Manikin-Recorded Cardiopulmonary Sounds Dataset Using Digital Stethoscope

Technical Summary: Manikin-Recorded Cardiopulmonary Sounds Dataset Using Digital Stethoscope

Problem Statement
Cardiopulmonary diseases remain a leading cause of global mortality, necessitating accurate analysis of heart and lung functions. While auscultation is a vital diagnostic tool, the acoustic signals involved are weak and require sophisticated analysis. Although Artificial Intelligence (AI) and machine learning have advanced clinical decision-making, their effectiveness is constrained by the scarcity of high-quality, diverse datasets containing both heart and lung sounds. Existing datasets often focus exclusively on either cardiac or respiratory sounds, lack mixed recordings, or rely on post-hoc mixing of separate files rather than simultaneous capture. Furthermore, many prior works utilize older electronic stethoscopes with lower resolution compared to modern digital alternatives.

Methodology
To address these gaps, the authors developed a dataset using a controlled, risk-free environment involving a clinical skills manikin and advanced recording hardware.

• Data Source: Recordings were generated using the CAE Juno™ mid-fidelity patient simulator, which features interchangeable male and female chest skins and is capable of simulating synchronized cardiac cycles and ventilation. The simulator was controlled via the CAE Maestro software on a tablet connected via Wi-Fi, allowing real-time adjustment of simulation parameters.
• Recording Hardware: The study utilized the 3M™ Littmann® CORE Digital Stethoscope, connected via Bluetooth to a mobile application for visualization and storage. This device offers up to 40x amplification, active noise cancellation (85% reduction), and built-in frequency filters.
• Recording Protocol:
  • Environment: Recordings were conducted in a quiet, noise-free professional practice collaboratory with the manikin in a sitting position.
  • Locations: Sounds were captured from 12 distinct anatomical landmarks, including standard heart auscultation sites (Apex, RUSB, LUSB, LLSB, RC, LC) and lung zones (Upper, Middle, and Lower Anterior on both right and left sides).
  • Filters: Three filter modes were employed based on the target sound: Bell mode (20–200 Hz) for low-frequency heart sounds, Diaphragm mode (100–500 Hz) for high-frequency lung sounds, and Midrange mode (50–500 Hz) for mixed recordings.
  • Duration: Each recording lasted 15 seconds, captured at a 22,050 Hz sample rate in .wav format.
• Data Composition: The final dataset comprises 210 recordings (101 female, 109 male) categorized into three groups:
  1. Heart Sounds (HS): 50 recordings covering normal sounds and 9 abnormalities (e.g., murmurs, atrial fibrillation, tachycardia, S3/S4).
  2. Lung Sounds (LS): 50 recordings covering normal sounds and 5 abnormalities (e.g., wheezing, crackles, rhonchi, pleural rub, gurgling).
  3. Mixed Sounds (Mix): 110 recordings capturing simultaneous heart and lung sounds with various combinations of the aforementioned pathologies.
• Validation: A nursing team verified the identification of auscultation landmarks. Qualitative analysis was performed by listening to audio files to ensure alignment with expected physiological patterns. The manikin undergoes annual calibration to ensure accuracy.

Key Contributions
The primary contribution of this work is the release of a novel dataset that offers simultaneous, naturally occurring mixed cardiorespiratory recordings, a feature the authors note is absent in previous major datasets (e.g., those by Tsai et al., Fraiwan et al., and Oliveira et al.).

• Unique Mixed Recordings: Unlike datasets where heart and lung sounds are recorded separately and mixed algorithmically, this dataset captures the natural overlap of signals, which is critical for developing unsupervised separation techniques like blind source separation.
• Diversity: The dataset includes a broader range of abnormalities than prior simulator-based studies, covering ten distinct heart sound types and six lung sound types.
• High Fidelity: The use of a digital stethoscope with advanced filtering and noise cancellation provides higher resolution and cleaner signals compared to datasets recorded with older electronic stethoscopes.
• Metadata: Each file is accompanied by a CSV file containing detailed metadata, including gender, specific sound types, and anatomical location, following a structured naming convention.

Results and Validation
The study successfully generated 210 high-quality audio files. The authors validated the data through:

1. Technical Specifications: Confirmation of the stethoscope's frequency response and noise cancellation capabilities (Table 1).
2. Qualitative Analysis: Expert verification that the recorded waveforms and spectrograms (visualized in Figures 5 and 6) matched expected physiological patterns.
3. Reproducibility: The use of a standardized manikin and recording protocol ensures that the dataset can be reproduced in similar scenarios.

Significance and Claims
The authors position this dataset as a valuable resource for the biomedical engineering and AI research communities. They claim its specific utility lies in:

• Algorithm Development: It supports both supervised learning (for classification of distinct heart/lung sounds) and unsupervised learning (for clustering and blind source separation of mixed signals).
• Clinical Applications: The dataset is intended for use in automated cardiopulmonary disease detection, sound classification, and signal processing tasks such as Principal Component Analysis (PCA), Independent Component Analysis (ICA), Support Vector Machines (SVM), and Convolutional Neural Networks (CNNs).
• Foundation for Future Work: The authors note that they have already utilized this dataset to propose a modified affine non-negative matrix factorization (NMF) method for blind separation, demonstrating its immediate applicability to advanced signal processing research.

The paper concludes that by providing clean, isolated, and mixed recordings in a realistic simulation environment, this dataset addresses a critical bottleneck in training robust AI models for cardiopulmonary monitoring.