Leveraging Deep Learning for Respiratory Sound Analysis in Anomalies and Disease Detection

Respiratory diseases are a major global health concern. Traditional diagnosis often relies on stethoscopes, a method known as auscultation. However, this approach can be subjective and vary depending on a doctor's experience. Our research addresses this limitation by proposing objective, deep learning-based models for analysing respiratory sounds. These models can be integrated into a telemedicine framework, enabling remote examination and analysis of respiratory sounds for both anomaly detection and disease diagnosis. This offers significant advantages in terms of accessibility and potential for early intervention. Our work introduces two distinct models. The first one is a CNN-LSTM model for anomaly classification. This model combines Convolutional Neural Networks (CNNs) for capturing spatial features and Long Short-Term Memory (LSTM) networks for analysing temporal patterns within respiratory sounds. It allows for the identification of anomalies like wheezes and crackles. The second model is a multi-feature CNN model for disease detection. This model utilizes a multi-feature CNN architecture to classify respiratory sounds into various disease categories. It extracts multiple features from the audio data (MFCCs, Chroma, Mel Spectrogram) to achieve accurate disease detection. Both models were trained and evaluated using the benchmark ICBHI dataset. Our results are promising, surpassing existing methods in terms of ICBHI scores. Specifically, the CNN-LSTM model achieved a score of 0.83, while the Multi-feature CNN achieved an impressive score of 0.93.