metrics, using a Welch’s t-test for these analyses. All data sets were identified as normally distributed using Shapiro-Wilk tests. Statistically significant effects were determined as p-values less than 0.05 and are indicated in the text and figures when applicable.

For the proof-of-concept controller testing, two-way ANOVA, post hoc Fisher’s least significant difference test was used to assess difference between controller configurations for each performance metrics. The two factors for this analysis were DT tuning (aggressive vs. conservative) and ARC state (on or off). Each noise configuration and performance metric were analyzed separately. Statistically significant effects were determined as p-values less than 0.05 and are indicated in the text and figures when applicable.

Funding Declaration

This work was funded by the U.S. Army Medical Research and Development Command (CO230029). This project was supported in part by an appointment to the Science Education Programs at the National Institutes of Health (NIH), administered by ORAU through the U.S. Department of Energy Oak Ridge Institute for Science and Education (M.D.L, J.M.B., R.G., T.M.R., and C.G.S.).

Conflict of Interest

Eric J. Snider is a co-inventor on a patent filed on the adaptive resuscitation controller technology (ISR 23-45-US02: 18/628,287). The other authors declare no conflict of interests for this work.

DoD Disclaimer

The views expressed in this article are those of the authors and do not reflect the official policy or position of the U.S. Army Medical Department, Department of the Army, DoD, or the U.S. Government.

Author Contributions

The initial concept was developed by D.B. and E.J.S. The animal data was captured by T.M.R., S.H.T., C.G.S., E.R., and E.J.S. Animal data analyses were performed by D.B., R.G., S.H.T., and E.J.S. Configuration and characterization of the VCM module were performed by D.B., M.D.L., A.R., and J.M.B. Proof-of-concept testing was performed by D.B., M.D.L., and A.R. Visualizations for the manuscript were by D.B., S.H.T., and E.J.S. Manuscript was drafted, reviewed, and edited by all authors. Funding and supervision were by E.J.S. All authors read and approved the final manuscript.

Data Availability Statement

The data presented in this study are not publicly available because they have been collected and maintained in a government-controlled database located at the U.S. Army Institute of Surgical Research. This data can be made available through the development of a Cooperative Research and Development Agreement (CRADA) with the corresponding author.

Gutierrez G, Reines HD, Wulf-Gutierrez ME. Clinical review: hemorrhagic shock. Crit Care. 2004;8(5):373–81.
Bonanno FG. Management of Hemorrhagic Shock: Physiology Approach, Timing and Strategies. J Clin Med. 2022;12(1):260.
Wise R, Faurie M, Malbrain ML, Hodgson E. Strategies for intravenous fluid resuscitation in trauma patients. World journal of surgery. 2017;41(5):1170–83.
Ramesh GH, Uma JC, Farhath S. Fluid resuscitation in trauma: what are the best strategies and fluids? International Journal of Emergency Medicine. 2019;12(1):38.
Vincent JL, Annoni F. Vasopressor therapy. Journal of Clinical Medicine. 2024;13(23):7372.
Ogbu OC, Murphy DJ, Martin GS. How to avoid fluid overload. Current opinion in critical care. 2015;21(4):315–21.
Rossaint R, Afshari A, Bouillon B, Cerny V, Cimpoesu D, Curry N, et al. The European guideline on management of major bleeding and coagulopathy following trauma: sixth edition. Critical Care. 2023;27(1):80.
Gupta B, Garg N, Ramachandran R. Vasopressors: Do they have any role in hemorrhagic shock? J Anaesthesiol Clin Pharmacol. 2017;33(1):3–8.
Sperry JL, Minei JP, Frankel HL, West MA, Harbrecht BG, Moore EE, et al. Early use of vasopressors after injury: caution before constriction. Journal of Trauma and Acute Care Surgery. 2008;64(1):9–14.
Burke EG, Hartley BW, Succar B, Dumas RP. Rethinking vasopressor use in the trauma bay: a shifting perspective. Trauma Surgery & Acute Care Open. 2025;10(Suppl 1).
Rinehart J, Ma M, Calderon MD, Cannesson M. Feasibility of automated titration of vasopressor infusions using a novel closed-loop controller. Journal of clinical monitoring and computing. 2018;32(1):5–11.
Rinehart J, Joosten A, Ma M, Calderon MD, Cannesson M. Closed-loop vasopressor control: in-silico study of robustness against pharmacodynamic variability. Journal of clinical monitoring and computing. 2019;33(5):795–802.
Wang S, Liu X, Yuan S, Bian Y, Wu H, Ye Q. Artificial intelligence based multispecialty mortality prediction models for septic shock in a multicenter retrospective study. npj Digit Med. 2025;8(1):228.
Jentzer JC, Patel S, Gajic O, Herasevich V, Lopez-Jimenez F, Murphree DH, et al. Early prediction of shock in intensive care unit patients by machine learning using discrete electronic health record data. Journal of Critical Care. 2025;88:155093.
Samei E. The future of in silico trials and digital twins in medicine. PNAS Nexus. 2025;4(5):pgaf123.
Riahi V, Diouf I, Khanna S, Boyle J, Hassanzadeh H. Digital Twins for Clinical and Operational Decision-Making: Scoping Review. J Med Internet Res. 2025;27:e55015.
Kadry K, Gupta S, Nezami FR, Edelman ER. Probing the limits and capabilities of diffusion models for the anatomic editing of digital twins. npj Digit Med. 2024;7(1):354.
Michael J, Cleophas L, Zschaler S, Clark T, Combemale B, Godfrey T, et al. Model-Driven Engineering for Digital Twins: Opportunities and Challenges. Systems Engineering. 2025;28(5):659–70.
Tasmurzayev N, Amangeldy B, Imanbek B, Baigarayeva Z, Imankulov T, Dikhanbayeva G, et al. Digital Cardiovascular Twins, AI Agents, and Sensor Data: A Narrative Review from System Architecture to Proactive Heart Health. Sensors. 2025;25(17):5272.
Snider EJ, Berard D, Vega SJ, Hernandez Torres SI, Avital G, Boice EN. An Automated Hardware-in-Loop Testbed for Evaluating Hemorrhagic Shock Resuscitation Controllers. Bioengineering. 2022;9(8):373.
Berard D, Vega SJ, Torres SIH, Polykratis IA, Salinas J, Ross E, et al. Development of the PhysioVessel: a customizable platform for simulating physiological fluid resuscitation. Biomed Phys Eng Express. 2022;8(3):035017.
Snider EJ, Berard D, Vega SJ, Ross E, Knowlton ZJ, Avital G, et al. Hardware-in-Loop Comparison of Physiological Closed-Loop Controllers for the Autonomous Management of Hypotension. Bioengineering. 2022 Sept;9(9):420.
Berard D, Vega SJ, Avital G, Snider EJ. Dual Input Fuzzy Logic Controllers for Closed Loop Hemorrhagic Shock Resuscitation. Processes. 2022;10(11).
National Research Council (US) Committee for the Update of the Guide for theCare and Use of Laboratory Animals. Guide for the Care and Use of Laboratory Animals [Internet]. 8th ed. Washington (DC): National Academies Press (US); 2011 [cited 2025 Sept 16]. (The National Academies Collection: Reports funded by National Institutes of Health). Available from: http://www.ncbi.nlm.nih.gov/books/NBK54050/
Rodgers TM, Berard D, Gonzalez JM, Vega SJ, Gathright R, Bedolla C, et al. In Vivo Evaluation of Two Hemorrhagic Shock Resuscitation Controllers with Non-Invasive, Intermittent Sensors. Bioengineering. 2024;11(12):1296.
Swindle MM, Makin A, Herron AJ, Clubb FJ, Frazier KS. Swine as Models in Biomedical Research and Toxicology Testing. Vet Pathol. 2012;49(2):344–56.
Jia H, Chang Y, Song J. The pig as an optimal animal model for cardiovascular research. Lab Animal. 2024;53(6):136–47.
Snider EJ, Vega SJ, Nessen IA, Hernandez Torres S, Salazar S, Berard D, et al. In Vivo Evaluation of an Adaptive Resuscitation Controller Using Whole Blood and Crystalloid Infusates for Hemorrhagic Shock. Frontiers in Bioengineering and Biotechnology. 2024;12:1420330.
Snider EJ, Vega SJ, Ross E, Berard D, Hernandez-Torres SI, Salinas J, et al. Supervisory Algorithm for Autonomous Hemodynamic Management Systems. Sensors. 2022;22(2):529.
Underwood W, Anthony R. AVMA guidelines for the euthanasia of animals: 2020 edition. Retrieved on March. 2020;2013(30):2020–1.

Competing interest reported. Eric J. Snider is a co-inventor on a patent filed on the adaptive resuscitation controller technology (ISR 23-45-US02: 18/628,287). The other authors declare no conflict of interests for this work.

Download PDF

Editorial decision: Revision requested
13 Nov, 2025
Reviews received at journal
12 Nov, 2025
Reviews received at journal
11 Nov, 2025
Reviews received at journal
07 Nov, 2025
Reviews received at journal
07 Nov, 2025
Reviews received at journal
02 Nov, 2025
Reviews received at journal
29 Oct, 2025
Reviews received at journal
27 Oct, 2025
Reviews received at journal
23 Oct, 2025
Reviewers agreed at journal
21 Oct, 2025
Reviewers agreed at journal
21 Oct, 2025
Reviewers agreed at journal
19 Oct, 2025
Reviewers agreed at journal
19 Oct, 2025
Reviewers agreed at journal
18 Oct, 2025
Reviewers agreed at journal
16 Oct, 2025
Reviewers agreed at journal
16 Oct, 2025
Reviewers agreed at journal
15 Oct, 2025
Reviewers agreed at journal
15 Oct, 2025
Reviewers agreed at journal
14 Oct, 2025
Reviewers agreed at journal
14 Oct, 2025
Reviewers invited by journal
14 Oct, 2025
Editor assigned by journal
14 Oct, 2025
Submission checks completed at journal
14 Oct, 2025
First submitted to journal
09 Oct, 2025

You are reading this latest preprint version

Vasopressor Control Module Development and Integration into the Hardware-in-Loop Automated Testbed for Resuscitation Controller Evaluation

Status:

Version 1

Abstract

Introduction:

Methods and Results

Conclusions

Figures

Introduction

Results

Discussion

Conclusion

Methods

Animal Data Capture and Analysis

Preprocessing and Data Analysis

Modifications of the HATRC Platform

Vasopressor Control Module

Noise Generator Module

HATRC Platform and Communications

Vasopressor Controller Proof-of-Concept Testing

ARC Overview

Decision Table Logic Overview

Scenario Development

Statistical Methods

Declarations

References

Additional Declarations

Status:

Version 1