Deep Learning-Based Analytic Models Based on Flow-Volume Curves for Identifying Ventilatory Patterns.

Area of Science:

• Pulmonary Medicine
• Artificial Intelligence
• Medical Diagnostics

Background:

• Spirometry is increasingly used in primary care for pulmonary function testing.
• Accurate identification of ventilatory patterns (normal, obstructive, restrictive, mixed) is crucial and guided by ATS/ERS standards.
• Lung volume assessments are typically required alongside spirometry for pattern classification.

Purpose of the Study:

• To evaluate the accuracy of deep learning models in classifying ventilatory patterns using flow-volume curves.
• To compare the performance of the best deep learning model against human physicians in interpreting these patterns.

Main Methods:

• Ten deep learning models, including VGG and ResNet architectures, were developed using flow-volume curve data.
• The gold standard for classification was based on American Thoracic Society (ATS) and European Respiratory Society (ERS) guidelines.
• The best-performing model's results were cross-checked against interpretations by 90 physicians from various healthcare settings.

Main Results:

• The VGG13 deep learning model achieved the highest accuracy of 95.6% on the test set.
• Physicians achieved an average accuracy of 76.9%, with primary care physicians showing lower accuracy at 56.2%.
• The VGG13 model correctly identified ventilatory patterns in 92.0% of cases interpreted by physicians, demonstrating superior performance (P < 0.0001).

Conclusions:

• The VGG13 model accurately identifies ventilatory patterns from flow-volume curves independently.
• This AI tool shows potential to aid physicians, particularly in primary care, in reducing diagnostic errors and variability.
• The model offers a reliable method for classifying ventilatory patterns without necessitating additional parameters.