EchoNext-Mini Open-Sourced for ECG-Based Heart Disease Detection

NewYork-Presbyterian Hospital has released EchoNext-Mini, an open-source artificial intelligence model and dataset designed to detect structural heart disease (SHD) from electrocardiograms (ECGs), according to a study published April 16 in NEJM AI. The model builds on prior work from researchers at the hospital and its academic affiliates, including Columbia University Irving Medical Center and Weill Cornell Medicine.

The EchoNext-Mini dataset includes 100,000 ECGs from 36,286 patients, paired with echocardiogram-derived diagnoses and demographic data. The dataset is a subset of a larger internal dataset used to train the original EchoNext model. Researchers made both the dataset and the model publicly available to support further development in ECG-based AI screening tools.

Structural heart disease, which affects the heart’s valves, walls, and chambers, is often underdiagnosed because screening depends on echocardiography, a resource-intensive imaging test that is not uniformly accessible. By contrast, ECGs are widely used and low-cost, but have historically not been sufficient for detecting structural abnormalities. The EchoNext line of models is designed to bridge that gap by using deep learning to infer structural disease from ECG waveforms.

Model Performance and Dataset Design

The EchoNext-Mini model uses the same architecture as the original EchoNext system, a deep learning model trained on ECG waveform data along with demographic and clinical inputs. It outputs predictions for 11 structural heart disease subdiagnoses, along with a composite classification.

In testing, the model achieved an area under the receiver operating characteristic curve (AUROC) of 82.0, compared with 85.2 for the larger EchoNext model. At a sensitivity threshold of 70%, the model reached a specificity of 77.9%. Performance across individual diagnoses ranged from AUROCs of 73.4 to 86.6, depending on the condition

The dataset includes cases where 52% of ECGs are positive for at least one form of structural heart disease. The most common conditions in the dataset include elevated left ventricular wall thickness, heart failure with reduced ejection fraction, and pulmonary hypertension. The data spans inpatient, outpatient, emergency, and procedural settings, and includes demographic diversity across patient populations.

The model is designed as a baseline system for benchmarking rather than a production-ready deployment tool. The open-source release allows researchers to test alternative architectures, training methods, and clinical applications using a standardized dataset.

Healthcare AI systems are increasingly evaluated based on measurable clinical performance rather than experimental novelty, as seen in prior deployments across operations and care delivery.

Open Dataset Signals Shift Toward Research Infrastructure

The release of EchoNext-Mini reflects a broader shift toward open datasets as infrastructure for medical AI development. In contrast to proprietary clinical models, publicly available datasets enable reproducibility and comparative benchmarking across research groups.

The EchoNext-Mini model follows a task-specific approach, focusing on ECG-based detection of structural heart disease rather than general-purpose clinical AI. This aligns with a growing body of work showing that specialized models often outperform broader systems in clinical settings.

The original EchoNext model was trained on more than 1 million ECG-echocardiogram pairs across multiple hospitals in the NewYork-Presbyterian system. That model demonstrated the ability to identify patients with structural heart disease using routine ECG data, including cases that may not have been flagged through standard clinical workflows.

By releasing a smaller, curated dataset and a corresponding model with similar architecture and performance characteristics, the researchers provide a foundation for further study without requiring access to large proprietary datasets.

Healthcare organizations are also expanding AI systems beyond single-use models into broader platforms that integrate care, operations, and research.

The EchoNext-Mini release focuses on the research layer of that stack, providing tools for model development and evaluation rather than direct clinical deployment.

The study authors stated that the dataset’s size, label diversity, and open availability are intended to support a wide range of future research into ECG-based detection and cardiovascular risk assessment.