Automated Classification of Respiratory Sounds

Automated classification of respiratory sounds harnesses advanced signal-processing techniques and artificial intelligence to identify normal and pathological lung sounds without reliance on subjective human interpretation. Digital stethoscopes capture auscultatory recordings, which are subsequently transformed into time–frequency representations such as spectrograms. Machine learning algorithms, notably convolutional neural networks and hybrid neural architectures, then extract discriminative features to distinguish between normal breathing, crackles, wheezes, rhonchi and other adventitious sounds. This automated approach addresses inter-observer variability, enhances diagnostic consistency, and supports remote and telemedicine applications. Recent advances have focused on improving noise robustness, optimising model interpretability via attention mechanisms and integrating temporal context using recurrent networks. By facilitating early detection and monitoring of conditions such as pneumonia, asthma and chronic obstructive pulmonary disease, these systems have significant global implications for resource-limited settings and routine clinical practice.

Research from Nature Portfolio

A deep learning-based system was developed using pretrained image feature extractors combined with a convolutional neural network classifier to categorise over 1,900 clinical recordings into normal, crackles, wheezes and rhonchi. The model achieved an overall accuracy of approximately 86% and an area under the receiver-operator characteristic curve near 0.93. Comparative evaluation revealed that automated classification outperformed junior clinicians and matched the performance of experienced fellows, underscoring its potential to complement traditional auscultation and accelerate diagnosis.

Automated Classification of Respiratory Sounds publication trend

The graph below shows the total number of articles in automated classification of respiratory sounds across all publications each year (not limited to Nature Index journals).

Technical terms

Auscultation: The practice of listening to internal body sounds, typically using a stethoscope, to assess respiratory or cardiac function.

Adventitious respiratory sounds: Abnormal lung sounds, including crackles, wheezes and rhonchi, indicative of pathological conditions in the airways or lung parenchyma.

Spectrogram: A visual representation of an audio signal showing how its frequency content varies over time, often used as input for deep learning models.

Convolutional Neural Network (CNN): A deep learning architecture that applies convolutional filters to extract spatially localised features, widely used for image and spectrogram analysis.

Temporal attention: A mechanism in sequential models that assigns varying weights to different time steps, enhancing interpretability by highlighting critical segments of the input sequence.

References

1. Respiratory sound classification for crackles, wheezes, and rhonchi in the clinical field using deep learning. Scientific Reports (2021).
2. DeepBreath—automated detection of respiratory pathology from lung auscultation in 572 pediatric outpatients across 5 countries. npj Digital Medicine (2023).
3. Automated Lung Sound Classification Using a Hybrid CNN-LSTM Network and Focal Loss Function. Sensors (2022).
4. A review on lung disease recognition by acoustic signal analysis with deep learning networks. Journal of Big Data (2023).