A scalable approach to determine intracardiac pressure from mechanical circulatory support device signals

Abstract

Objective:

Effective mechanical circulatory support (MCS) relies on cardiac function measures to guide titration. Left ventricular end diastolic pressure (LVEDP) is a useful measure that is indirectly estimated using pulmonary artery catheters (PACs). PACs require additional intervention and provide intermittent and unreliable estimations. MCS device signals can estimate LVEDP but are prone to inter-device variability and require rigorous specialized characterization. We present a scalable and implementable approach to calculate LVEDP continuously using device signals.

Methods:

LVEDP was calculated from MCS device measured aortic pressure and motor current, which approximates the pressure head between the aorta and left ventricle. This motor current-pressure head relationship is device-specific but approximated using existing flow calibration and assumed physiologic relationships. Performance was evaluated with comparison from direct measurement of LVEDP in a series of acute animal models.

Results:

LVEDP measures (n=178,279) from 18 animals had good correlation (r=0.84) and calibration (Bland-Altman limits of agreement −7.77 to 7.63 mmHg; mean bias −0.07 ± 0.02 mmHg). The total mean error prediction interval was −3.42 to 3.32 mmHg and RMS error was 3.85 mmHg.

Conclusion:

LVEDP can be continuously calculated using device signals without specialized characterization. Calculated LVEDP values improved upon PAC estimations and were found using a scalable and manufacturer-accessible method.

Significance:

This method improves upon existing LVEDP measures without the need for rigorous characterization, external calibration, or additional intervention; this allows widescale deployment of continuous LVEDP measurement for patients on MCS and demonstrates key considerations necessary to translate research-grade technologies.

I. INTRODUCTION

Mechanical circulatory support (MCS) devices have been increasingly used to deliver hemodynamic stability for patients in heart failure and cardiogenic shock [1]–[3]. By reducing reliance on endogenous cardiac ejection, these devices can provide adequate systemic perfusion without increasing stress on the failing heart. Furthermore, many forms of MCS, such as indwelling percutaneous ventricular assist devices that are the focus of this work, also provide direct unloading of the ventricle to reduce wall stress and metabolic demands [3]–[6]. Several trials have shown improved outcomes from the use of this form of MCS for patients in cardiogenic shock [7]–[9]. While traditional medical therapy yields acute mortality rates as high as 40% [10], the use of MCS has presented the potential to support and recover endogenous cardiac function [11]–[15]. More effective use of indwelling MCS devices has been partially impeded by the challenge of titrating appropriate amount of cardiovascular support and degree of ventricular unloading [16]–[18]. Inadequate levels of support do not provide the benefit these devices may have, and excess impedes the recovery of native cardiac contractility [15], [19]–[22]. Consequently, frequent and accurate assessment of endogenous cardiac function is one of many critical measures that help guide clinicians in determining the appropriate operation of indwelling MCS devices.

Limitations in current clinically deployed methods to assess endogenous cardiac function have led to increased interest in more reliable and higher resolution diagnostic methods. The unique characteristics posed during the use of MCS have led to efforts that use the MCS device’s operation to better assess cardiovascular state [22]–[26]. Using MCS device signals reduces the need for additional intervention and can potentially provide an avenue for better assessment than other clinically used surrogate measures [27]. We and others have previously proposed methods that use device dynamics [28] and hysteretic relationships to predict different cardiac metrics, including left ventricular end diastolic pressure (LVEDP) [27].

LVEDP, normally ranging from 8 to 15 mmHg, is a useful measure as elevation can indicate ventricular volume dependence and congestion that occurs with depressed cardiac function as described by the Frank-Starling relationship. Current clinical practice infers LVEDP from intermittent estimates of the surrogate pulmonary capillary wedge pressure measured using a venous pulmonary artery catheter [29]. While pulmonary artery catheters may offer many forms of diagnostic information, the intermittent pulmonary capillary wedge pressures it provides are least reliable in disease states and yield errors as large as 5–6 mmHg when compared with direct measurement of left ventricular pressure, especially during concurrent use with MCS devices [30]–[32]. Consequently, LVEDP posed an ideal measure to estimate continuously using MCS device signals.

Our prior work demonstrated the feasibility of calculating near real-time LVEDP using device-heart hysteretic interactions with root mean square errors (RMSE) of <2 mmHg when compared with direct measurement [27]. This method required careful quantification of each support device using a mock circulatory loop before deployment – a situation that requires well-controlled research environments to recapitulate. While demonstrating the feasibility of using the device signals, the prior approach remained susceptible to error from inter-device and signal gating variability. Without compensating for these sources of variability, the aggregated error increases beyond limits of clinical utility. The need for such specificity to produce clinically meaningful results poses a challenge towards scalable clinical translation for this and many other techniques demonstrated in limited research settings.

The methods discussed here allow for near real-time LVEDP estimation across production MCS devices by quantifying device-heart interaction without device-specific characterization. Significantly, the approach only uses current standard characterization performed by most device manufacturers and does not need any external clinical calibration during deployment. We demonstrate our results with a number of devices across a wide range of stressed physiologic conditions using a variety of pharmaceutical and mechanical interventions. Furthermore, this work not only outlines the often missing but necessary steps that consider scalability for translating a research-grade technique to clinical deployment but also demonstrates improved performance over pulmonary artery catheter estimates for measuring LVEDP without additional intervention during MCS deployment.

II. METHODS

A. Impella and Device-Heart Interactions

While the approach is generalizable to other forms of MCS, we use the Impella (Abiomed, Danvers, MA) as a paradigmatic device given its approved clinical indications for use and percutaneous deployment. The Impella is a mixed-flow percutaneous ventricular assist device that consists of an axially rotating impeller on the end of a 9 Fr catheter. Blood is drawn from an inflow area on the distal end of the catheter that resides within the left ventricle, through a lumen, and out of the lumen via an outflow port that is positioned within the ascending aortic arch (Fig 1a). The axially rotating impeller is near the outflow area and acts by suctioning blood out of the left ventricle and into the aorta. The device has an optical pressure sensor near the impeller that measures the proximal aortic pressure when appropriately positioned (Fig 1a). The device operates at a clinician-determined speed ranging from 23,000 RPM to 46,000 RPM (corresponding to P-1 to P-9 settings) during routine clinical use and provides a flow rate of up to 3.5 L/min [33]. This flow rate is estimated from motor current using a clinically deployed algorithm that relies on a validated motor current calibration that is performed by the manufacturer during final quality checks. The external end of the catheter is connected to a controller to allow clinician control and recording of the optical pressure sensor, motor current, and motor speed signals.

Fig. 1.

Schematics of Impella percutaneous ventricular assist device placement and hysteretic interactions that occur with endogenous cardiac contractility. (a) The Impella is a mixed flow pump mounted on the end of a 9 Fr catheter that is typically placed in a retrograde approach from femoral artery access. An axially rotating impeller sits within the ascending aorta and pulls blood flow from the left ventricle (LV) through a small cannula that bridges the aortic valve. An optical pressure sensor near the outflow in the ascending aorta measures the proximal aortic pressure. (b) A modified pump performance curve can visualize the influence of the heart on device motor current by plotting the motor current against the pressure head across the device (the difference between aortic pressure and left ventricular pressure). Influence from the heart manifests as a hysteresis loop around the typical static pump performance curve (grey) that can be separated into different cardiac phases of LV contraction (red), LV relaxation (blue), and LV filling (green). The key points, LV end diastole and peak systole, are readily identifiable points (orange) on the hysteresis loop when it is measured.

During most use cases, the Impella provides partial support in conjunction with remaining endogenous cardiac contractility. Consequently, the Impella is exposed to variable loading conditions and adjusts its power draw via motor current according to cardiac pulsatility and cardiac phase [27]. Influence from endogenous cardiac contractility appears on the pump-performance curve (pressure head versus pump flow) as a hysteresis loop centered around standard static characterization; our prior work demonstrated an alternative visualization of hysteretic pump-performance curves by using the device motor current as a surrogate for pump flow [27]. In this case, the pressure head across the device is defined as the difference between aortic pressure and left ventricular pressure. This hysteretic relationship between motor current and pressure head can be seen cycling around a static relationship and separated by cardiac phases of ventricular contraction, relaxation, and filling with the point corresponding to left ventricular end diastolic pressure being readily identifiable and stable across variable contractility (Fig 1b).

B. Left Ventricular End Diastolic Pressure Calculation

Pump performance can be quantified using the previously described motor current-pressure head relationship [27]. Known motor current can determine an estimated pressure head from this relationship. Consequently, the motor current and aortic pressure at the time of end diastole can be used together to determine LVEDP as follows.

$$LVEDP=Ao{P}_{ED}-\text{Δ}P\left(i_{\text{motor},ED}\right)$$ (1)

In the above equation, the LVEDP is determined by the difference of the aortic pressure at end diastole (AoP_ED) and the pressure head (ΔP) in mmHg as a function of the motor current at end diastole (i_motor,ED) in A. This approach requires accurate prior knowledge of the motor current-pressure head relationship at end diastole. Our prior work has shown that this can be characterized across a range of loading conditions and remains stable with varying contractility [27]. Here we present two combined approaches that allow for scalability by accounting for inter-device variability: a generalized motor current-pressure head curve that is built off of existing single-point flow characterization and a peak offset method that only uses internal methods for calibration.

Instead of recreating the entire motor current-pressure head relationship with a mock circulatory loop, the relationship can be estimated by a pre-determined spline with a two-element quadratic and linear portion (Fig 2a). A set of two splines, which correlates to the non-hysteretic static condition that the hysteresis moves around (Fig 1b), are pre-determined using flow rate estimation curves as a basis and selected based on motor speed and a characteristic motor current. The Impella and other MCS devices are currently characterized during manufacturing by measuring individual pump motor performance (i_FQ) in the context of historical distributions of pump performance, which is used for flow estimation and appropriate estimation via the controller. Together with the motor speed setting from the controller (e.g., P-2, P-6, etc.), this leads to a specific motor current-pressure head relationship that corresponds to the device but is not device-specific (Fig 2a).

Fig. 2.

Representations of the motor current-pressure head relationships valid at end diastole and the key identifying points used in the gating algorithm for detecting end diastole for estimation of left ventricular end diastolic pressure.(a) Left ventricular end diastolic pressure can be estimated by using the Impella measured aortic pressure and a pressure head (difference between aortic pressure and left ventricular pressure) determined from the motor current. This pressure head-motor current relationship is pump and operating speed specific (P-2, P-6, P-9, etc.). Pump specificity is estimated by using an already performed manufacturer measurement of a characteristic motor current (iFQ) to select a pre-determined spline that is valid at end diastole and peak systole. This selection is further refinement by ensuring that peak systolic pressure head remains physiologic. (b) Timing of end diastole is determined by a robust gating method that sub-windows each beat using a surrogate estimate of left ventricular pressure. The ventricular filing phase window is bound by the minimum pressure and peak slope of contraction (Window 1). A second window bound by the last and second to last extrema of the curvature defines the range of end diastolic values (Window 2). End diastole is gated by finding the minimum value between the local minimum extremum of the second window and the end of the window (Window 3).

The motor current-pressure head relationship selected using the motor speed and iFQ is more specific to, but not necessarily characteristic of, a given pump. Thus, an internal calibration is performed to fine-tune the relationship for a given device. The pump hysteresis is based around a true motor current-pressure head relationship that is estimated by the set of two splines selected using the motor speed and i_FQ. Notably, the device operates directly on the curve at end diastolic pressure and peak systolic ejection (Fig 1b). Consequently, the same motor current-pressure head relationship can also be used to calculate left ventricular peak systolic pressure in the context of aortic pressure measured during ventricular ejection. The motor current waveform is linearly scaled using this relationship such that the pressure head remains physiologically consistent with the measured aortic pressure to provide further device specificity.

Combined, these two approaches can estimate the motor current-pressure head relationship to allow for an approximation of a device-specific characterization using general relationships. The first uses a single point motor current measurement that is typically performed by most MCS device manufacturers for flow estimation. The second uses an internal calibration to provide more device-specificity. Both approaches lead to an LVEDP calculation that can be performed automatically by the algorithm for ease of deployment at the bedside and does not require external calibration by the clinician.

C. Gating Methods

Left ventricular end diastolic pressure calculation is reliant on accurate signal gating to determine the time of end diastole for the correct aortic pressure and motor current. Because of the rapid temporal changes that come with contraction, inaccurate gating can lead to substantial errors that are independent of the calculation method itself. Here we gate signals by using features in the motor current waveform transformed using the previously determined motor current-pressure head waveform. This signal has a resemblance to a left ventricular tracing and is recorded by the Impella controller at 250 Hz. Several sub-windows are selected to refine the detection of end diastole. The minimum waveform value defines the end of relaxation, and contraction is defined using the maximum of the time derivative to bound the ventricular filling phase (Fig 2b). The second to last and last local extrema of the curvature found via the second time derivative of this segment is used to define a second window to search for end diastolic values. End diastolic pressure is then identified by using the minimum value between the local minimum extremum and the end of this window (Fig 2b). Method precision is assessed by comparison with the electrocardiogram QRS complex. The motor current, motor speed, and aortic pressure at this time are then used in the previously described algorithm to calculate LVEDP. A similar approach is taken to determine the time of peak systolic ejection for the internal motor current scaling previously described. Results are then presented after a 60 second median filter to address any unaccounted motor current noise.

D. Animal Models

A series of 18 acute porcine animals were used for assessment and validation of LVEDP estimation methods across a range of physiologic conditions induced by pharmaceutical and mechanical interventions. All animals were young adult castrated male Yorkshire swine (~75 kg) maintained in accordance with NIH and AAALAC guidelines (CBSET, Lexington, MA). Animals were intubated and maintained in anesthesia using inhaled isoflurane after induction via intramuscular injection of Telazol (6 mg/kg). Body temperature, oxygen saturation, and electrocardiogram were monitored throughout the duration of all studies. Animals were separated into a training set (n=8) for algorithm refinement and test set (n=10) for validation.

Vascular access was obtained at both femoral arteries and veins, the left jugular vein, and the left carotid artery. An Impella CP was introduced via the right femoral artery and was placed in the left ventricle. A pigtail catheter to measure left ventricular pressure was placed via the left carotid artery. A Judkins left catheter was used to access the left main coronary artery via the left femoral artery. Pacing wires to the right atrium were placed via the left jugular vein. Pharmaceutical interventions were delivered via the femoral venous sheath. Placement of all catheters and aortic valve competency was confirmed via fluoroscopy. In a subset of animals (n=8), a pulmonary artery catheter was placed via the right jugular vein for intermittent pulmonary capillary wedge pressure estimation of LVEDP that is confirmed by expert clinician review.

Pharmaceutical and mechanical interventions were used to induce a wide range of LVEDP values across a variety of loading conditions to simulate the range of disease conditions that the algorithm may face with use in patients. Measurements at all conditions were performed at four Impella speeds (P-2, P-4, P-6, and P-8) ventricular filling volume permitting; high speeds were only used if there was adequate ventricular volume to prevent pump suction. Data at stable baseline conditions was measured for at least two minutes per Impella speed prior to interventions. The impact of heart rate was investigated with otherwise baseline physiology using right atrial pacing at 100, 120, and 150 bpm. Ventricular filling was reduced using a mechanical occlusion of the inferior vena cava with a balloon-tipped catheter introduced via the right femoral, which induced low LVEDP values. Pharmaceutical interventions were given for positive inotropic (norepinephrine drip at 0.1 μg/kg/min), negative inotropic (esmolol boluses at 1 mg/kg), and high afterload (phenylephrine drip at 3 μg/kg/min) conditions for a range of LVEDP values and loading conditions. Finally, acute cardiogenic shock was induced by injecting compressible 45–105 μm diameter microspheres (Hydropearl, Terumo, Tokyo, Japan) into the left main coronary artery. Boluses of 0.25 mL of microspheres mixed with 10 mL of isotonic saline and 10 mL of contrast were serially injected until LVEDP was elevated >20 mmHg or the heart could no longer maintain ventricular-vascular coupling.

During all interventions, a single-lead electrocardiogram, the left ventricular pressure, the femoral venous pressure via sheath, and the femoral arterial pressure via sheath were continuously recorded by an external research-grade data acquisition system (ADInstruments, Dunedin, New Zealand). Direct measurement of left ventricular pressure with electrocardiogram gating was used as a gold standard reference for both calculated and pulmonary artery catheter derived measures. Impella signals, including aortic pressure, motor current, and motor speed, were recorded and retrieved from the Impella controller.

E. Statistical Analysis

Error analysis and algorithm improvements were performed using the training set of data and validated with the test set. Statistical analyses for algorithm performance are presented for the separate data sets and the pooled data. Values are reported as a mean ± standard deviation with errors reported using root mean square differences. Errors are determined using direct measurement of left ventricular pressure and electrocardiogram gating as a reference value. Differences in error distribution between data sets were assessed using a two-sided Student’s t-test. Overall performance is assessed using sample weighted meta-analysis approaches with a prediction interval [34]. The sum data is presented using correlation and Bland-Altman plots [35]. For all methods, standard confidence intervals (95%) and significance values (p<0.05) were used. Data and statistical analyses were performed in MATLAB (Mathworks, Natick, MA).

III. RESULTS

Data from the series of 18 acute animals resulted in 178,279 separate LVEDP calculations across a wide range of physiologic conditions with mean aortic pressure ranging from 23.7 to 140.4 mmHg, heart rate from 41 to 207 bpm, and measured LVEDP from 0.90 to 39.8 mmHg.

Our approach resulted in an improved calculation of LVEDP despite no device-specific characterization. The end diastolic gating method was robust with negligible differences when compared with the timing of the electrocardiogram (0.004 ± 0.028 s). In representative animal time courses from each data set that were selected due to their extreme values and availability of pulmonary artery catheter derived estimations, a wide range of LVEDP values was calculated with trends and rapid temporal changes captured by the algorithm (Fig 3). For all the animals, the average mean error was −0.04 ± 0.37 mmHg and can be seen for each animal and set as a mean and confidence interval in a forest plot; the prediction interval for the total mean error is −3.42 to 3.32 mmHg (Fig 4). The training set resulted in a weighted RMSE of 3.64 mmHg, and the test set resulted in a weighted RMSE of 4.00 mmHg. Using all animals resulted in a total weighted RMSE of 3.85 mmHg with no significant difference between training and test set error distribution (p = 0.26).

Fig. 3.

Left ventricular end diastolic pressure time course over the duration of acute animal trials with algorithm calculated values (blue) and direct measurement using a pig-tail catheter as a reference (red). Intermittent pulmonary artery catheter (PAC) derived estimates were also taken (grey). (a) A representative training set animal shows variable trends in left ventricular end diastolic pressure that are generally captured by the calculated method over time with a range from 0.90 to 21.9 mmHg. The total training set data resulted in a RMSE of 3.64 mmHg. (b) A representative test set animal shows a generally rising trend in left ventricular end diastolic pressure that is captured over time with a range from 5.15 to 39.8 mmHg. The total testing set data resulted in a RMSE of 4.00 mmHg. There is no significant difference in RMSE distribution between test and training set (p = 0.26) with a total combine RMSE of 3.85 mmHg.

Fig. 4.

Forest plot of mean difference between calculated and reference left ventricular end diastolic pressure for all animals and sets. The mean difference for each animal is shown with a 95% confidence interval for the mean. While the mean errors differed between individual animals, the effect size was small with a range of mean absolute error from 0.161 to 2.68 mmHg. Confidence intervals for the mean of each set (dark blue) and the total animals (light blue) is shown overlaying the respective prediction interval. There were no statistically significant differences between the training and test animal sets with a mean error across all animals of −0.04 ± 0.37 mmHg and a prediction interval for the mean from −3.42 to 3.32 mmHg.

When taken as a whole, the total data pool shows good correlation (r = 0.84, p < 0.001) (Fig 5a) and is illustrated on a Bland-Altman diagram with a mean bias of −0.07 ± 0.02 mmHg and limits of agreement from −7.77 to 7.63 mmHg (Fig 5b). There were no significant error biases based on pump speed, heart rate, or LVEDP, but high mean arterial pressure (>120 mmHg) yielded errors larger than the error threshold set by typical pulmonary artery catheter estimation.

Fig. 5.

Correlation and Bland-Altman diagrams for left ventricular end diastolic pressure calculation compared with direct measurement as a reference. Data is shown from all 18 animals with a total of 178,297 independent calculated measurements with corresponding reference values ranging from 0.90 to 39.8 mmHg.(a) The correlation plot showed strong agreement between the measurements (r = 0.84, p<0.001) with minimal deviation from a unity relationship. (b) The Bland Altman diagram is shown with a small mean bias of −0.07 ± 0.02 mmHg and limits of agreement from −7.77 to 7.63 mmHg. There are no clear biases and trends across the range of end diastolic pressure measurements.

For comparison, pulmonary artery catheter derived wedge pressure estimation of LVEDP was obtained in 8 animals with 96 separate comparisons to reference LVEDP measurements ranging from 2.82 to 32.9 mmHg. Intermittent estimations were performed concurrently with continuously calculated and directly measured LVEDP through the course of each trial (Fig 3) and yielded a total RMSE of 7.20 mmHg. As a whole, the correlation was weaker (r = 0.37, p < 0.001) and Bland-Altman analysis yielded a mean bias of −2.38 ± 0.70 mmHg and limits of agreement from −15.8 to 11.0 mmHg (Fig 6).

Fig. 6.

Correlation and Bland-Altman diagrams for pulmonary artery catheter derived wedge pressure estimation for LVEDP compared with direct measurement as a reference. Data from 8 animals with 96 intermittent measurements have corresponding reference values ranging from 2.82 to 32.9 mmHg. (a) The correlation plot showed a weak correlation (r = 0.37, p<0.001) with a large deviation from a unity relationship indicating poor calibration. (b) The Bland Altman diagram is shown with a significant mean bias of −2.38 ± 0.70 mmHg and limits of agreement from −15.8 to 11.0 mmHg that is similar to previously reported comparisons.[36]

IV. DISCUSSION

Mechanical circulatory support devices have the potential to improve outcomes for patients in heart failure by providing hemodynamic stability and direct ventricular unloading [3]–[6]. Less invasive devices, such as the Impella and other percutaneous ventricular assist devices, may also have the added benefit of promoting recovery of endogenous cardiac function [11]–[15]. However, these devices are only positioned to foster “bridge-to-recovery” if the degree of support and unloading is appropriately titrated to the patient state. This clinician-guided titration relies on many markers including accurate determination of cardiac function.

Left ventricular end diastolic pressure, which is calculated in the work here, is an often-used clinical marker of ventricular function. Elevations in LVEDP act as markers of increased reliance on preload to compensate for reductions in forward cardiac output as described by the Frank-Starling relationship. Consequently, pulmonary artery catheters are widely deployed to monitor cardiac function in part by estimating LVEDP via wedging a balloon-tipped catheter in a distal pulmonary artery to measure left atrial pressures indirectly [29]. The resulting atrial-like pressure tracing must then be interpreted to estimate the corresponding end diastolic pressure. Though providing many other different diagnostic measures, pulmonary artery catheter estimation of LVEDP not only requires additional intervention but also is an unreliable and only intermittently obtained marker with substantial errors and wide limits of agreement [30]–[32], [36]. Thus, LVEDP is an ideal measure to estimate using new continuous techniques without additional intervention for patients undergoing treatment with MCS devices.

Mechanical circulatory support device signals and operation can be leveraged to estimate a variety of cardiac markers including LVEDP [37]–[41]. Our prior work showed the capability of estimating pressure head from motor current using device-specific characterization from a specialized mock circulatory loop [27]. However, all of this work is reliant on labor-intensive characterization and controlled testing conditions that pose an often-overlooked barrier to widescale translation of technologies to clinical practice. Disregard for the feasibility of scalability cripples the value of novel approaches. Indeed, the application of our previously described method for estimation of LVEDP in more varied test conditions with only manufacturer-capable levels of device characterization led to more than four-fold increases in RMSE. Consequently, our work here not only presents a new approach to estimating LVEDP using MCS device signals but also demonstrates the importance of scalability when developing new clinical tools for MCS.

A. Approach and Physiologic Validation

Here we present a method to estimate LVEDP from Impella device signals that was developed to only use manufacturer available device characterization. In this paradigmatic case with the Impella, the final quality motor current value (i_FQ) is an existing manufacturer-performed measure used for quality evaluation and flow rate estimation. The i_FQ is a single point characterization that is easily performed and used as a guiding parameter to select from a family of pre-generated two-phase motor current-pressure head curves (Fig 2a). Curve selection from i_FQ and subsequent motor current scaling to further improve pump specificity was refined using a training set of data and independently validated on the test data set, for which there was no statistically significant difference in errors (p = 0.26). This is used with a gating method that yielded negligible temporal errors (0.004 ± 0.028 s) when considering the Impella 0.004 second data sampling period (Fig 2b). These were evaluated using acute animal models with a variety of pharmaceutical and mechanical interventions, including a cardiogenic shock state that induced large changes in LVEDP to stress calculation with values as high as 39.8 mmHg (Fig 3). Importantly, there is no bedside or post-deployment calibration for pressure measurements, eliminating need or reliance on even temporary secondary diagnostic tools.

Pulmonary artery catheters are the existing clinical standard for estimating LVEDP and thus set our performance benchmark. Pulmonary artery catheters were deployed in a subset of animals and yielded results that are comparable with existing literature. The RMSE of 7.20 mmHg is in line with previously reported estimations of ~5.0 mmHg [32], [42]. Bland-Altman analysis was also comparable with a mean bias of −2.38 ± 0.70 mmHg and limits of agreement from −15.8 to 11.0 mmHg that is very similar to the previously reported mean bias of −2.87 mmHg and limits of agreement from −17.2 to 11.4 mmHg [36]. In comparison to pulmonary artery catheter estimation, the pooled calculated LVEDP with 178,279 separate comparisons across 18 animals yielded a lower RMSE (3.85 mmHg), improved limits of agreement (−7.77 to 7.63 mmHg), and lower mean bias (−0.07 ± 0.02 mmHg). While these comparisons rely on data pooling and unbalanced sample count, the pulmonary artery catheter performance seen here is in line with existing literature and provides appropriate clinically relevant context to assess the performance of the calculated LVEDP values.

To best portray the calculated data, results per animal and data set are shown in a forest plot. There is no significant difference between test and training set performance, but there are significant differences in the mean error for individual animals (Fig 4). This is a product of the high number of samples per animal and slight variations in i_FQ-based curve estimations with each device. Usually, this would be prohibitive to pooling results from each animal; however, the effect size is small (0.161 to 2.68 mmHg) when compared to the residuals for each animal and standard deviation of the residuals for pooled LVEDP comparisons. To better interpret the expected performance of our results we also report a prediction interval for the mean difference. Because there was no significant difference between the test and training set, mean differences for all animals were used to find a prediction interval from −3.42 to 3.32 mmHg (Fig 4). The prediction interval indicates that 95% of future measurements would yield a mean error between these bounds with high reliability due to the central limit theorem and are a good indicator of expected future performance of calculations.

B. Limitations and Future Work

This new approach for calculating LVEDP using MCS device signals yielded scalable results that improve upon pulmonary artery catheter estimations of LVEDP. Deployment of this technique to existing Impella controllers can allow reporting of clinically actionable LVEDP values without the need for additional intervention and with no need for bedside calibration. Measurements are taken near real-time using a 60-second window, allowing for continuous measurement of what was only an intermittent estimation by pulmonary artery catheters. There was no statistically significant impact on error from pump speed, heart rate (41–207 bpm), or LVEDP (0.90–39.8 mmHg), however mean arterial pressure in excess of 120 mmHg yielded an increased error. The reason for this is likely because high mean arterial pressure uses the iFQ-based curves higher regions that are either extrapolated or less linear (Fig 2a). This condition is less clinically relevant and unlikely to occur with patients requiring hemodynamic stability from mechanical circulatory support. Furthermore, this scenario can be readily identified with devices like the Impella that also contain aortic pressure measurements. Our LVEDP calculations do improve upon pulmonary artery catheter measurements, but there are still potentially relevant degrees of error. Less drastic device-specific characterization than that previously reported may allow for reduced error.

The approach and constraint of employing scalable characterization methods can be generalized to other forms of mechanical circulatory support. However, the specifics of the algorithm remain reliant on direct access to the left ventricle and some features from the Impella, such as the aortic pressure measurement and high frequency data capture. Applying this approach to devices that can be deployed in lower-acuity care settings would be beneficial for chronic and longitudinal monitoring and is a necessary step towards self-actuated physiologic control devices. Acute animal models simulated a wide range of physiologic states however the method should still be evaluated in humans for clinical validation, ideally during MCS-assisted percutaneous coronary intervention for comparison with directly measured LVEDP values.

V. CONCLUSION

Mechanical circulatory support devices are a promising therapeutic avenue for patients requiring hemodynamic stability and ventricular unloading. The methods proposed here not only potentially eliminate the need for further intervention with diagnostic instruments but also improve on existing surrogate estimates of LVEDP for patients with these devices by providing more reliable and continuous measures. This new approach is developed with scalability in mind to allow for translation from controlled research setting to the clinic for rapid application to improve patient outcomes.