Acute Respiratory Infections Identification With Cough Sounds and Overlapping Patch Modulated Vision Transformers

Rapid detection of Acute Respiratory Infections (ARI) is crucial to reduce breathing difficulties and severe life-threatening conditions. Automatic cough identification is being conducted using speech frequency analysis and machine learning models. Learning models trained on Mel frequency spectrum(MFCC) features of cough sounds represented as images have recorded an average binary classification accuracy of 68%. Variable cough sound vs silent intervals between samples of a class in MFCC spectral images has shown to influence training algorithms to learn meaningful patterns for classification. To learn all possible local patterns in the MFCC cough images using a vision transformer model (ViT), we propose an image patch overlapping vision transformer $IPO-ViT$ . The patch overlapping factor $k\ $ controls the quantity of common pixels between them. The $IPO-ViT$ patch encoder computes all possible local pixel pattern relationships by breaking the image into overlapping patches and equating them across all classes making a balanced augmented dataset. The $IPO-ViT$ is evaluated on our own 511 – sound cough dataset (IndiCough_2024) with 5 classes captured at AJ Institute of Medical Sciences, paediatric division along with benchmarks EPFL COUGH VID, Coswara for COVID-19 Diagnosis and Covid19-Cough. The $IPO-ViT$ achieved higher accuracies of around 92.33% over the state-of-the-art cough sound-based disease identification networks.