A device-invariant multi-modal learning framework for respiratory disease classification

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Signal Processing

Background:

• Smartphone-based cough analysis shows promise for remote respiratory disease screening.
• Existing deep learning models face limitations due to device variability, diverse populations, and multimodal data integration challenges.

Purpose of the Study:

• To develop a device-invariant, multimodal deep learning framework for accurate multi-label classification of adult respiratory diseases using cough acoustics, demographics, and symptoms.
• To enhance the robustness and generalizability of AI models for cough-based respiratory diagnostics.

Main Methods:

• Proposed a multimodal deep learning framework incorporating an adversarial branch for device-invariant audio feature learning.
• Employed invariant risk minimization-augmented loss to improve robustness against non-structural shifts.
• Utilized a real-world, multi-center dataset of over 10,000 cases across seven respiratory conditions.

Main Results:

• Achieved superior performance in identifying chronic obstructive pulmonary disease (COPD) (AUROC 0.9698), lower respiratory tract infection (LRTI) (AUROC 0.8483), and pulmonary shadows (PS) (AUROC 0.8720).
• Demonstrated promising results in identifying comorbidities for 7 respiratory diseases (overall AUROC 0.8907).
• Effectively mitigated device effects and improved cross-device generalization for cough-based diagnoses.

Conclusions:

• The developed AI framework offers a scalable and transferable approach for cough-driven respiratory screening.
• Multimodal fusion and robust representation learning are crucial for advancing the clinical applicability of AI in respiratory diagnostics.
• The method shows significant potential for self-management care in home settings.