DiemaNet: An automated detection model for three typical respiratory diseases

Respiratory diseases are leading causes of serious illness and mortality globally. Rapid and accurate diagnosis is crucial for effective treatment. Cough analysis offers a simple, low-cost, and accessible solution for detecting respiratory diseases anytime, anywhere, with significant public health implications for early intervention. However, existing cough datasets for respiratory diseases are limited. Most studies focus on binary detection of specific diseases, with fewer exploring multi-disease detection. Additionally, single-feature input models have limited performance. To address these challenges, this paper constructs a respiratory disease cough dataset and proposes DiemaNet, a deep learning method combining temporal dimensionality reduction and attention mechanisms for automatic cough-based detection. The temporal dimension reduction module (DiffRes) enhances detailed information extraction with lower computational cost and improves attention to critical temporal frames. Efficient Multi-Scale Attention (EMA) further boosts performance by highlighting important channel and spatial information. To evaluate the performance of the proposed method, we constructed a cough dataset named Cough-Exp, which includes four types of cough data: Health, COVID-19, Asthma, and chronic obstructive pulmonary disease (COPD). Experiments show DiemaNet achieves the best performance on the dataset, with precision, recall, F1 score, and accuracy of 86.2%, 85.6%, 85.9%, and 84.7%, respectively.