Lightweight Hybrid Deep Learning Models for Accurate Classification of Respiratory Conditions from Raw Lung Sounds

Area of Science:

• Artificial Intelligence
• Medical Diagnostics
• Respiratory Medicine

Background:

• Deep learning has significantly improved medical diagnosis.
• Analyzing raw lung auscultation sounds for respiratory pathologies presents challenges.
• Need for efficient and accurate diagnostic tools for respiratory diseases.

Purpose of the Study:

• To evaluate lightweight deep learning models for detecting eleven respiratory pathologies using raw lung auscultation sounds.
• To compare the performance of different deep learning models on original and augmented datasets.
• To identify the most effective deep learning model for respiratory sound analysis.

Main Methods:

• Utilized a deep learning framework with multiple lightweight models.
• Analyzed raw lung auscultation sounds without feature engineering or preprocessing.
• Assessed model performance on original and augmented datasets.
• Employed a hybrid Convolutional Neural Network (CNN)-Long Short-Term Memory (LSTM) model.

Main Results:

• Deep learning models achieved remarkable accuracy (>99%) in detecting respiratory pathologies from raw lung sounds.
• The hybrid CNN-LSTM model outperformed standalone CNN and LSTM models.
• Data augmentation enhanced the performance of deep learning models.
• The lightweight hybrid CNN-LSTM model reached 100% accuracy on the augmented dataset.

Conclusions:

• Raw lung auscultation sounds can be reliably used to detect multiple respiratory pathologies with lightweight deep learning models.
• The hybrid CNN-LSTM model offers an efficient and accurate solution for respiratory disease diagnosis.
• Further external validation is needed to assess model generalization across diverse clinical datasets.