Transfer learning in spirometry: CNN models for automated flow-volume curve quality control in paediatric populations

Area of Science:

• Pulmonary function testing
• Artificial intelligence in healthcare
• Machine learning for medical diagnostics

Background:

• Spirometry interpretation requires manual assessment of acceptability criteria, leading to variability.
• Automating spirometry quality control can enhance diagnostic efficiency and consistency.

Purpose of the Study:

• To apply transfer learning with convolutional neural networks (CNNs) for automated classification of spirometry flow-volume curves.
• To evaluate the performance of various CNN models in identifying acceptable and non-acceptable spirometry tests.

Main Methods:

• 5287 spirometry curves were categorized into acceptable, early termination, and non-acceptable.
• Six CNN models (VGG16, InceptionV3, Xception, ResNet152V2, InceptionResNetV2, DenseNet121) were trained using augmented data.
• Model performance was assessed using accuracy, precision, recall, and F1-score.

Main Results:

• VGG16 achieved the highest accuracy (93.9%).
• Non-acceptable curves were easiest to classify (precision ≥87.7%).
• Early termination curves presented the most challenge (precision 75.0%–90.3%).

Conclusions:

• CNN models, especially VGG16, demonstrate potential for automating spirometry quality control.
• This automation can reduce reliance on manual inspection by specialists.
• The approach promises streamlined, consistent spirometry diagnostics, beneficial for non-specialized settings.