AI-Powered Tool Revolutionizes Medical Audio Analysis for Enhanced Clinical Diagnostics

On May 2, 2025, researchers Tsai-Ning Wang, Lin-Lin Chen, Neil Zeghidour, and Aaqib Saeed published CaReAQA: A Cardiac and Respiratory Audio Question Answering Model for Open-Ended Diagnostic Reasoning, introducing an innovative AI system that integrates audio analysis with large language models to improve medical diagnostics. The model achieved 86.2% accuracy in open-ended tasks, demonstrating its potential for clinical decision support.

Analyzing medical audio signals like heart and lung sounds remains challenging due to reliance on handcrafted features or supervised deep learning requiring extensive labeled data. To address this, researchers developed CaReAQA, an audio model integrating foundation audio with reasoning capabilities for open-ended diagnostic responses. They also introduced CaReSound, a benchmark dataset of annotated medical audio recordings enriched with metadata and paired question-answer examples. Evaluation shows CaReAQA achieves 86.2% accuracy on open-ended tasks and generalizes well to closed-ended classification, achieving 56.9% accuracy on unseen datasets. This demonstrates how integrating audio models with reasoning advances medical diagnostics for efficient AI systems in clinical decision support.

In an era where artificial intelligence (AI) is reshaping healthcare, a novel tool called CaReAQA is emerging as a significant advancement in heart diagnostics. This innovative system utilises large language models (LLMs) to interpret heart sounds and lung crackles, offering a fresh approach compared to traditional methods that rely on imaging or electronic health records.

CaReAQA distinguishes itself by focusing on audio data, an area fraught with challenges due to variability in recording quality and patient conditions. The model was trained on extensive datasets of heart murmurs and lung crackles, demonstrating superior performance compared to other models like Qwen and SmolLM after specialised training.

The research involved fine-tuning the base Llama model for specific medical tasks, a strategy that proved more effective than broad instruction-tuning. This approach yielded high accuracy in diagnosing conditions such as mitral regurgitation and aortic stenosis, with performance exceeding competitors by 10-20% points.While the results are promising, challenges remain. The model struggles with conditions sharing similar acoustic features and less common diseases due to limited training data. This underscores the need for diverse datasets and improved handling of ambiguous cases.

CaReAQA is poised to become an invaluable tool for healthcare professionals, particularly beneficial in areas with scarce specialist resources. Its integration could enhance early detection of heart conditions, potentially saving lives. The success of CaReAQA highlights the importance of data curation and domain-specific training in AI development.

Future research may explore noise reduction techniques to improve robustness against poor-quality recordings. Addressing these limitations will be crucial for maximising the tool’s impact in clinical practice.

CaReAQA represents a promising step forward in AI-driven healthcare diagnostics. While it complements human expertise rather than replacing it, its potential to enhance patient outcomes is substantial. As the field evolves, addressing current limitations will be crucial for maximising the tool’s impact in clinical practice.