Published Paper Details:

The advancement of intelligent systems has significantly impacted the healthcare industry, enabling the development of automated diagnostic tools for early disease detection and monitoring. Respiratory diseases, particularly in children, are among the most common health concerns, requiring timely and accurate diagnosis. Traditional methods rely on clinical evaluations and laboratory tests, which can be time-consuming and resource-intensive. In this study, we introduce RESPITECH, a deep learning-based framework for automated cough sound classification. By leveraging audio signal processing techniques and machine learning, RESPITECH aims to differentiate between healthy and pathological coughs in individuals, with a focus on respiratory conditions such as asthma, upper respiratory tract infection (URTI), and lower respiratory tract infection (LRTI), COPD, Pneumonia, Bronchitis and also Healthy lungs. Our approach employs a Convolutional Neural Network model (CNN) trained on Mel-Frequency Cepstral Coefficients (MFCCs) extracted from a carefully curated dataset of cough sounds labeled by clinicians. The CNN model is designed to capture temporal dependencies in cough sounds, enhancing classification performance. When trained to distinguish between healthy and pathological coughs, the model achieves an accuracy exceeding 84%, aligning closely with physician diagnoses. To improve classification performance for specific respiratory pathologies, multiple cough epochs per subject are aggregated, resulting in an overall accuracy surpassing 91% for detecting asthma, URTI, and LRTI. However, classification performance declines when distinguishing between four separate cough categories, as certain pathological coughs share overlapping acoustic characteristics, leading to misclassification. A detailed analysis of the MFCC feature space through a longitudinal study indicates that pathological coughs, regardless of the underlying condition, exhibit similar acoustic patterns, making it challenging to differentiate between specific respiratory diseases using MFCCs alone. This suggests that while MFCCs provide valuable spectral information, additional features or multimodal data integration may be necessary for finer disease discrimination. Despite these challenges, RESPITECH demonstrates strong potential as an automated, non-invasive screening tool for respiratory disease monitoring. Future work will explore advanced feature extraction techniques, multimodal learning approaches, and real-world deployment strategies to enhance diagnostic accuracy and clinical applicability.