A device-invariant multi-modal learning framework for respiratory disease classification - PubMed
3 hours ago
- #deep learning
- #respiratory disease
- #multimodal fusion
- Proposes a device-invariant, multimodal deep learning framework for respiratory disease classification using cough acoustics, demographic data, and symptom descriptions.
- Addresses device heterogeneity with an adversarial branch in the audio encoder and invariant risk minimization-augmented loss for robust feature learning.
- Evaluated on a multi-center dataset with over 10,000 cases covering seven major respiratory conditions, achieving high AUROC scores for diseases like COPD (0.9698), LRTI (0.8483), and PS (0.8720).
- Demonstrates effectiveness in identifying comorbidities for seven respiratory diseases with an overall AUROC of 0.8907.
- Shows promise in mitigating device effects and improving cross-device generalization for cough-based diagnoses.
- Highlights the importance of multimodal fusion and robust representation learning for clinical applicability in respiratory screening.