Automated Multiclass Classification of Lung Diseases with Deep Learning: A Hybrid CNN-LSTM Approach

The subjective and error-prone traditional methods for detecting respiratory diseases necessitate developments for diagnostic technologist. This research examines deep learning algorithms, especially Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, which analyze breathing sounds to diagnose respiratory ailments. We attempt to develop a strong and precise model that can classify diseases based on the 2017 ICBHI dataset with its 920 audio recordings of different types of respiratory sounds. Our method involves extracting relevant features from the audio files using Librosa library, followed by hybrid CNN-LSTM modeling. In this approach, strengths of CNNs in feature extraction are combined with those of LSTMs in temporal pattern recognition. The final model was able to attain an accuracy of 95℅ indicating the possibility of its real-life application in clinics. Our results show that deep-learning models are effective in analyzing respiratory sound, paving way for better medicine tools within our health care system.