Automated Cough Detection System based on Vision Transformers

Cough is a common symptom with significant public health implications. Objective cough detection is crucial disease monitoring, the development of effective therapies and improving patient care. This study aimed to develop an automated, accurate, and generalisable system for detecting cough events using machine learning. We developed an automated cough detection model based on the Vision Transformer (ViT) architecture. The dataset comprising 232 24-hour recordings across nine diagnostic categories was used for model training and evaluation. Recordings were segmented into one-second clips, converted into spectrograms, and classified using the ViT model. We evaluated the model using sensitivity, specificity, precision, and F1 score as key metrics. Our model achieved a sensitivity of 91.5%, specificity of 99.0%, precision of 73.9%, and F1 score of 0.82 on the test dataset, demonstrating strong performance. Bland-Altman analysis revealed an average difference (bias) of 97 cough events per 24h recording. The model effectively automates cough detection, offering a significant efficiency improvement over manual method. Future improvements may involve enhancing preprocessing techniques, reducing model complexity, and refining the labelling process to align better with the model's detection approach.