Multimodal Cardiac Phenotype Estimation Using ECG and Clinical Data

Q: What is the clinical evidence from this study?

Study design: Observational. Population: Structural heart disease / cardiac phenotype estimation (n=82543). Intervention: MultiCard-PNet (Multimodal Cardiac Phenotype Prediction Network using 12-lead ECG and clinical data with Clinical-Guided Feature Gating) vs. Single-modality baselines (ECG-only, Clinical-only) and alternative fusion/deep learning models. Primary outcome: Mean Absolute Error (MAE) for Left Ventricular Ejection Fraction (LVEF) estimation on the held-out test set (MAE 7.179 (R² 0.395, Pearson r 0.645)).

Q: What does this research mean for the field?

A multimodal deep learning framework integrating 12-lead ECG and clinical data predicts quantitative cardiac echocardiographic phenotypes (LVEF, LVWT, and PASP) with lower error than single-modality models. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Kavitha Bai A. S.; J Somasekar

doi:10.14569/ijacsa.2026.0171060

January 1, 2026International Journal of Advanced Computer Science and ApplicationsOpen Access

Multimodal Cardiac Phenotype Estimation Using ECG and Clinical Data

View Full Paper

Ask AI

Bookmark

Share

Key result

MultiCard-PNet estimated cardiac phenotypes from 12-lead ECG and clinical data with test MAEs of 7.179% for LVEF, 0.168 cm for LVWT, and 9.537 mmHg for PASP, outperforming single-modality baselines.

Why the study?

Does the Multimodal Cardiac Phenotype Prediction Network (MultiCard-PNet) improve the estimation accuracy of LVEF, LVWT, and PASP from 12-lead ECG and clinical variables compared to single-modality baselines and standard fusion methods?

Key Points

To develop and evaluate a multimodal deep learning framework, MultiCard-PNet, for estimating quantitative echocardiographic phenotypes directly from 12-lead ECG signals and clinical data.
Designed MultiCard-PNet integrating a 1D-ResNet ECG encoder, a multi-layer perceptron clinical encoder, and a Clinical-Guided Feature Gating (CGFG) mechanism optimized with a weighted Huber multi-task loss.
Evaluated performance using the EchoNext dataset partitioned at the patient level across test subsets with valid reference labels (N=4,827 for LVEF, N=4,825 for LVWT, and N=2,621 for PASP).
MultiCard-PNet achieved an MAE of 7.1793 (RMSE 10.49, R² 0.3951, Pearson r = 0.6449) for LVEF, an MAE of 0.168271 (RMSE 0.244327, R² 0.0867, Pearson r = 0.3174) for LVWT, and an MAE of 9.5375 (RMSE 13.604, R² 0.1228, Pearson r = 0.3649) for PASP.
Multimodal CGFG integration consistently outperformed single-modality models, reducing MAE compared to ECG-only (LVEF MAE 8.222; LVWT MAE 0.175; PASP MAE 9.787) and clinical-only baselines (LVEF MAE 8.592; LVWT MAE 0.170; PASP MAE 9.932).
The proposed gating mechanism demonstrated lower prediction error and higher correlation than feature concatenation, FiLM fusion, and attention-based fusion across all three phenotypes.

Study Design

Type

Observational (n=82,543)

Multicenter

Yes

Structured PICO

Does the Multimodal Cardiac Phenotype Prediction Network (MultiCard-PNet) improve the estimation accuracy of LVEF, LVWT, and PASP from 12-lead ECG and clinical variables compared to single-modality baselines and standard fusion methods?

P

Population

82,543 active 12-lead ECG records (from a 100,000-record dataset) paired with clinical and echocardiographic data from patients at Columbia and Allen hospitals, evaluated across training (n=72,475), validation (n=4,626), and test (n=5,442) splits.

E

Exposure

Multimodal Cardiac Phenotype Prediction Network (MultiCard-PNet) integrating a 1D-ResNet-based 12-lead ECG encoder and a Multilayer Perceptron (MLP)-based clinical encoder (using age, sex, ventricular rate, atrial rate, PR interval, QRS duration, and QTc interval) with a Clinical-Guided Feature Gating (CGFG) mechanism and weighted Huber multi-task regression loss.

C

Comparator

Single-modality baselines (clinical-only and ECG-only models) and alternative multimodal fusion strategies (feature concatenation, FiLM fusion, and attention-based fusion).

O

Outcome

Quantitative estimation of echocardiographic cardiac phenotypes: Left Ventricular Ejection Fraction (LVEF, %), Left Ventricular Wall Thickness (LVWT, cm), and Pulmonary Artery Systolic Pressure (PASP, mmHg), evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), coefficient of determination (R²), and Pearson's correlation coefficient (r).surrogate

Main Result

Effect estimate: MAE 7.179 (R² 0.395, Pearson r 0.645)

Absolute Event Rate: 7.179% vs 8.222%

Limitations

Lack of paired acquisition timestamps in the dataset metadata prevented calculation of the exact time interval between ECG recordings and echocardiographic reference measurements.
Missing reference values for LVEF, LVWT, and PASP across ECG records required target-specific evaluation subsets (4,827 for LVEF, 4,825 for LVWT, and 2,621 for PASP in the test set).
Limited agreement and higher prediction uncertainty for extreme LVWT and PASP values, with a tendency to underestimate high PASP measurements.
A separate single-task versus multi-task ablation study was not performed to independently isolate the contribution of multi-task learning.
Additional validation using independent data splits and external datasets is required to confirm generalizability.
Upstream preprocessing details such as ECG sampling frequency, digital filtering methodology, and clinical feature imputation were not specified in the dataset loader.
A separate single-task versus multi-task ablation is required to determine the contribution of multi-task learning independently.
Limited agreement in prediction for extreme PASP and LVWT values, with a tendency to underestimate high PASP measurements.
Moderate correlation for LVWT and PASP reflects the difficulty of estimating structural and hemodynamic parameters from ECG-derived information.
Missing reference values for LVEF, LVWT, and PASP were not available for every ECG record.

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

Integrating 12-lead ECG waveforms with clinical and ECG-derived parameters via a clinical-guided feature gating mechanism improves non-invasive continuous estimation of LVEF, LVWT, and PASP compared to single-modality models.

Abstract

Existing ECG-based cardiac assessment methods primarily focus on disease classification, providing limited information about quantitative cardiac function, structural remodeling, and hemodynamic status. The complex relationship between ECG patterns and echocardiographic phenotypes remains insufficiently explored. This study proposes a Multimodal Cardiac Phenotype Prediction Network (MultiCard-PNet) for estimating clinically relevant cardiac phenotypes, including left ventricular ejection fraction (LVEF), left ventricular wall thickness (LVWT), and pulmonary artery systolic pressure (PASP), using 12-lead ECG signals and clinical information. The proposed framework integrates a 1D-ResNet-based ECG encoder for extracting temporal and morphological representations and an MLP-based clinical encoder for learning patient-specific characteristics. A Clinical-Guided Feature Gating (CGFG) mechanism adaptively recalibrates ECG features using clinical context before multimodal fusion. The fused representation is optimized through a multi-task regression framework with a weighted Huber loss for robust phenotype estimation. Evaluation on the EchoNext dataset demonstrated that the proposed MultiCard-PNet framework achieving a MAE value of 7.179 for LVEF, 0.168 for LVWT, and 9.537 for PASP. Although multimodal feature integration improved the reported metrics compared with single-modality baselines. The proposed MultiCard-PNet provides a non-invasive and comprehensive approach for ECG-based cardiac phenotype estimation, enabling improved cardiac assessment beyond conventional ECG interpretation and supporting future clinical decision-support applications.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

S. et al. (2026) conducted an observational in Structural heart disease / cardiac phenotype estimation (n=82,543). MultiCard-PNet (Multimodal Cardiac Phenotype Prediction Network using 12-lead ECG and clinical data with Clinical-Guided Feature Gating) vs. Single-modality baselines (ECG-only, Clinical-only) and alternative fusion/deep learning models was evaluated on Mean Absolute Error (MAE) for Left Ventricular Ejection Fraction (LVEF) estimation on the held-out test set (MAE 7.179 (R² 0.395, Pearson r 0.645)). MultiCard-PNet estimated cardiac phenotypes from 12-lead ECG and clinical data with test MAEs of 7.179% for LVEF, 0.168 cm for LVWT, and 9.537 mmHg for PASP, outperforming single-modality baselines.

synapsesocial.com/papers/6af0dba03c845ca7837cede0 https://doi.org/10.14569/ijacsa.2026.0171060

View Full Paper

Ask AI

Bookmark

Share

Multimodal Cardiac Phenotype Estimation Using ECG and Clinical Data

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Multimodal Cardiac Phenotype Estimation Using ECG and Clinical Data

Key Points

Study Design

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading