Assessing the Feasibility of Cough Detection Using Statistical Features Extracted from Accelerometry Data

The work presented in this paper investigates the feasibility of using data collected from a free-field accelerometer placed on a shirt’s collar for cough detection. The data collected captured body movement related to coughing, sneezing, laughing, talking, and resting while seated. Statistical features from the tri-axial data were extracted and used to train four different neural networks. The first two classification models addressed cough detection as a multi-class problem, differentiating between the five different activities; the first model using 21 input features, while the second model using 15 input features. The other two classification models targeted cough detection as a binary problem, clustering all non-cough activities under a single class. Similarly, the first binary model was trained using 21 input features, while the second using 15 input features. The performance of the four models were compared in terms of average accuracy, cough sensitivity, specificity, and F1-score. The best performing model in terms of cough sensitivity (100%) and F1-score (0.95) was the multi-class model using 15 input features. This model was finally deployed on Nordic Thingy:53 using Edge Impulse, reporting the estimated memory requirement, and model performance for on-board inference. This work provides a proof-of-concept for a TinyML cough detection system based on contact-free accelerometer.