An Ensemble of Deep Learning Frameworks for Predicting Respiratory Anomalies

Area of Science:

• Artificial Intelligence
• Biomedical Signal Processing
• Medical Informatics

Background:

• Respiratory diseases present a significant global health burden.
• Accurate and early detection of respiratory anomalies like wheezes and crackles is crucial for effective patient management.
• Traditional diagnostic methods can be invasive or require specialized equipment, highlighting the need for non-invasive approaches.

Purpose of the Study:

• To evaluate and compare various deep learning frameworks for the automated detection of respiratory anomalies from audio recordings.
• To develop a high-performance system capable of classifying respiratory cycles into medically relevant categories.
• To achieve state-of-the-art performance on a recognized benchmark dataset for respiratory sound analysis.

Main Methods:

• Respiratory audio recordings were transformed into time-frequency spectrograms.
• Deep learning networks, including inception-based and transfer learning models, were trained on spectrogram segments.
• A late fusion strategy combining multiple high-performance frameworks was employed to enhance classification accuracy.

Main Results:

• The proposed late fusion model achieved an ICBHI score of 57.3%, setting a new state-of-the-art benchmark.
• The fusion of inception-based and transfer learning deep learning frameworks significantly outperformed individual models and existing systems.
• The system demonstrated high efficacy in classifying respiratory cycles, indicating potential for clinical application.

Conclusions:

• Deep learning, particularly through the fusion of advanced frameworks, offers a powerful and accurate approach for non-invasive respiratory anomaly detection.
• The developed method shows significant promise for improving the diagnosis and monitoring of respiratory diseases.
• This research contributes a novel, high-performing computational tool for respiratory sound analysis.