Comparison of the Performance of Statistical and Spectral Feature Based Models for Embedded Cough Detection Using Accelerometry Data

The work presented in this paper compares the performance of different machine learning approaches, based on spectral and statistical features, in identifying coughs from accelerometry data sensed via a wearable attached to a shirt’s collar. The extracted features are separately used to train and evaluate neural network models for cough detection - first as a multi-class problem, second as a binary problem. The models’ performance was compared in terms of overall accuracy, cough sensitivity, specificity, and F1-score. It is concluded that the model using statistical features for a multi-class cough detection achieved the best cough sensitivity of 100% and F1-score of 0.95 with cough specificity of 97.3%. The four classification models were further evaluated for on-board performance as a TinyML cough system. They were all successfully deployed on Nordic Thing:53 using Edge impulse, and their memory requirements and estimated time per inference are reported. In terms of memory and time, the statistical- multi-class model was the smallest model occupying 13.7 KB of Flash memory and 1.1 KB of RAM; it was also the fastest model, requiring an estimated 1 ms per inference.