HearCough: Enabling continuous cough event detection on edge computing hearables.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Wearable Technology

Background:

• Cough event detection is crucial for monitoring pulmonary illnesses.
• Existing methods often require significant computational resources.
• Wearable devices offer potential for continuous health monitoring.

Purpose of the Study:

• To develop a lightweight, continuous cough event detection system for edge computing hearables.
• To leverage active noise cancellation (ANC) microphones in commodity hearables for cough detection.
• To enable low-cost pulmonary health monitoring through hearable technology.

Main Methods:

• Proposed Tiny-COUNET, a lightweight end-to-end neural network model for cough detection.
• Utilized transfer learning for training the Tiny-COUNET model.
• Implemented and deployed the Tiny-COUNET model on consumer hearables, creating the HearCough system.
• Conducted a field study with 8 patients to evaluate HearCough's effectiveness and reliability.

Main Results:

• Tiny-COUNET achieved comparable detection performance to cutting-edge methods with reduced computational resources and storage.
• HearCough demonstrated effective and reliable continuous cough event detection in a real-world setting.
• Achieved 90.0% accuracy and 89.5% F1-score for 2 Hz cough event detection.
• Consumed only an additional 5.2 mW of power for continuous operation.

Conclusions:

• HearCough provides an effective, low-power solution for continuous cough event detection using commodity hearables.
• The system is suitable for integration into future hearables for enhanced pulmonary health monitoring.
• Tiny-COUNET offers a computationally efficient approach for on-device cough detection.