Multimodal Latent Fusion of ECG Leads for Early Assessment of Pulmonary Hypertension

Introduction

This repository contains the PyTorch implementation of the LS-EMVAE framework. LS-EMVAE is a lead-specific electrocardiogram multimodal variational autoencoder model to pre-train on a large unlabeled 12L-ECG dataset and transfer its latent representations to fine-tune on smaller task-specific labeled 6L-ECG datasets. Unlike existing approaches that consider multi-lead ECG as a single modality, \textsc{LS-EMVAE} treats each lead as a separate modality. It uses a hierarchical modality expert (HiME) fusion mechanism comprising a mixture of experts and a product of experts to enable flexible, adaptive latent fusion. \textsc{LS-EMVAE} also introduces a latent representation alignment loss to improve coherence between individual leads and shared representations. We validate \textsc{LS-EMVAE} across two retrospective cohorts in a 6L-ECG setting: $892$ subjects from the ASPIRE registry for (1) PH detection and (2) phenotyping pre-/post-capillary PH, and $16,416$ subjects from UK Biobank for (3) predicting elevated pulmonary atrial wedge pressure, where it consistently outperforms baseline methods and demonstrates strong generalizability and interpretability.

Framework

System Requirements

The source code developed in Python 3.11 using PyTorch 2.3. LSEMVAE has been pre-trained on an NVDIA A100 NVLink 80GB GPU with 256GB RAM.

Datasets

The datasets for pre-training can be accessed through PhysioNet. MIMIC-IV-ECG is an open-access dataset, and you can find it here.

The in-house ASPIRE registry dataset contains sensitive patient data which cannot be made public because of the General Data Protection Regulation (GDPR).

References

[1] Wu, M., Goodman, N.: Multimodal generative models for scalable weaklysupervised learning. Advances in Neural Information Processing Systems 31 (2018)
[2] M. N. Suvon, P. C. Tripathi, W. Fan, S. Zhou, X. Liu, S. Alabed, V. Osmani, A. J. Swift, C. Chen, and H. Lu, “Multimodal variational autoencoder for low-cost cardiac hemodynamics instability detection,” in International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2024, pp. 296–306.
[3] Garg, P., Gosling, R., Swoboda, P., Jones, R., Rothman, A., Wild, J.M., Kiely, D.G., Condliffe, R., Alabed, S., Swift, A.J.: Cardiac magnetic resonance identifies raised left ventricular filling pressure: prognostic implications. European Heart Journal 43(26), 2511–2522 (2022)
[4] K. McKeen, L. Oliva, S. Masood, A. Toma, B. Rubin, and B. Wang, “Ecg-fm: An open electrocardiogram foundation model,” arXiv preprint arXiv:2408.05178, 2024.
[5] A. Das, W. Kong, R. Sen, and Y. Zhou, “A decoder-only foundation model for time-series forecasting,” in Forty-first International Confer- ence on Machine Learning, 2024.