Determination of cardiac output (CO) is essential to the clinical management of cardiovascular compromise. However, the invasiveness, procedural risks, and reliance on specialized infrastructure limit accessibility and scalability of standard-of-care right heart catheterization (RHC). Seismocardiography (SCG), a non-invasive technique which records subtle chest wall vibrations generated by cardiac mechanical activity, may offer a promising alternative for CO determination. To explore this potential, we developed and evaluated a deep learning model for estimating CO directly from SCG, electrocardiogram (ECG), and body mass index (BMI) in heart failure patients undergoing RHC. We trained a deep convolutional neural network for CO estimation using an open-access dataset comprising 73 heart failure patients with simultaneous RHC, SCG, and ECG recordings. Model performance was evaluated on 64 patients using pairwise nested leave-pair-out cross-validation. When estimating CO in patients with a reference output < 6 L/min, the deep learning model achieved a mean bias of −0.01 L/min with LoA from −0.88 to 0.87 L/min. When predicting cardiac index in patients with a reference index < 2.2 L/min/m 2, the model yielded a mean bias of 0.07 L/min/m 2 with LoA from −0.35 to 0.48 L/min/m 2. This study demonstrates the feasibility of using deep learning in combination with wearable SCG sensors to non-invasively estimate CO. Model performance was particularly strong in low-output states. These findings highlight the potential of SCG-based monitoring to augment clinical decision-making in settings where invasive measurements are impractical or unavailable. Prospective multicenter validation is needed to confirm generalizability and assess clinical impact.
Accurate bedside assessment of cardiac output (CO) is fundamental to the diagnosis and management of cardiovascular compromise and plays a central role in clinical decisions involving inotropic medication, mechanical circulatory support, and treatment escalation. ,, Despite the clinical value of CO measurement, its dependence on invasive right heart catheterization (RHC)—which is associated with procedural risks while requiring skilled operators and advanced instrumentation—limits the availability of CO as a clinical tool. In fact, a recent prospective multicenter study reported that cardiogenic shock patients initially evaluated at facilities without invasive monitoring capabilities had higher mortality rates compared to patients who were directly admitted to tertiary care centers. Although echocardiography and magnetic resonance imaging (MRI) can provide certain noninvasive measures of CO, both modalities have limitations: echocardiography relies on geometric assumptions and consistent image quality, , while MRI may be restricted by high cost and uneven geographic availability. , These findings underscore the need for accurate, accessible, and noninvasive solutions to expand timely hemodynamic evaluation in settings where RHC is unavailable or contraindicated.
Recent evidence indicates that seismocardiography (SCG) may enable a scalable, noninvasive estimation of key hemodynamic parameters. ,,, Seismocardiography employs multiaxial accelerometers affixed to the chest wall to capture microvibrations of the thoracic surface arising from the mechanical activity of the heart, including myocardial contraction, valvular motion, and blood flow dynamics. Prior studies have suggested the potential of SCG-based models to predict Doppler- and MRI-derived stroke volume (SV) in perioperative and pediatric cohorts. , However, it is unknown whether SCG can be leveraged for the determination of CO in adults with heart failure.
In this study, we address this gap by evaluating SCG in a publicly available adult heart failure cohort undergoing invasive hemodynamic evaluation. We developed and validated a deep learning model to estimate CO directly using SCG, electrocardiogram (ECG), and body mass index (BMI) as inputs ( Central Illlustration ). All reference CO values were obtained from RHC, the clinical standard of care.
We describe a novel algorithm utilizing (1 and 2) publicly available wearable patch-derived electrocardiographic (ECG) and triaxial seismocardiographic (SCG) signals, combined with body mass index (BMI), in a (3) deep learning model to (4) predict cardiac output in heart failure.
Methods
Dataset and code availability . The dataset used in this study is publicly available on the PhysioNet repository at https://physionet.org/content/scg-rhc-wearable-database/1.0.0/13 . The dataset was collected under a protocol reviewed and approved by the University of CA San Francisco (UCSF) Institutional Review Boards (IRB number: 16-20442) on December 20, 2016. Patients were recruited from the catheterization laboratory at UCSF and all patients provided written consent. The code used for model development and analysis is publicly available at https://github.com/jwang6174/scg-co .
Study population . The deidentified, open-source dataset released by Chan et al contained wearable patch signals and RHC measurements for 73 patients referred for hemodynamic evaluation for primary diagnosis of heart failure. , As nineteen patients had multiple encounters, a total of 83 encounters were included. Each encounter included simultaneously recorded SCG, ECG, and RHC signals. Only baseline hemodynamic parameters before physiologic challenge were analyzed. Patient demographics, clinical characteristics, hemodynamic parameters, and patch signals were collected at the time of RHC. CO was calculated as the average of Fick and thermodilution (TD) when both were available ( n = 75 encounters) and was otherwise considered to be the sole Fick ( n = 1 encounter) or TD ( n = 7 encounters) measurement. Additional details on data preprocessing and signal augmentation are provided in the Supplemental Methods .
Model design and implementation. Triaxial SCG signals were used to capture interaxis mechanical motion and improve robustness to variability in sensor placement. A single-lead ECG, obtained from the wearable patch, was incorporated to provide precise temporal alignment with cardiac events and to enhance model learning of the electromechanical relationship. Prior work has shown improved SV prediction with combined SCG and ECG inputs. BMI was included as an additional input to account for interindividual differences in signal attenuation related to body habitus. All features were fed into a deep convolutional neural network model. The model was implemented using the PyTorch 2.5.1 programming library. Model training was performed on hardware comprising an Intel Core i7-4790K processor, NVIDIA GeForce GTX 1080 Ti GPU, and 32 GB system memory. Additional details on training and validation strategy are provided in the Supplemental Methods .
Evaluation metrics and statistical analysis . Model performance was evaluated using encounter-level , target-matched leave-pair-out cross-validation (LPOCV) with a nested design. Root mean squared error (RMSE), relative error, and Pearson correlation coefficient (PCC) were calculated using averaged model predictions for each catheterization encounter. To quantify uncertainty in RMSE, nonparametric bootstrapping was performed by resampling 10,000 bootstrap replicates with replacement from the full set of reference–predicted CO pairs. The 95% confidence interval (CInt) was defined as the 2.5th and 97.5th percentiles of the resulting RMSE distribution. Relative error was expressed as RMSE% and calculated as the RMSE divided by the mean of the reference values then multiplied by 100. For PCC, the 95% CInt was computed via Fisher’s z-transformation, with percentile bounds of the transformed distribution inverse-transformed to the correlation scale. Bland-Altman analysis was conducted to assess agreement between predicted and reference CO values, with mean bias and 95% limits of agreement (LoA). Summary metrics were also calculated across stratified ranges of CO and CI. Statistical analyses were performed with the SciPy 1.14.1 programming library.
Results
Demographic and hemodynamic profile of a high-risk heart failure cohort . We first examined patient characteristics to understand how the study population compares with the broader heart failure population. Patient demographic and hemodynamic characteristics were summarized at the encounter level ( Table 1 ). The mean age of the cohort was 54.9 ± 13.3 years, and 66.3% of participants were male. The average BMI was 29.2 ± 6.6 kg/m², and the average body surface area (BSA) was 2.0 ± 0.3 m 2. Among patients in the cohort, 19.3% were NY Heart Association (NYHA) class II, 62.7% were NYHA class III, 79.5% had heart failure with reduced ejection fraction (HFrEF), and 18.0% had heart failure with preserved ejection fraction (HFpEF). The average heart rate (HR) was 74.3 ± 16.5 bpm. Cardiac rhythms discerned from ECG recordings were 61.8% sinus rhythm, 9.6% atrial fibrillation, 10.8% dual AV paced, and 14.5% V paced.
Table 1
Overview of patient characteristics and hemodynamics.
|
All
( n = 83) |
Validation-test set ( n = 64) |
Training-only set
( n = 19) |
|
|---|---|---|---|
| Age (years) | 54.9 ± 13.3 | 55.9 ± 13.8 | 51.6 ± 10.9 |
| Sex | |||
| Male | 55 (66.3%) | 45 (70.3%) | 10 (52.6%) |
| Female | 28 (33.7%) | 19 (29.7%) | 9 (47.4%) |
| BMI (kg/m 2) | 29.2 ± 6.6 | 29.9 ± 7.0 | 26.9 ± 4.2 |
| BSA (m 2) | 2.0 ± 0.3 | 2.1 ± 0.3 | 1.9 ± 0.2 |
| NYHA | |||
| I | 6 (7.2%) | 6 (9.4%) | 0 (0%) |
| II | 16 (19.3%) | 14 (21.9%) | 2 (10.5%) |
| III | 52 (62.7%) | 40 (62.5%) | 12 (63.2%) |
| IV | 6 (7.2%) | 4 (6.2%) | 2 (10.5%) |
| Not reported | 3 (3.6%) | 0 (0%) | 3 (15.8%) |
| Outpatient | |||
| Yes | 67 (80.7%) | 52 (81.2%) | 15 (78.9%) |
| No | 16 (19.3%) | 12 (18.8%) | 4 (21.1%) |
| EF | |||
| Preserved | 15 (18.0%) | 11 (17.2%) | 4 (21.1%) |
| Reduced | 66 (79.5%) | 51 (79.7%) | 15 (78.9%) |
| Not reported | 2 (2.4%) | 2 (3.1%) | 0 (0%) |
| Other history | |||
| CABG | 10 (12.0%) | 10 (15.6%) | 0 (0%) |
| CRT-D | 24 (28.9%) | 20 (31.2%) | 4 (21.1%) |
| ICD | 25 (30.1%) | 19 (29.7%) | 6 (31.6%) |
| OHTx | 6 (7.2%) | 2 (3.1%) | 4 (21.1%) |
| PCI | 9 (10.8%) | 7 (10.9%) | 2 (10.5%) |
| VAD | 2 (2.4%) | 0 (0%) | 2 (10.5%) |
| RAP (mmHg) | 8.2 ± 5.3 | 8.7 ± 5.4 | 6.3 ± 4.8 |
| mPAP (mmHg) | 28.6 ± 10.6 | 30.1 ± 10.9 | 23.3 ± 7.5 |
| PAWP (mmHg) | 17.1 ± 7.7 | 17.7 ± 8.0 | 14.7 ± 6.1 |
| HR (bpm) | 74.3 ± 16.5 | 74.4 ± 15.8 | 74.1 ± 19.2 |
| Rhythm | |||
| Sinus rhythm | 51 (61.4%) | 39 (60.9%) | 12 (63.2%) |
| Atrial fibrillation/flutter | 8 (9.6%) | 6 (9.4%) | 2 (10.5%) |
| Paced rhythm | 24 (28.9%) | 19 (29.7%) | 5 (26.3%) |
| A-Paced | 1 (1.2%) | 1 (1.6%) | 0 (0%) |
| AV-paced | 9 (10.8%) | 7 (10.9%) | 2 (10.5%) |
| V-paced | 14 (16.9%) | 11 (17.2%) | 3 (5.3%) |
| SV (ml) | 64.8 ± 20.7 | 64.6 ± 21.1 | 65.6 ± 19.7 |
| CO (L/min) | 4.7 ± 1.7 | 4.6 ± 1.4 | 5.0 ± 2.4 |
| CI (L/min/m 2) | 2.3 ± 0.9 | 2.3 ± 0.6 | 2.6 ± 1.5 |
Values summarized at the encounter level and presented as mean ± SD for normally distributed values and n (%) for counts.
BMI = body mass index; BSA = body surface area; CABG = coronary artery bypass graft; CI = cardiac index; CO = cardiac output; CRT = cardiac resynchronization therapy; EF = ejection fraction; HR = heart rate; ICD = implantable cardioverter defibrillator; mPAP = mean pulmonary artery pressure; NYHA = New York heart association; OHTx = orthotic heart transplant; PAWP = pulmonary artery wedge pressure; PCI = percutaneous coronary intervention; RAP = right atrial pressure; SV = stroke volume.
Demographic, clinical, and hemodynamic characteristics for all catheterization encounters, validation-test patients, and training-only patients.
Participants had an average CO of 4.7 ± 1.7 L/min and an average cardiac index (CI) of 2.3 ± 0.9 L/min/m 2. The distributions of CO and CI values were approximately normal ( Figure 1 A and Figure 1 B). An outlier, defined as greater than three standard deviations from the mean, was observed with Fick CO of 9.4 L/min and TD CO of 17.8 L/min, resulting in an average CO of 13.6 L/min. For encounters where both measurements were performed, Fick and TD were strongly correlated with a PCC of 0.84 (95% CInt: 0.76–0.90, p <0.001, Figure 1 C). Two RHC encounters were performed while the patient was supported by a ventricular assistive device (VAD) and were both included in the training-only set. Overall, the cohort reflected a high-risk heart failure population characterized by advanced functional limitation, a substantial burden of pacing dependence, and reduced CI.
