AI-Powered ECG Monitoring Could Transform Heart Failure Detection

By Soham NandiReviewed by Bethan DaviesFeb 20 2025

A new artificial intelligence-powered model could change the way heart failure is monitored, offering a non-invasive solution for detecting elevated left atrial pressure (LAP). Researchers have developed a deep neural network that analyzes single-lead electrocardiogram (ECG) data from wearable patch monitors, demonstrating strong performance across multiple validation datasets. The findings, published in Nature, highlight the model’s potential for outpatient and home-based cardiac monitoring.

Background

Heart failure is a serious condition with high rates of hospital readmission, morbidity, and mortality. Elevated LAP, often measured via mean pulmonary capillary wedge pressure (mPCWP), is a key indicator of worsening HF.

The current gold standard for assessing LAP, right heart catheterization (RHC), is an invasive hospital-based procedure. Non-invasive alternatives, such as cardiac Doppler ultrasound, require skilled operators, while clinical assessments alone can be unreliable. Implantable devices like the CardioMEMS HF system provide remote monitoring but require surgical implantation, posing risks.

While artificial intelligence (AI) and machine learning have been successfully used to analyze single-lead ECG data for arrhythmias and other cardiac issues, their potential in LAP estimation remains largely unexplored. This study aimed to fill that gap by developing a deep neural network model that utilizes wearable ECG data to detect elevated LAP non-invasively, making it a promising tool for early intervention and reducing avoidable hospitalizations.

Methods and Data Acquisition

The researchers developed and validated a machine learning model called the Cardiac Hemodynamic AI Monitoring System (CHAIS) to estimate hemodynamic parameters, specifically detecting mPCWP greater than 18 mmHg. The study incorporated both retrospective and prospective ECG data to ensure robustness.

For retrospective data, the researchers gathered 6739 samples from Massachusetts General Hospital (MGH) and 4620 samples from Brigham and Women’s Hospital (BWH). Each cardiac catheterization record was paired with a same-day 12-lead ECG. The model underwent pre-training using a large dataset of 242,216 MGH patients, where it learned to predict PR, QRS, QT intervals, and heart rate from 12-lead ECGs before being fine-tuned for single-lead ECG data.

The dataset was then split into three groups: development (80 %), internal-holdout (20 %), and external-validation, with BWH samples serving as the external validation set. For prospective data, ECGs were collected from MGH inpatients using a QOCA Portable ECG patch-monitor, recorded within 24 hours before cardiac catheterization. High-quality 10-second ECG segments were selected for analysis. The pre-trained CHAIS model was directly applied to the patch-monitor data using zero-shot transfer learning, eliminating the need for additional fine-tuning.

Performance was then assessed through AUROC, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Additionally, an entropy-based trustworthiness analysis was conducted to identify uncertain predictions, enhancing the reliability of the model’s assessments.

Findings

The CHAIS model was evaluated across four datasets: an internal-holdout set from MGH, an external validation set from BWH, and prospective patch-monitor data from MGH patients.

• Internal-holdout dataset: CHAIS achieved an AUROC of 0.80 for detecting elevated mPCWP. Predictions with lower entropy were more reliable (AUROC 0.80), while high-entropy predictions had reduced accuracy (AUROC 0.52). When using sensitivity thresholds of 70 % and 80 %, specificities were 75 % and 66 %, respectively. The model also demonstrated NPVs exceeding 95 % for low pre-test probabilities.
• External validation dataset: The model maintained strong performance with an AUROC of 0.76. Consistent with internal results, low-entropy predictions performed better (AUROC 0.76) than high-entropy ones (AUROC 0.52). Sensitivities of 55 % and 68 % corresponded to specificities of 82.3 % and 75 %, respectively, with NPVs surpassing 90 % for low pre-test probabilities.
• Prospective patch-monitor data: Performance improved when the ECG recording occurred closer to the RHC procedure. The highest AUROC (0.875) was observed when the ECG was recorded 1 hour and 25 minutes before RHC.

Further subgroup analysis showed slightly enhanced performance in heart failure and transplant patients. Additionally, CHAIS predictions reflected intra-patient trends in mPCWP changes over time, reinforcing its potential for continuous monitoring.

Conclusion

The CHAIS model demonstrated strong performance in detecting elevated LAP using single-lead ECG data from wearable patch monitors. Its ability to provide non-invasive, real-time monitoring makes it a promising tool for the early detection of worsening heart failure in outpatient and home settings. The model’s high NPV and trustworthiness analysis enhance its clinical utility, potentially reducing hospital admissions through early intervention.

Future research should focus on expanding its applicability across broader patient populations and integrating it into remote cardiac care systems to improve heart failure management.

Journal Reference

Schlesinger, D. E., Alam, R., Ringel, R., Pomerantsev, E., Srikanth Devireddy, Shah, P., Garasic, J., & Stultz, C. M. (2025). Artificial intelligence for hemodynamic monitoring with a wearable electrocardiogram monitor. Communications Medicine, 5(1). DOI:10.1038/s43856-024-00730-5 ‌https://www.nature.com/articles/s43856-024-00730-5

Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.