Artificial intelligence-guided detection of under-recognised cardiomyopathies on point-of-care cardiac ultrasonography: a multicentre study

Evangelos K. Oikonomou, Akhil Vaid, Gregory Holste, Andreas Coppi, Robert L. McNamara, Cristiana Baloescu, Harlan M. Krumholz, Zhangyang Wang, Donald J. Apakama, Girish N. Nadkarni, Rohan KheraThe Lancet Digital Health 2025 · DOI

TL;DR

This paper introduces a novel AI framework that can accurately detect under-diagnosed cardiomyopathies (hypertrophic cardiomyopathy and transthyretin amyloid cardiomyopathy) from low-quality, point-of-care cardiac ultrasound (POCUS) videos. This AI-driven approach has the potential to enable scalable, opportunistic screening, leading to earlier diagnosis and improved patient outcomes.

What Problem Does This Paper Address?

Point-of-care ultrasound (POCUS) is a powerful tool for bedside cardiac imaging, offering accessibility and portability. However, its utility for diagnosing complex conditions like cardiomyopathies is limited by abbreviated protocols, variable image quality, and operator experience. Traditional echocardiography, while more detailed, is less accessible and can suffer from referral and selection biases. Cardiomyopathies, particularly hypertrophic cardiomyopathy (HCM) and transthyretin amyloid cardiomyopathy (ATTR-CM), are often under-diagnosed due to their subtle or asymptomatic presentation, leading to delayed treatment and increased morbidity/mortality. This paper aims to bridge this gap by developing and validating an AI model that can reliably extract diagnostic information from POCUS, even from suboptimal acquisitions, thereby democratizing early detection.

Methods

The core of this research is a multi-label, view-agnostic, video-based deep learning algorithm. Key methodological innovations include:

  • POCUS-Adapted Architecture: The model was specifically designed to handle the challenges of POCUS, such as lower image quality and off-axis views, which are common in real-world clinical settings.
  • Augmentation Strategies: A combination of natural and synthetic data augmentation techniques were employed to simulate variations in view quality and orientation, making the model more robust.
  • View-Confidence Weighted Loss Function: The training process incorporated a loss function that assigns higher weights to inferences from views where the model has lower confidence in its anatomical identification. This helps the model learn from less ideal acquisitions.
  • View-Agnostic Approach: Instead of relying on specific views, the model was trained to process a pooled dataset of various views, allowing it to generalize better to the diverse POCUS acquisitions.
  • Multi-Label Classification: The AI model is capable of simultaneously identifying multiple conditions of interest: HCM, ATTR-CM, and severe aortic stenosis.

Datasets

The study utilized two primary datasets:

  • Development Cohort (Yale-New Haven Health System - YNHHS): This dataset comprised transthoracic echocardiographic (TTE) studies from five hospitals and affiliated clinics. It was used to train the deep learning algorithms.
  • Testing Cohorts (YNHHS and Mount Sinai Health System - MSHS): These datasets consisted of real-world POCUS videos acquired in the emergency departments of two large health systems. YNHHS served as an internal testing set, while MSHS provided an external validation set.

The study involved over 90,000 unique POCUS videos from these two health systems, representing a significant real-world benchmark for POCUS-based AI diagnostics.

Key Insights and Ideas

  • POCUS as a Screening Tool: The research strongly suggests that POCUS, when augmented with AI, can serve as a powerful and scalable tool for opportunistic screening of under-diagnosed cardiomyopathies, extending its utility beyond acute clinical decision-making.
  • Robustness to Low-Quality Data: The AI model's ability to perform well even with low-quality and off-axis POCUS views is a significant breakthrough, addressing a major limitation of current POCUS applications.
  • Early Detection Potential: The study found that positive AI-POCUS screens can occur years before confirmatory diagnosis, highlighting the potential for earlier intervention and improved prognosis.
  • Prognostic Value of AI Phenotypes: Even in individuals without a diagnosed cardiomyopathy, AI-identified phenotypes suggestive of HCM or ATTR-CM were associated with increased mortality risk, suggesting these AI markers can identify individuals at higher risk.
  • Democratizing Cardiac Care: The higher proportion of Black and Hispanic individuals in the POCUS cohort compared to the TTE cohort underscores the potential of AI-guided POCUS to improve access to advanced cardiac diagnostics in diverse populations.

Key Results

  • The AI classifier demonstrated strong performance in discriminating HCM and ATTR-CM from controls using POCUS videos. For example, in the YNHHS, the AUROC for HCM screening using apical-4-chamber views was 0.903, and for ATTR-CM using parasternal views, it was 0.907.
  • In YNHHS, 58% of HCM cases and 46% of ATTR-CM cases had a positive AI-POCUS screen a median of 2.1 and 1.9 years, respectively, before their eventual diagnosis.
  • Among individuals without known cardiomyopathy, those with the highest AI-derived probabilities for HCM or ATTR-CM phenotypes had a 17% and 32% higher adjusted risk of mortality, respectively.
  • The view-agnostic model outperformed view-specific models in POCUS cohorts, demonstrating its adaptability to varied acquisition quality.

Limitations and Future Work

  • Retrospective Design: The study was retrospective, analyzing clinically indicated scans. Future prospective studies are needed to validate these findings and investigate diagnostic performance and lead time in a prospective setting.
  • Under-diagnosis Challenge: The known under-diagnosis of HCM and ATTR-CM may lead to some false positives, a hypothesis supported by the observed lead times and adverse outcomes in some cases.
  • Model Not for Acute Management: The AI model is intended as a screening and risk stratification tool, not for immediate acute management decisions in the emergency department.
  • Saliency Maps for Isolated Cases: Saliency maps, used for explainability, are noted to perform poorly for complex phenotypes.
  • Outcomes Analysis Limitations: The outcomes analysis was limited to the YNHHS due to data availability, and potential confounding factors in the retrospective observational design cannot be entirely excluded.
  • Future Work: The authors suggest expanding POCUS libraries for in-domain finetuning of models, investigating the model's utility in prospective settings, and exploring its integration into clinical workflows for opportunistic screening and risk stratification. They also envision a vendor-agnostic, end-to-end pipeline for processing DICOM data and performing multi-label classification.

Summary updated

BibTeX
@inproceedings{oikonomou2025artificial,
  author = {Evangelos K. Oikonomou and Akhil Vaid and Gregory Holste and Andreas Coppi and Robert L. McNamara and Cristiana Baloescu and Harlan M. Krumholz and Zhangyang Wang and Donald J. Apakama and Girish N. Nadkarni and Rohan Khera},
  title = {{Artificial intelligence-guided detection of under-recognised cardiomyopathies on point-of-care cardiac ultrasonography: a multicentre study}},
  booktitle = {The Lancet Digital Health},
  year = {2025},
  doi = {10.1016/s2589-7500(24)00249-8},
}
Figure