References and Resources

This challenge asked teams to identify potential Chagas disease cases from standard 12-lead ECG recordings, producing both a binary prediction and a probability; ECG screening is intended to help prioritize confirmatory serological testing rather than replace it.

An adjacent challenge series using pediatric respiratory recordings collected with digital stethoscopes. Tasks span sound classification, recording compression, and event detection, with released audio, annotations, and evaluation resources.

Participants developed open-source algorithms to reconstruct ECG time series from scanned or photographed paper ECGs and/or classify cardiac conditions from those images or reconstructed signals. Useful resources include ECG-Image-Kit for generating realistic synthetic ECG images and ECG-Image-Database, containing 35,595 images derived from 1,977 distinct ECG records.

Detect murmurs (present, absent, or unknown) and predict normal versus abnormal clinical outcomes using digital-stethoscope recordings from multiple auscultation locations plus routine demographic data. The public CirCor DigiScope training set contains 3,163 recordings from 942 patients, and the challenge provides open-source baselines, scoring code, and participant solutions.

A pediatric and young-adult phonocardiogram collection underlying the PhysioNet 2022 Challenge, with recordings from multiple auscultation locations, heart-cycle segmentation, detailed murmur annotations, demographics, and clinical-outcome labels. The full collection comprises 5,272 recordings from 1,568 participants. Of those, the public challenge training release represents 60% of the patients with the remainder reserved for validation and testing.

This collection combines 176 recordings acquired through the iStethoscope Pro smartphone app with 656 recordings from clinical DigiScope acquisition, including labeled training data and unlabeled challenge test data. Labels distinguish normal sounds, murmurs, extra heart sounds, artifacts, or extrasystoles depending on the subset. Selected normal recordings include S1/S2 segmentation annotations.

BMD-HS contains 864 twenty-second phonocardiogram recordings from 108 participants, including normal heart sounds and several valvular diseases. Echocardiogram-confirmed diagnoses and multi-disease annotations make it particularly relevant to disease-specific and multi-label classification rather than murmur detection alone.

The public challenge training set contains 3,126 heart-sound recordings collected in clinical and nonclinical settings from children and adults, with normal-versus-abnormal labels. Recordings span approximately 5–120 seconds and are supplied as WAV files resampled to 2,000 Hz, providing a benchmark for heart-sound classification across heterogeneous acquisition conditions.

A pediatric electronic-stethoscope dataset developed for normal-versus-pathological heart-sound classification and congenital heart disease research with diagnoses reviewed by cardiac experts. The study distinguishes clean, high-quality recordings from noisy, lower-quality recordings, supporting research into robustness under different acquisition conditions.

A compact collection of 1,000 WAV recordings distributed evenly across normal heart sounds, aortic stenosis, mitral regurgitation, mitral stenosis, and mitral valve prolapse. The repository also includes feature-extraction and classification code, making it useful for reproducible disease-classification experiments.

This open-access database contains 69 synchronized ECG and phonocardiogram recordings from 24 healthy young adults during rest and exercise, with recordings lasting either 30 seconds or 30 minutes. It supports multimodal timing, heart-cycle analysis, and signal-quality research, but its healthy-only population does not provide a disease-classification benchmark.

A large collection of approximately 21,800 ten-second, 12-lead ECGs from 18,869 patients, annotated with 71 diagnostic, rhythm, and waveform-related statements. Rich metadata and cardiologist-reviewed labels support multilabel cardiac classification and evaluation across patient subgroups.

A classic arrhythmia benchmark containing 48 half-hour, two-channel ambulatory ECG recordings from 47 subjects, with reference beat annotations. It is relevant to rhythm monitoring, heartbeat classification, and testing temporal models on longer recordings rather than short diagnostic ECG snapshots.

This open-access collection contains 12-lead ECGs from 45,152 patients, sampled at 500 Hz and labeled by clinical experts. It covers multiple common rhythms and additional cardiovascular conditions, supporting larger-scale diagnostic classification and external validation.

An openly available subset of the Clinical Outcomes in Digital Electrocardiography study, containing 345,779 ECG examinations from 233,770 patients alongside patient-level annotations. It is also a public training-data source for the PhysioNet 2025 Chagas challenge, where Chagas labels are self-reported rather than uniformly confirmed by serology. This is a 15% sample of the CODE dataset.

This public release provides first-examination, 12-lead ECG recordings from 1,631 participants in a chronic Chagas cardiomyopathy cohort, accompanied by age, mortality, and other annotations. It supports Chagas-related cardiac research and serves as a public training-data source for the PhysioNet 2025 Challenge.

This collection contains approximately 800,000 ten-second, 12-lead ECGs from nearly 160,000 patients, with links to the wider MIMIC-IV clinical database. It supports large-scale ECG modeling and research connecting electrical measurements with clinical history and outcomes; access is governed by PhysioNet’s credentialing and data-use requirements.

A companion dataset linking MIMIC-IV-ECG examinations to emergency-department and hospital-discharge diagnoses expressed as ICD-10-CM codes. It provides clinical prediction targets rather than additional waveforms and requires credentialed access, relevant training, and the associated MIMIC datasets.

A dataset designed to test and evaluate ECG digitization systems, containing diverse ECG images and their corresponding ECG information. Its metadata links each image to the associated reference information, supporting research into recovering usable cardiac signals from image-based records.

A downloadable ECG-image collection created with the Ch. Pervaiz Elahi Institute of Cardiology in Multan, Pakistan, for cardiovascular disease research. It offers an image-based complement to waveform datasets and is distributed under a CC BY 4.0 license.

A collection of 10,030 apical four-chamber echocardiography videos with expert left-ventricular tracings, volume measurements, and ejection-fraction labels. It supports cardiac-function estimation and ultrasound segmentation, complementing acoustic research with imaging-based measurements of ventricular function.

A collection of 7,643 labeled pediatric echocardiography videos with expert measurements, ventricular tracings, and cardiac-function annotations. Its pediatric population and inclusion of apical four-chamber and parasternal short-axis views make it relevant to childhood cardiac-function research alongside pediatric heart-sound datasets.

The Cardiac Acquisitions for Multi-structure Ultrasound Segmentation dataset contains apical two- and four-chamber ultrasound sequences from 500 patients, with expert reference contours and measurements. It supports segmentation of cardiac structures and estimation of ventricular volumes and function; the dataset remains downloadable although the online challenge evaluation platform has closed.

An adjacent digital-stethoscope dataset released with respiratory-sound challenges, with annotated recordings and challenge-specific test sets across multiple editions. Although it does not provide cardiovascular diagnoses, it is relevant to shared auscultation-audio methods such as event detection, compression, and classification.

This clinical validation study evaluated digital-stethoscope recordings from 962 patients against echocardiograms and expert annotations, achieving 76.3% sensitivity and 91.4% specificity for murmur detection. Sensitivity differed substantially by disease—93.2% for clinically significant aortic stenosis versus 66.2% for mitral regurgitation—illustrating why detecting a murmur is not equivalent to reliably excluding all valvular disease.

A semisupervised model used digital-stethoscope recordings to screen for elevated pulmonary artery systolic pressure, with echocardiography-derived estimates as reference labels. Testing in 196 patients produced sensitivity of 71% and specificity of 73%, supporting feasibility as a screening approach rather than a replacement for definitive pulmonary hypertension assessment.

This study combined deep learning with conventional machine learning to detect pathological murmurs in real-world cardiac acoustic recordings. It reported participant-level accuracy of 90.0%, sensitivity of 88.8%, and specificity of 91.2%, while separately examining recording-level performance and validation on PhysioNet data.

An exploratory study of 40 patients with chronic cardiovascular disease examined quantitative acoustic differences between consecutive heartbeats and their association with heart failure and NT-proBNP. It also found that body mass index affected sound intensity, highlighting a potential confounder for acoustic biomarkers and machine-learning models.

This study developed an explainable deep-learning framework using the HeartWave dataset, whose recordings were manually segmented into S1, systole, S2, and diastole. Its emphasis on clinically meaningful cardiac-cycle components makes it relevant to interpretable murmur classification and assessing whether models use plausible acoustic evidence.

This dataset paper describes multi-location pediatric heart-sound acquisition, cardiac-cycle annotations, and detailed murmur characteristics, establishing the foundation for the PhysioNet 2022 Challenge. It is particularly relevant to designing annotation schemes and distinguishing murmur detection from characterization and clinical-outcome prediction.

The challenge report describes tasks for detecting murmurs and predicting abnormal cardiac function from heart-sound recordings, together with the associated evaluation framework. It provides a reference for comparing algorithms under standardized conditions rather than relying on independently reported accuracy figures from incompatible datasets.

This review surveys deep-learning approaches to phonocardiogram segmentation and classification, including convolutional and recurrent architectures. It cautions that inconsistent evaluation procedures limit comparisons between studies, including reports of exceptionally high classification accuracy.

A broad review of machine-learning and deep-learning methods across heart-sound analysis subtasks, including murmur detection and diagnostic classification. It provides methodological background for developing automated auscultation systems and identifying research gaps across preprocessing, segmentation, and classification.

This prospective multicentre study evaluated 1,050 patients and found that AI applied to single-lead ECGs acquired with an ECG-enabled stethoscope could identify left-ventricular ejection fraction of 40% or less, with an AUROC of 0.85. Despite the stethoscope hardware, the model analyzed ECG rather than heart sounds, making this an important complementary—not acoustic—diagnostic approach.

This pragmatic trial randomized 205 primary-care practices to routine care or implementation of a stethoscope combining ECG and phonocardiogram algorithms for reduced ejection fraction, atrial fibrillation, and valvular disease. The intention-to-treat analysis found no significant increase in heart-failure detection after 12 months, providing an important distinction between diagnostic model performance and demonstrated benefit from real-world implementation.

This systematic review examines cardiac acoustic biomarkers for heart-failure detection and longitudinal monitoring, particularly the third heart sound and electromechanical activation time. It finds promising associations with disease severity and clinical events while emphasizing the need for further validation of predictive performance across patient populations.

This randomized study of 225 patients tested post-discharge management guided by electromechanical activation time, an acoustic-cardiography measure combining ECG timing and heart sounds, against conventional symptom-guided treatment. It reported a lower risk of the composite of heart-failure rehospitalization and mortality in the acoustic-guided group, making it directly relevant to sound-informed treatment rather than diagnosis alone.

This research paper introduces 35,595 software-labeled ECG images incorporating imaging and scanning distortions for evaluating digitization and analysis methods. It supports the related problem of extracting diagnostic cardiac signals from paper or photographed ECGs rather than directly acquired waveforms or sounds.

In 4,744 patients with symptomatic heart failure and reduced ejection fraction, dapagliflozin reduced worsening heart failure or cardiovascular death compared with placebo, irrespective of diabetes status. The primary outcome occurred in 16.3% versus 21.2% over a median 18.2 months, providing treatment context for research aimed at earlier detection of systolic dysfunction.

This randomized trial enrolled 6,263 patients with heart failure and ejection fraction above 40%, finding that dapagliflozin reduced the combined risk of worsening heart failure or cardiovascular death. It complements reduced-ejection-fraction studies and highlights that clinically important heart failure extends beyond the target population of many ECG-based screening algorithms.

Among 9,361 adults at increased cardiovascular risk without diabetes or previous stroke, an intensive systolic blood-pressure target below 120 mm Hg reduced major cardiovascular events and mortality compared with a target below 140 mm Hg. Serious adverse events, including hypotension, electrolyte abnormalities, and acute kidney injury, were more frequent with intensive treatment, documenting the benefit–harm trade-off in preventive management.

This corrected and republished analysis of 7,447 high-risk adults without established cardiovascular disease found fewer major cardiovascular events with Mediterranean diets supplemented by olive oil or nuts than with advice to reduce dietary fat. The article addresses departures from the randomization protocol and reports analyses accounting for them, making the 2018 publication preferable to the superseded original report.

In 17,604 adults with overweight or obesity and established cardiovascular disease but no diabetes, semaglutide reduced cardiovascular death, nonfatal myocardial infarction, or nonfatal stroke by 20% relative to placebo. This is secondary-prevention evidence in an established-disease population, not evidence that the treatment prevents first cardiovascular events in otherwise disease-free adults.