Neural Cough Classification Method for Embedded Devices

This technical report outlines the development of a neural cough classifier designed for edge devices by Hyfe Inc. The objective is to provide an accurate and efficient method for monitoring cough frequency and severity, which is crucial for managing respiratory illnesses. The report details the use of a microphone embedded in wearable devices, leveraging the Edge Impulse platform for deploying machine learning models. It describes the methodology for training neural networks using Mel-Filterbank Energies (MFE) for feature extraction, achieving high sensitivity and specificity in cough detection. The experiments demonstrate that the classification can be performed in under 100 milliseconds with a sensitivity of approximately 91% and specificity of 99.7%. The report also discusses the integration of a Software Development Kit (SDK) that enables third-party applications to utilize the cough detection capabilities in real-time. The findings suggest that reliable cough detection is feasible on low-power embedded devices.