Resp-Agent: An Agent-Based System for Multimodal Respiratory Sound Generation and Disease Diagnosis
New agent-based system weaves audio with medical records to spot hard-to-diagnose respiratory conditions.
Deep Dive
Researchers Pengfei Zhang, Tianxin Xie, and Minghao Yang developed Resp-Agent, an autonomous multimodal system for respiratory disease diagnosis. It combines a 229k recording dataset (Resp-229k) with EHR data via a Modality-Weaving Diagnoser and uses a Flow Matching Generator to synthesize rare cases. The system's Active Adversarial Curriculum Agent identifies diagnostic weaknesses, improving accuracy in data-scarce, imbalanced scenarios where traditional methods fail.
Why It Matters
Could enable earlier, more accurate detection of respiratory diseases using accessible audio data, especially where medical imaging is limited.