To develop a device-invariant, multimodal deep learning framework for classifying respiratory diseases using cough acoustics, demographic data, and symptom descriptions.
Key Findings:
Achieved AUROC of 0.9698 for COPD, 0.8483 for LRTI, and 0.8720 for pulmonary shadows.
Overall AUROC of 0.8907 for identifying comorbidities across seven respiratory diseases.
Demonstrated mitigation of device effect and enhanced cross-device generalization.
Interpretation:
The proposed AI-based approach shows promise for scalable and transferable cough-driven respiratory screening, highlighting the importance of multimodal fusion and robust representation learning.
Limitations:
Datasets are not publicly available due to sensitive clinical information and proprietary components.
The developed code is proprietary and cannot be shared.
Conclusion:
This work advances the clinical applicability of cough sound analysis for respiratory disease diagnosis through a robust multimodal deep learning framework.
