Screening Patients for Elevated Left Ventricular End‐Diastolic Pressure Using a Novel Noninvasive Brachial Cuff–ECG System: A Multicenter Validation Study

Abstract

Background

Measurement of left ventricular end‐diastolic pressure (LVEDP) is an established diagnostic method to evaluate heart failure but requires an invasive procedure. A noninvasive technique for detecting elevated LVEDP would improve the diagnosis of heart failure. Herein, we present the results of a multicenter study to validate a noninvasive system to detect elevated LVEDP.

Methods

The Vivio System includes a modified blood pressure cuff and single‐lead ECG to capture brachial artery waveforms. A model was created to identify patients with elevated LVEDP (>18 mm Hg). For the invasive cohort, all patients were referred for coronary angiography and left heart catheterization for clinical indications. A group of 321 patients with no significant health problems were enrolled as a control cohort. Invasive LVEDP measurements were performed using Millar catheters. The training data set (n=262) contained 101 patients with LVEDP measurements (n=44 with LVEDP >18 mm Hg) and 161 controls. The validation data set (n=155) contained 75 patients with LVEDP measurements (n=40 with LVEDP >18 mm Hg) and 80 controls.

Results

Leave‐one‐out cross‐validation on the training data set yielded a sensitivity of 0.84 (95% CI, 0.70–0.93]) and a specificity of 0.84 (95% CI, 0.79–0.89). The validation data set showed a sensitivity of 0.80 (95% CI, 0.64–0.91) and a specificity of 0.83 (95% CI, 0.75–0.90).

Conclusions

The Vivio System can accurately detect elevated LVEDP and has the potential to significantly improve early detection of heart failure in the outpatient setting.

Nonstandard Abbreviation and Acronym

LVEDP

left ventricular end‐diastolic pressure

Clinical Perspective.

What Is New?

• The Vivio System can accurately identify patients with elevated left ventricular end‐diastolic pressure (>18 mm Hg) with a sensitivity of 0.80 and a specificity of 0.83.

What Are the Clinical Implications?

• The Vivio System may enable the earlier detection of heart failure in high‐risk patients within the primary care setting, addressing a critical gap in timely diagnosis.

• By identifying patients before heart failure decompensation occurs, the Vivio System may support earlier initiation of guideline‐directed medical therapy, which can improve outcomes and reduce hospitalizations.

Heart failure (HF) affects millions of people worldwide and is a leading cause of morbidity and death. ¹ , ² , ³ The annual cost to care for patients with HF is estimated at nearly $30 000 per patient in the United States, with a total economic burden in 2020 of ≈$200 billion. ⁴ A simple noninvasive tool to aid in the diagnosis of HF would enable health care providers to accelerate access to diagnostic pathways and initiate treatment for at‐risk patients. Elevated cardiac filling pressures are one of the earliest signs of HF and often precede the development of symptoms, physical findings, and structural remodeling of the left ventricle. ⁵ , ⁶ Elevated filling pressures, assessed by the direct measurement of left ventricular end‐diastolic pressure (LVEDP), is independent of the type of HF but requires an invasive procedure and therefore is not practical to apply in the context of public health. ⁷ , ⁸ A simple and rapid noninvasive method to measure LVEDP would be beneficial from a public health perspective.

In the current study, we sought to validate a novel, noninvasive method to measure LVEDP. This noninvasive method relies on the Vivio System that includes a modified blood pressure cuff and a single‐lead ECG, which was cleared by the Federal Drug Administration in 2023 for screening patients to aid in the diagnosis of HF in adults aged ≥21 years. ⁹

Methods

Due to the sensitive and protective nature of the data collected for this study, the data, analytical methods, and study materials will not be made available to other researchers for the purpose of reproducing the results or replicating the study procedures.

Study Population

The study population included 728 patients aged ≥21 years from 8 study sites across the United States and was composed of 2 separate cohorts: an invasive cohort and a healthy control cohort. The invasive cohort consisted of 407 patients referred for nonemergent left heart catheterization inclusive of direct measurement of LVEDP as part of routine clinical management. For the invasive cohort, patients with acute coronary syndrome, cardiogenic shock, the need for intravenous inotropic, myocardial infarction within a week of the scheduled procedure, or those requiring mechanical circulatory support were excluded from enrollment as a part of the study protocol. The healthy control cohort consisted of 321 participants with no known significant health problems. All participants provided written informed consent, and the study protocol was approved by the following institutional review committees: Advarra; the Spectrum Health Human Research Protection Program Office of the Institutional Review Board; MetroWest Medical Center Institutional Review Board; University of California, San Diego Human Research Protections Program; and the Institutional Review Board at Louisiana State University Health Shreveport.

Data Collection

Noninvasive data were collected using the Vivio System, which includes a modified brachial cuff and single‐lead ECG (Figure 1), developed by Ventric Health. ¹⁰ In addition to providing brachial pressure measurements, the cuff also acquires 40 seconds of brachial pulse wave data at suprasystolic blood pressure (systolic blood pressure plus 35 mm Hg). It is important to note that this system is distinct from the wireless optical Vivio handheld device presented in previous publications that measures the carotid artery pulse waveform. ¹¹ , ¹² Left ventricular pressure data were acquired using Mikro‐Cath catheters (Millar, Houston, TX). A sample of the study signal data is shown in Figure 2. All signal data were acquired simultaneously at a sampling rate of 1 kHz from a PowerLab data acquisition system (ADInstruments, Colorado Springs, CO).

Figure 1. The Vivio System: a synchronized single‐lead ECG and modified blood pressure cuff that communicates via Bluetooth to a tablet (iPad).

Figure 2. An example recording of the study data showing the synchronized invasive and noninvasive signals.

A, Thirty seconds of synchronized left ventricular pressure, brachial pulse waveform, and ECG signals. B, A magnified view of ≈1 cardiac cycle showing the locations of LVEDP, the foot of the brachial pulse waveform and the onset of the QRS. LVEDP indicates left ventricular end‐diastolic pressure.

Signal Processing

Vivio System data were down‐sampled to 500 Hz before analysis. A conventional band‐pass filter was applied to the ECG signal to mitigate any baseline wander and environmental noise present in the signal. After filtering, ECG beats were identified, and fiducial point detection was performed to identify fiducial points within each ECG beat. Baseline wander was removed from the brachial pulse waveform through the application of a finite impulse response high‐pass filter with a cutoff at 0.1 Hz. This step was followed by a wavelet low‐pass filter with a cutoff at 30 Hz to remove high‐frequency noise. Cardiac cycles in the brachial pulse waveform were identified using the slope sum function. ¹³

For each cardiac cycle, a combination of first to fourth derivative calculations were used to locate the following fiducials: systolic and diastolic peaks, notch, maximum slope, and inflection point. Figure 2B shows several key fiducials located on the signal data such as the QRS onset of the QRS complex (green dot), the LVEDP value at the R‐peak of the ECG (red dot), and the foot of the brachial pulse waveform (blue dot). ECG beats were then matched to their corresponding cardiac cycles by identifying the brachial pulse cycle and ECG beat pairs with a time interval between the QRS onset of the ECG (green dot) and foot of the brachial pulse waveform (blue dot) ≤400 milliseconds. Only matched cycles and their fiducials were used to compute features for modeling. LVEDP values were determined by measuring the pressure value from the left ventricular pressure catheter at the time corresponding to the R wave of the ECG signal, ¹⁴ as depicted by the red dot and dashed line in Figure 2B. Healthy controls were labeled as having nonelevated LVEDP, as no catheter data were available.

Exclusion Criteria for Model Development

Patients were excluded from the model development on the basis of medical history as well as data quality issues. Most of these issues originated from poor quality or missing catheter or Vivio System signals, including issues such as no signal from the catheter, low‐amplitude ECG signals (<0.1 mV) or discontinuities and artifacts due to the participants moving during data acquisition. As a result, LVEDP could not be determined from the catheter tracing, or features could not be extracted from the Vivio measurements. Patients with severe valvular disease were excluded from the invasive cohort. Healthy cohort participants were excluded if they had a history of regular smoking or any medical conditions that may affect their LVEDP such as diabetes, cardiomyopathy, hypertension, kidney disease, or lung disease. Additionally, several patients were also excluded on the basis of their arm sizes being outside the range of the blood pressure cuff (22–42 cm) and/or systolic blood pressures lower than 90 mm Hg. The breakdown of the study population based on the exclusion criteria from the model development requirements is shown in Figure 3.

Figure 3. Flow diagram detailing the process for generating the training and validation data sets from the study population including exclusions.

Prediction Algorithm

A logistic regression classifier with log‐F (m, m) penalty (m=1), ¹⁵ and embedded conic solver ¹⁶ was used to develop the prediction model. The log‐F (m, m) penalty ¹⁵ was used to minimize the prediction bias and improve overall predictive performance. Features computed from the ECG and brachial blood pressure signals were explored as inputs to the model. Feature selection was performed to reduce the model complexities, training time, data noise, and chances of overfitting (Figure S1).

Only LVEDP values from postexpiratory cardiac cycles ¹⁷ were used for model development. In this study, postexpiratory cardiac cycles are defined as those above the 75th percentile of LVEDP measurements, The upper quartile value of postexpiratory LVEDP measurements was used as the final ground‐truth target for a patient. The LVEDP threshold of 18 mm Hg ¹⁸ , ¹⁹ , ²⁰ was chosen to generate prediction targets, which defined elevated LVEDP as >18 mm Hg and nonelevated LVEDP as ≤18 mm Hg. To compute a prediction for a patient, cycle‐level predictions were collapsed using the upper quartile values.

Results

The training data set contained 161 healthy controls, 57 patients with nonelevated LVEDP (≤18 mm Hg), and 44 patients with elevated LVEDP (>18 mm Hg). The validation data set contained 80 healthy controls, 35 patients with nonelevated LVEDP (≤18 mm Hg), and 40 patients with elevated LVEDP (>18 mm Hg). For the invasive cohort, training data were collected from September 22, 2021, to May 7, 2022, and validation data from May 8, 2022, to July 8, 2022. For the healthy control group, training data were gathered between June 16, 2022, and July 8, 2022, while validation data were collected from July 18, 2022, to August 26, 2022. A summary of the demographics of the training and validation data sets is shown in Table 1.

Table 1.

Study Population for the Training and Validation Cohorts

	Training				Validation
	Overall (n=262)	Invasive assessment		Healthy control (n=161)	Overall (n=155)	Invasive assessment		Healthy control (n=80)
		LVEDP >18 mm Hg (n=44)	LVEDP ≤18 mm Hg (n=57)			LVEDP >18 mm Hg (n=40)	LVEDP ≤18 mm Hg (n=35)
Age, y, mean±SD	45±19	66±9	67±10	32±10	46±21	65±11	65±11	28±8
Male sex, n (%)	141 (54)	29 (66)	34 (60)	78 (48)	96 (62)	26 (65)	25 (71)	45 (56)
Race and ethnicity, n (%)
White	136 (52)	29 (66)	34 (60)	73 (45)	105 (68)	30 (75)	27 (77)	48 (60)
Asian	60 (23)	4 (9)	6 (11)	50 (31)	12 (8)	1 (3)	1 (3)	10 (13)
Black	34 (13)	11 (25)	16 (28)	7 (4)	20 (13)	7 (18)	5 (14)	8 (10)
Body mass index, kg/m2, mean±SD	26±5	30±5	28±5	25±4	27±6	32±6	27±3	25±5
Hypertension, n (%)	83 (32)	35 (80)	48 (84)	N/A	66 (43)	38 (95)	28 (80)	N/A
Hyperlipidemia, n (%)	75 (29)	32 (73)	43 (75)	N/A	62 (40)	35 (88)	27 (77)	N/A
Lung disease, n (%)	51 (19)	20 (45)	31 (54)	N/A	39 (25)	21 (53)	18 (51)	N/A
Diabetes, n (%)	37 (14)	19 (43)	18 (32)	N/A	27 (17)	12 (30)	15 (43)	N/A
Kidney disease, n (%)	13 (5)	5 (11)	8 (14)	N/A	12 (8)	5 (13)	7 (20)	N/A
Angina, n (%)	53 (20)	23 (52)	30 (53)	N/A	45 (29)	26 (65)	19 (54)	N/A
Myocardial infarction, n (%)	19 (7)	9 (20)	10 (18)	N/A	10 (6)	7 (18)	3 (9)	N/A
Coronary artery bypass grafting, n (%)	11 (4)	7 (16)	4 (7)	N/A	8 (5)	6 (15)	2 (6)	N/A
Congestive heart failure, n (%)	21 (8)	11 (25)	10 (18)	N/A	15 (10)	8 (20)	7 (20)	N/A

LVEDP indicates left ventricular end‐diastolic pressure; and N/A, not applicable.

A total of 180 features were extracted from the matched ECG and pulse cycles (see Data S1). After performing feature selection (Figure S1), a total of 3 features that describe the preload, afterload, and contractility were selected for the model. A leave‐one‐subject‐out cross‐validation was performed on the training data set, resulting in a sensitivity of 0.84 (95% CI, 0.70–0.93) and a specificity of 0.84 (95% CI, 0.79–0.89). The leave‐one‐subject‐out cross‐validation performance on the invasive cohort only yields a specificity of 0.67 (95% CI, 0.53–0.79). The results of both analyses are summarized in Table 2. The confusion matrix and the receiver operating characteristic curve with the area under the the receiver operating characteristic curve (AUROC) are shown in Figure 4A.

Table 2.

Model Performance for the Training and Validation Sets

		All patients (95% CI)	Invasive cohort (95% CI)
Sensitivity	Training	0.84 (0.70–0.93)	0.84 (0.70–0.93)
	Validation	0.80 (0.64–0.91)	0.80 (0.64–0.91)
Specificity	Training	0.84 (0.79–0.89)	0.67 (0.53–0.79)
	Validation	0.83 (0.75–0.90)	0.66 (0.48–0.81)

Figure 4. Receiving operator characteristic curves with area under the curve and confusion matrices for (A) the training data and (B) the validation data.

Model performance was assessed by applying the model on the holdout validation data set, resulting in a sensitivity of 0.80 (95% CI, 0.64–0.91) and a specificity of 0.83 (95% CI, 0.75–0.90). When the model performance was restricted to the invasive cohort, a specificity of 0.66 (95% CI, 0.48–0.81) was observed. A summary of these results is shown in Table 2. The resulting confusion matrix and the receiver operating characteristic curve with AUROC curve for the validation data set are shown in Figure 4B.

Discussion

Multiple investigations have been explored to evaluate noninvasive methods to estimate LVEDP. The VeriCor system ²¹ by CVP Diagnostics Inc (Boston, MA) estimates LVEDP by measuring the radial blood pressure and respiratory pressure during the Valsalva maneuver. This study reported a significant correlation with catheter‐measured LVEDP (r=0.86) and received Federal Drug Administration clearance in 2010. The use of a Valsalva maneuver, however, can put patients with unstable angina or uncontrolled hypertension at risk and may be difficult to consistently perform. In addition, there were excessive exclusion criteria applied (women excluded), thus limiting availability. Pahlevan et al ²² proposed a novel method called cardiac triangle mapping based on the electromechanical response of the left ventricle to compute LVEDP (r=0.76). This approach, however, requires a central blood pressure waveform, and the preservation of the relative timing between the QRS complex and the arterial pulse. Inherently, this limits this methodology to catheter‐based studies where LVEDP values could otherwise be directly measured or to waveforms captured via carotid tonometry, which requires substantial training. Fathieh et al proposed a method for predicting elevated LVEDP (≥20 mm Hg) using 3‐dimensional orthogonal voltage gradient and photoplethysmography. ²³ The 94‐feature model was trained on 690 patients and validated on 172 patients, which yielded an AUROC of 0.89. However, the ratio between the number of features and patients used for training suggests a reasonable chance of overfitting. More recently, Bhavnani et al applied a similar approach for predicting elevated LVEDP (≥25 mm Hg) also using orthogonal voltage gradient and photoplethysmography data. ²⁴ The model included over 100 features and was trained on 337 patients and validated on 79 patients, screening patients with LVEDP ≥25 versus ≤12 mm Hg, with AUROCs of 0.81 and 0.79, respectively.

When developing the model to estimate the risk of elevated LVEDP, priority was given to features that reflect physiological measurements related to preload, afterload, and/or contractility and therefore the performance of the left ventricle. Several recent studies have independently explored noninvasive estimation of preload, afterload, and contractility including using machine learning methodologies. ²⁴ , ²⁵ , ²⁶ , ²⁷ , ²⁸ These investigations have used a diverse array of sensing modalities, such as echocardiography and ultrasound, ²⁹ , ³⁰ , ³¹ arterial tonometry, ²⁵ , ²⁶ and seismocardiography. ³² , ³³ Similar to other studies involving noninvasive pulse waveforms, this study parameterized the data by calculating derivatives, integrals, and time intervals between fiducial markers on the 2 noninvasive signals acquired by the Vivio System. Training a model by leave‐one‐subject‐out cross‐validation produced an AUROC of 0.91 with a sensitivity of 0.84 and a specificity of 0.84. A validation data set containing 155 patients was created to achieve an overall power of 79% assuming a true sensitivity of the validation of 0.74 and true specificity of 0.80. The trained model was frozen and tested on the validation data set yielding an area under the curve of 0.85, with a sensitivity of 0.80 and a specificity of 0.83. These results suggest that the model is robust and performs excellently given that both sensitivities and specificities in training and validation are >0.80, and areas under the curve are >0.85. Negligible decreases in performance were observed when applying the validation data set; the sensitivity decreased by 0.04 and specificity decreased by 0.01.

Elevated cardiac filling pressures are one of the first signs that a patient is no longer hemodynamically stable and frequently precedes the onset of HF symptoms. ³⁴ , ³⁵ HF is frequently diagnosed at its later stages, with about 65% of patients initially presenting to the emergency department or requiring inpatient care. ³⁶ , ³⁷ Identifying HF is particularly important in high‐risk patients, such as those with diabetes or chronic kidney disease, while they are still in the primary care setting. The Vivio System can thus help identify patients with potential HF earlier in the progression of the disease, giving health care professionals more time to mitigate a patient’s risk by initiating medical therapy and earlier diagnostic testing. This early detection enables earlier intervention and the initiation of guideline‐directed medical therapy, which may improve patient outcomes. Furthermore, the use of serial measurements by the Vivio System may also enable titration of medical therapy, including diuretic therapy. Using LVEDP as the underlying diagnostic measurement suggests that the Vivio System may be able to screen patients for both HF with reduced ejection fraction and HF with preserved ejection fraction. In general, wider access to a noninvasive method to assess cardiac filling pressure would be expected to help prioritize patients in need of medical therapy and/or additional diagnostic testing. This approach has the potential to reduce delays in the diagnosis of HF and to improve in the missed diagnosis of HF. Future studies will evaluate the use of this device for the noninvasive diagnosis of HF with preserved ejection fraction and with longitudinal HF management.

Limitations

There were multiple limitations to the current study. Due to the potential risks of catheterization, invasive measurement of LVEDP was not performed in healthy participants and a nonelevated LVEDP was assumed. Referral of patients for cardiac catheterization and use of procedural sedation were not standardized across the clinical sites. This lack of standardization may have affected LVEDP measurements. Issues with signal and measurement quality resulted in the exclusion of data measurements, thus introducing potential bias. Another potential limitation of this study was the use of the R peak of the ECG rather than the “Z” point at the end of the “A” wave ³⁸ to determine the LVEDP value. In specific cases, this approach may overestimate the LVEDP value due to the rapid rise in pressure caused by the contraction of the left ventricle. However, the R peak was chosen to standardize the approach to LVEDP value determination across all patients since the A wave was not always present. The specificity in the invasive cohort being lower than the overall specificity was also a limitation. However, this was likely because there were other cardiovascular comorbidities present in patients from the invasive cohort with nonelevated LVEDP, which could impact the model. In addition, signal quality of both invasive and noninvasive measurements may differ, depending on the environment and operator, thus introducing additional bias.

Conclusions

This study demonstrates the capability of the Vivio System to identify patients with elevated LVEDP and has the potential to provide screening for asymptomatic or symptomatic patients in primary care, outpatient, or home care settings.

Sources of Funding

This study was funded by Ventric Health.

Disclosures

Dr Shavelle discloses grant support from Ventric Health, Penumbra, Neurescue, Inquis Medical, Alleviant Medical, V‐Wave Medical, and Cardiac Dimensions. Dr Rinderknecht discloses providing consultative services to and incentive compensation with Ventric Health. Dr Jin discloses providing consultative services to and incentive compensation with Ventric Health; Dr Chiu discloses providing consultative services to and incentive compensation with Ventric Health. A. Krupa is a paid employee and has incentive compensation with Ventric Health; Dr Jerdonek discloses providing consultative services to and incentive compensation with Ventric Health. K. Cook is a paid employee and has incentive compensation with Ventric Health; Dr Pahlevan holds equity in Avicena LLC (Ventric Health) and has a consulting agreement with Avicena LLC (Ventric Health). Dr Reeves has served as a consultant to Abbott and is a member of the Data and Safety Monitoring Board for the TWIST trial; Dr Madder has received speaker honoraria from Abbott Vascular, Boston Scientific, and Corindus; has served as a consultant to Abbott Vascular, Angiowave Imaging, Nanoflex Robotics, RapidAI, and Spectrawave; has received research support from Angiowave Imaging, Corindus, Microbot Medical, and Nanoflex Robotics; and serves on the advisory boards of Boston Scientific, Gentuity, Medtronic, and Spectrawave. The remaining authors have no disclosures to report.

Supporting information

Figure S1

Supplemental Methods

References 39 and 40