Deep Learning Estimation of Forced Expiratory Volume in One Second/Forced Vital Capacity and Obstructive Lung Disease...

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Pulmonary Diagnostics

Background:

• Spirometry is the gold standard for diagnosing airflow obstruction but is underutilized and unavailable in many regions.
• Chest radiographs are widely accessible and may serve as a complementary screening tool.
• The performance of deep learning models on chest radiographs across diverse populations is not well-examined.

Purpose of the Study:

• To develop and evaluate a deep learning model for estimating the forced expiratory volume in 1 second (FEV₁)/forced vital capacity (FVC) ratio from chest radiographs.
• To classify airflow obstruction (FEV₁/FVC <0.70) using the model.
• To assess the model's performance across various demographic subgroups.

Main Methods:

• A retrospective cohort of 3537 adults with concurrent spirometry and chest radiography was used.
• A ConvNeXt-Base deep learning architecture was trained to predict FEV₁/FVC ratio.
• Model performance was evaluated using regression and classification metrics, with subgroup analyses conducted for age, sex, and ethnicity.

Main Results:

• The model achieved a mean absolute error of 0.08 and RMSE of 0.10 for FEV₁/FVC prediction.
• Binary classification of airflow obstruction showed moderate sensitivity (0.70) and specificity (0.72).
• Performance varied significantly across subgroups, with reduced sensitivity observed in Asian patients and younger age groups.

Conclusions:

• Deep learning models can estimate FEV₁/FVC and identify airflow limitation from chest radiographs with moderate accuracy in a North American cohort.
• The model demonstrated good calibration for regression predictions and positive net benefit.
• Further validation across multiple sites and specific subgroup verification are crucial for clinical implementation.