Robust Multimodal Cough Detection With Optimized Out-of-Distribution Detection for Wearables.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Wearable Technology

Background:

• Continuous cough monitoring is vital for diagnosing respiratory diseases.
• Wearable devices offer remote symptom monitoring but face challenges like privacy and poor audio quality.
• Existing systems struggle with real-world acoustic complexities and background noise.

Purpose of the Study:

• To develop and optimize a compact multimodal cough detection system using audio and Inertial Measurement Unit (IMU) signals.
• To enhance the system with an Out-of-Distribution (OOD) detection algorithm for improved robustness.
• To address privacy concerns and poor audio quality in real-world remote monitoring settings.

Main Methods:

• Developed a multimodal system integrating audio and IMU sensors.
• Optimized the system using an enhanced dataset and a weighted multi-loss approach for the In-Distribution (ID) classifier.
• Implemented an OOD detection algorithm by reconstructing training data components.

Main Results:

• The optimized system demonstrated robustness across various window sizes (1-5 seconds) and efficiency at low audio frequencies.
• Maintained high accuracy (90.08%) and cough F1 score (0.7548) at 16 kHz, and 87.3% accuracy (0.7015 F1 score) at 750 Hz, even with 50% OOD data.
• The multimodal model with OOD detection outperformed single-modal models, showing significant improvement in cough detection.

Conclusions:

• The Audio-IMU multimodal model with OOD detection offers a robust solution for real-world acoustic cough detection.
• The system enhances user privacy through efficient low-frequency audio processing.
• Further improvements can be achieved by incorporating more diverse non-speech vocalization data.