Source-linked AI summary
The CirCor DigiScope Dataset: From Murmur Detection to Murmur Classification
Jorge Oliveira, Francesco Renna, Paulo Dias Costa, Marcelo Nogueira, Cristina Oliveira, Carlos Ferreira, Alipio Jorge, Sandra Mattos, Thamine Hatem, Thiago Tavares, Andoni Elola, Ali Bahrami Rad, Reza Sameni, Gari D Clifford, Miguel T. Coimbra
TL;DR
Cardiac auscultation systems are limited by publicly available datasets that provide only binary normal-versus-abnormal labels, despite auscultation’s role as a cost-effective screening tool. This paper presents a large pediatric heart sound dataset with detailed murmur annotations across recordings and auscultation locations, supporting research on murmur detection and analysis.
Problem
Publicly available heart sound datasets largely provide only binary normal-versus-abnormal labels, limiting detailed murmur analysis despite auscultation’s cost-effective screening role.
Method
The study constructs a pediatric heart sound dataset with multi-location recordings, manually segmented sounds, and expert murmur annotations covering timing, pitch, grading, shape, quality, and location.
Results
5282 recordings from 1568 patients and 215,780 manually segmented heart sounds form a detailed murmur characterization and classification database.
Takeaways & Limitations
The dataset supports future machine-learning research on detecting and analyzing murmur waves for diagnostic purposes across varied target populations.
Abstract
from arXiv · showhide
Cardiac auscultation is one of the most cost-effective techniques used to detect and identify many heart conditions. Computer-assisted decision systems based on auscultation can support physicians in their decisions. Unfortunately, the application of such systems in clinical trials is still minimal since most of them only aim to detect the presence of extra or abnormal waves in the phonocardiogram signal, i.e., only a binary ground truth variable (normal vs abnormal) is provided. This is mainly due to the lack of large publicly available datasets, where a more detailed description of such abnormal waves (e.g., cardiac murmurs) exists. To pave the way to more effective research on healthcare recommendation systems based on auscultation, our team has prepared the currently largest pediatric heart sound dataset. A total of 5282 recordings have been collected from the four main auscultation locations of 1568 patients, in the process, 215780 heart sounds have been manually annotated. Furthermore, and for the first time, each cardiac murmur has been manually annotated by an expert annotator according to its timing, shape, pitch, grading, and quality. In addition, the auscultation locations where the murmur is present were identified as well as the auscultation location where the murmur is detected more intensively. Such detailed description for a relatively large number of heart sounds may pave the way for new machine learning algorithms with a real-world application for the detection and analysis of murmur waves for diagnostic purposes.
I. INTRODUCTION
Cardiac auscultation is a cost-effective first-line screening tool, but computer-assisted systems are constrained by limited richly annotated public datasets. The presented dataset addresses this gap with detailed murmur characterization and recordings from multiple auscultation locations.
- 31% of all deaths globally are attributed to cardiovascular disease, which also increases morbidity, disability, hospital admissions, and healthcare burdens.
- Cardiac auscultation remains an important, cost-effective first-line screening tool, although it requires extensive training and clinical experience.
- Computer-aided auscultation systems are motivated by limited access to trained professionals and resources for timely diagnosis and referral.
- Large annotated heart-sound datasets are needed to represent and characterize murmurs and anomalies for modern machine-learning methods.
- Existing public phonocardiogram datasets commonly provide only normal-versus-abnormal labels or abnormal-sound presence, rather than full sound-signature characterization.
- The presented dataset collects sounds from pediatric screening campaigns and characterizes anomalies using clinical timing, pitch, grading, shape, quality, and multiple auscultation locations.
II. BACKGROUND
Cardiac auscultation interprets heart sounds generated by cardiac structures and blood flow, using standardized chest locations aligned with valve anatomy. Recordings distinguish fundamental sounds and cardiac-cycle periods.
- Normal heart sounds reflect valve vibrations, blood-flow turbulence, and the anatomical relationship between valves and the chest wall.
- The aortic, pulmonary, tricuspid, and mitral valves are auscultated at four standardized chest locations.
- The first heart sound S1 arises mainly from mitral and tricuspid valve closure at the beginning of systole.
- Figure 2 identifies S1 and S2 within a normalized recording and marks systolic and diastolic periods.
B. Clinical Applications
Additional heart sounds such as murmurs are associated with turbulent flow and specific cardiac conditions. Their timing and auscultation location provide clinical clues for identifying valve disease and related abnormalities.
- Murmurs, clicks, and snaps are additional sounds associated with turbulent blood flow and rapid movements of cardiac structures.
- Aortic stenosis produces a harsh crescendo-decrescendo systolic murmur best heard at the right upper sternal border, often radiating to the carotids.
- Aortic regurgitation produces a decrescendo-blowing diastolic murmur best heard at the left lower sternal border.
- Mitral stenosis produces a diastolic murmur best heard at the cardiac apex, where blood passage from the left atrium to the left ventricle is impeded.
- Mitral regurgitation produces a systolic murmur best heard at the apex and radiating to the left axilla because blood flows back into the left atrium.
4) Mitral regurgitation:
Several structural abnormalities have characteristic auscultatory findings, including clicks and murmurs associated with mitral valve prolapse, pulmonary stenosis, tricuspid disease, and septal defects.
- 4) Mitral regurgitation:: Mitral valve prolapse causes an early systolic click at the apex, often followed by a late systolic murmur.
- 4) Mitral regurgitation:: Pulmonary stenosis produces a crescendo-decrescendo systolic ejection murmur loudest at the upper left sternal border.
- 4) Mitral regurgitation:: Tricuspid stenosis produces a diastolic murmur best heard at the left lower sternal border.
- 4) Mitral regurgitation:: Tricuspid regurgitation produces a systolic murmur best heard at the left lower sternal border.
- 4) Mitral regurgitation:: Ventricular septal defects often generate holosystolic murmurs, whereas atrial septal defects can produce a fixed split S2.
9) Septal defects:
The reviewed open datasets vary in population, recording characteristics, and annotation scope, with several databases documenting heart-sound classes or fundamental sounds. Selection for analysis required public accessibility, study-population information, and recording details.
- The dataset review selected six open-access datasets using availability, accessibility, relevance, population information, and recording information criteria.
- One database contains 656 recordings from an unknown number of patients, sampled at 4000 Hz for durations of 1–30 seconds.Its two datasets were collected through a smartphone application and a digital stethoscope system.
- Its recordings were classified into normal, murmur, extra-heart-sound, artifact, or extra-systole classes, with fundamental heart sounds manually annotated in Dataset B.
- The PhysioNet/CinC 2016 dataset contains 2435 records from 1297 patients, collected at four auscultation locations and downsampled to 2000 Hz.
C. HSCT-11 (2016) [29]
HSCT-11 is described alongside several publicly available heart-sound resources, including a multi-database collection, pediatric recordings, and fetal heart sounds. These datasets differ substantially in size, population, recording conditions, and annotation scope.
- The biometric heart-sound dataset contains 412 recordings from 206 patients across mitral, pulmonary, aortic, and tricuspid locations.It was recorded with a ThinkLabs Rhythm digital electronic stethoscope at 11025 Hz and 16-bit resolution.
- The PhysioNet/CinC release officially contains 4430 recordings, differing from an earlier reported count because 338 normal-subject recordings were subdivided.
- A pediatric dataset contains 29 recordings from 29 patients aged six months to 17 years, collected at the mitral point at 4 kHz.Two cardiac physiologists manually annotated the beginning and ending of each fundamental heart sound.
- A fetal-heart-sound dataset contains 26 recordings from healthy pregnant women aged 25–35 years, sampled at 333 Hz with 8-bit resolution.
F. EPHNOGRAM: A Simultaneous Electrocardiogram and Phonocardiogram Database [32], [35]
The paper describes the screening campaigns, participant eligibility, clinical workflow, and resulting data organization. Participants underwent clinical assessment followed by electronic auscultation at four typical recording sites.
- The dataset was collected during two mass screening campaigns in Paraíba, Brazil, conducted in July–August 2014 and June–July 2015.The campaigns were named CC2014 and CC2015 and received institutional review-board approval.
- The CC2014 and CC2015 campaigns covered distinct sets of municipalities, with Santa Rita listed for CC2014 and Mamangrape listed for CC2015.
- Participants younger than 21 years with parental consent where appropriate were eligible; 2061 attended, 493 were excluded, and 116 attended both campaigns.
- Each participant completed questionnaires, clinical and nursing assessments, and cardiac investigations before electronic auscultation recordings were collected.Recordings were obtained from four typical auscultation spots by the same operator in a real clinical setting.
B. Demographic and Clinical Information
The dataset comprises a predominantly pediatric screening population with broadly balanced sex representation and generally favorable clinical status. Screening nevertheless identified diverse cardiopathies and generated substantial follow-up and intervention referrals.
- 988 children (63.0%) and 311 infants (19.8%) formed the largest age groups, while 110 pregnant women represented 8.1% of participants.
- The participants’ mean age was 73.4 ± 0.1 months, ranging from 0.1 to 356.1 months.
- 647 diagnoses were confirmed, including simple congenital cardiopathy (30.2%), acquired cardiopathy (3.3%), and complex congenital cardiopathy (3.9%).
- 834 participants (53.2%) were referred for follow-up, 27 (1.2%) for additional testing, and 35 (2.2%) for surgery or intervention.
- The mean weight, height, BMI, heart rate, oxygen saturation, temperature, and blood pressures were reported alongside their corresponding ranges or variability.For example, mean heart rate was 102 ± 20 bpm and mean oxygen saturation was 95% ± 5%.
C. Heart Sounds and Annotations
The dataset combines recordings from two screening campaigns with detailed acquisition metadata and murmur annotations. It includes extensive heart-sound segmentation and identifies murmur presence and timing across patients.
- 215780 heart sounds were collected across the CC2014 and CC2015 campaigns, including separately reported S1 and S2 annotations.CC2014 contributed 103853 heart sounds, while CC2015 contributed 111927.
- 5282 recordings were collected from the four main auscultation locations and additional unreported points across both campaigns.Recording counts differed by campaign and location, with multiple recordings typically collected per patient.
- Signals were acquired with a Littmann 3200 stethoscope and DigiScope Collector at 4 kHz and 16-bit resolution, then normalized to [−1, 1].Average recording durations were 28.7 seconds for CC2014 and 19.0 seconds for CC2015.
- The Collector GUI supports patient creation, clinical-data entry, and heart-sound acquisition from an auscultation location.The illustrated acquisition example uses the Pulmonary point.
- CC2014 recordings ranged from 5.3 to 80.4 seconds, whereas CC2015 recordings ranged from 4.8 to 45.4 seconds.The campaign averages were 28.7 and 19.0 seconds, respectively.
- 305 patients had murmurs: 294 had only systolic murmurs, 1 only a diastolic murmur, and 9 had both.These counts summarize the murmur distribution in the collected dataset.
V. DATA LABELLING
The labelling workflow combines automatic segmentation recommendations with independent physiologist review and manual correction. Murmurs receive structured clinical annotations, while grading and recording conditions impose explicit scope boundaries.
- Segmentation and review: Three segmentation algorithms generated recommendations identifying S1, S2, and their boundaries before independent physiologist inspection.The physiologists reviewed mutually exclusive portions of the data.
- Segmentation and review: Disagreements triggered additional recommended annotations or manual segmentation of at least five heartbeat cycles.The resulting audio and annotation files were saved after manual correction.
- Segmentation and review: Only high-quality representative sections were retained with validated labels, while the remaining signal may contain mixed-quality data.Users can choose whether to use the suggested manually inspected time window.
- Murmur characterization: Murmurs were screened at each auscultation location and classified at the point where they were most audible.The adopted grading scale defines Grades I, II, and III by audibility and loudness.
- Murmur characterization: Murmurs were characterized by timing, shape, pitch, quality, and grade, using categories such as systolic timing, crescendo or plateau shape, and Levine grading.The listed attributes include early-, mid-, and late-systolic or diastolic timing; high, medium, or low pitch; and blowing, harsh, or musical quality.
- Scope and limitations: Grade annotations can diverge from standard murmur grading when not all auscultation locations are available.Grade I/VI may be assigned by default, while Grade III/VI can include murmurs potentially graded III/VI or higher because thrills require physical examination.
- Scope and limitations: Ambulatory recordings include stethoscope rubbing and background crying or laughing, making automatic analysis difficult but reflecting real operating environments.The dataset is described as representative of environments where computer-aided decision systems must operate.
VI. DISCUSSION
The dataset represents pediatric populations across rural and urban Northeast Brazil and captures diverse clinical and murmur characteristics. Its detailed annotations reveal dominant murmur patterns while highlighting measurement challenges in pitch analysis.
- Study population: 63% of patients were children and 20% were infants, supporting the dataset’s focus on pediatric populations in rural and urban Northeast Brazil.A few young adults with complex congenital heart disease also participated voluntarily.
- Murmur timing: 96.8% of observed murmurs were systolic and 3.2% were diastolic, consistent with the higher pressure gradients during systolic ejection.The discussion also notes that diastolic murmurs are faint and technically harder to detect.
- Murmur shape and quality: 59.8% of murmurs had a Plateau shape, while Diamond-shaped murmurs accounted for 18.5%.The authors suggest digital filters may attenuate or modify murmur shape, potentially contributing to the observed distribution.
- Murmur shape and quality: 52.5% of murmurs were classified as Harsh, whereas Musical murmurs represented 1.3%.Harsh quality is associated in the discussion with high-velocity flow across a pressure gradient and significant semilunar stenosis or ventricular septal defect.
- Murmur grading: 27% of systolic murmurs and 10% of diastolic murmurs were grade III or greater on the intensity scale.The grade variable represents murmur loudness, which is affected by lesion location and distance from the stethoscope.
- Measurement considerations: Pitch interpretation is difficult because stethoscope transfer functions and preprocessing filters alter perceived sound characteristics.Digital equalizers may partially compensate for gain losses and amplitude/phase distortions in digital auscultation.
VII. CONCLUSIONS
CirCor DigiScope provides a distinctive cohort and age distribution that may support decision-support systems across target populations. Its rich annotations also enable multiple research applications involving heart-sound analysis and murmur reporting.
- The dataset includes pediatric and pregnant populations with significant congenital and acquired cardiac diseases.
- Its homogeneous age distribution may support robust decision-support systems spanning neonates to adults.
- Rich annotations enable multichannel multi-site PCG analysis, heart-murmur detection and classification, and automatic murmur-report generation.
- The authors plan comparative studies of heart-sound segmentation and classification using the CirCor DigiScope dataset.