Deep learning models for detecting respiratory pathologies from raw lung auscultation sounds

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Respiratory Medicine

Background:

• Deep learning models show promise in improving disease diagnosis, particularly for respiratory conditions.
• Accurate diagnosis of respiratory pathologies from lung auscultation sounds is crucial.

Purpose of the Study:

• To evaluate the performance of different deep learning models for detecting respiratory pathologies using raw lung auscultation sounds.
• To identify the optimal deep learning model for diagnosing respiratory conditions from digital sound recordings.

Main Methods:

• Three deep learning models were evaluated on both non-augmented and augmented datasets.
• Two distinct datasets were utilized, creating four sub-datasets for comprehensive analysis.
• The Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) hybrid model was investigated.

Main Results:

• All evaluated deep learning models achieved high performance in classifying raw lung sounds across various datasets.
• The CNN-LSTM model consistently outperformed other models, achieving accuracies up to 100% with augmentation.
• Data augmentation notably improved model performance, enhancing diagnostic accuracy.

Conclusions:

• The CNN-LSTM hybrid model is the most effective for diagnosing respiratory pathologies from lung sounds.
• Deep learning, especially hybrid CNN-LSTM architectures, offers a powerful tool for digital respiratory sound analysis.
• Data augmentation is a valuable technique for improving the robustness and accuracy of these diagnostic models.