AuscultaBase: A Foundational Step Towards AI-Powered Body Sound Diagnostics

Pingjie Wang, Zihan Zhao, Liudan Zhao, Miao He, Xin Sun, Ya Zhang, Kun Sun, Yanfeng Wang, Yu Wang

Abstract

Auscultation of internal body sounds is essential for diagnosing a range of health conditions, yet its effectiveness is often limited by clinicians' expertise and the acoustic constraints of human hearing, restricting its use across various clinical scenarios. To address these challenges, we introduce AuscultaBase, a foundational framework aimed at advancing body sound diagnostics through innovative data integration and contrastive learning techniques. Our contributions include the following: First, we compile AuscultaBase-Corpus, a large-scale, multi-source body sound database encompassing 11 datasets with 40,317 audio recordings and totaling 322.4 hours of heart, lung, and bowel sounds. Second, we develop AuscultaBase-Model, a foundational diagnostic model for body sounds, utilizing contrastive learning on the compiled corpus. Third, we establish AuscultaBase-Bench, a comprehensive benchmark containing 16 sub-tasks, assessing the performance of various open-source acoustic pre-trained models. Evaluation results indicate that our model outperforms all other open-source models in 12 out of 16 tasks, demonstrating the efficacy of our approach in advancing diagnostic capabilities for body sound analysis.

Introduction: The Challenge of Medical Auscultation

Auscultation—the clinical practice of listening to internal body sounds—remains a cornerstone of physical examination. Despite its ubiquity, the technique is inherently subjective. A clinician’s diagnostic accuracy is heavily dependent on their experience and the physiological limits of human hearing, which often leads to significant inter-observer variability.

While Artificial Intelligence (AI) has the potential to standardize and scale these diagnostics, existing models face two critical hurdles. First, there is a significant domain gap: most "foundation" audio models are trained on general datasets like AudioSet, which focus on music, speech, and environmental noise. These models often fail to capture the subtle, low-frequency, and rhythmic pathological variations unique to medical sounds. Second, data fragmentation has hindered progress; medical audio datasets are typically small, task-specific, and lack a unified benchmark for rigorous comparison.

The AuscultaBase Framework

To address these limitations, we developed a comprehensive three-part framework designed to bridge the gap between general audio AI and specialized medical diagnostics.

The AuscultaBase framework overview, illustrating database construction, model pre-training, benchmarking, and clinical evaluation.

AuscultaCorpus: Large-Scale Data Integration

We aggregated AuscultaCorpus, a massive repository integrating 11 datasets (10 public and one proprietary). This corpus contains 40,317 recordings totaling 322.4 hours of audio, covering three primary categories: lung sounds (e.g., SPRSound, ICBHI 2017), heart sounds (e.g., Cinc 2016, Circor 2022), and bowel sounds. This represents one of the most diverse and extensive collections of medical auscultation data to date.

AuscultaBase-Model: Self-Supervised Learning

The core of our framework is a transformer-based architecture (utilizing a Swin Transformer backbone) trained via self-supervised contrastive learning. By training the model to identify "positive pairs" (segments from the same recording) against "negative distractors," the encoder learns to discern fine-grained acoustic patterns without requiring exhaustive manual labeling for every second of audio. This approach allows the model to develop robust feature representations specifically tuned to the medical domain.

Benchmarking and Performance

To evaluate the model, we established AuscultaBench, a standardized benchmark consisting of 16 downstream tasks. These tasks range from broad abnormality detection to specific disease diagnoses, such as classifying the severity of Chronic Obstructive Pulmonary Disease (COPD).

Macro-F1 performance comparison across various task categories, showing AuscultaBase (red) consistently outperforming other acoustic pre-trained models.

Key Results

  • State-of-the-Art Performance: AuscultaBase achieved the highest Macro-F1 scores in 12 out of 16 tasks, consistently outperforming established general-purpose models like AudioMAE, CLAP, and PANN.
  • Superior Discriminative Power: In binary classification tasks (e.g., Healthy vs. Abnormal), the model reached an AUC of 0.920 in specific heart sound tasks, significantly higher than the best baseline (0.872).
  • Clinical Sensitivity: In comparative tests against human experts for detecting Congenital Heart Disease (CHD), AuscultaBase demonstrated a sensitivity of 80%, far surpassing the 38% achieved by clinicians. This high sensitivity makes the model an ideal screening tool, ensuring fewer pathological cases are missed.
  • Pediatric Accuracy: The model was particularly effective in challenging pediatric cases, achieving 77% accuracy in infants under one year old, compared to 58% for human experts.

Discussion and Future Directions

While AuscultaBase represents a significant leap forward, several areas for future work remain. The current corpus, while large, does not yet encompass every rare pathological variation. Furthermore, like many deep learning systems, the model currently operates as a "black box." Future iterations will focus on improving interpretability, providing clinicians with explainable markers that highlight specific abnormal segments within a cardiac or respiratory cycle.

Ultimately, AuscultaBase provides a foundational step toward integrating AI-powered diagnostics into clinical workflows, offering a path toward more objective, accessible, and accurate body sound analysis across diverse healthcare settings.

Summary updated

Abstract

Auscultation of internal body sounds is essential for diagnosing a range of health conditions, yet its effectiveness is often limited by clinicians' expertise and the acoustic constraints of human hearing, restricting its use across various clinical scenarios. To address these challenges, we introduce AuscultaBase, a foundational framework aimed at advancing body sound diagnostics through innovative data integration and contrastive learning techniques. Our contributions include the following: First, we compile AuscultaBase-Corpus, a large-scale, multi-source body sound database encompassing 11 datasets with 40,317 audio recordings and totaling 322.4 hours of heart, lung, and bowel sounds. Second, we develop AuscultaBase-Model, a foundational diagnostic model for body sounds, utilizing contrastive learning on the compiled corpus. Third, we establish AuscultaBase-Bench, a comprehensive benchmark containing 16 sub-tasks, assessing the performance of various open-source acoustic pre-trained models. Evaluation results indicate that our model outperforms all other open-source models in 12 out of 16 tasks, demonstrating the efficacy of our approach in advancing diagnostic capabilities for body sound analysis.
BibTeX
@misc{wang2024auscultabase,
  author = {Pingjie Wang and Zihan Zhao and Liudan Zhao and Miao He and Xin Sun and Ya Zhang and Kun Sun and Yanfeng Wang and Yu Wang},
  title = {{AuscultaBase: A Foundational Step Towards AI-Powered Body Sound Diagnostics}},
  year = {2024},
  eprint = {2411.07547},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.07547},
}
Figure