JMIR Publications

ABSTRACT

Background:

Spirometry is the gold standard test for diagnosing chronic obstructive pulmonary disease (COPD) and other obstructive lung diseases, but it requires calibrated equipment and trained personnel and is often unavailable in low and middle income settings. Consequently, airflow limitation is under detected in many regions. Because chest radiography is widely available and inexpensive, recent studies have explored whether deep learning can estimate spirometric indices from chest radiographs.

Objective:

To build on prior work by evaluating a deep learning model that predicts the FEV₁/FVC ratio and classifies obstructive lung disease from chest radiographs, and to assess model fairness across demographic subgroups.

Methods:

We retrospectively assembled a cohort of 3,537 unique patients who underwent both pre bronchodilator spirometry and chest radiography at a single Canadian hospital. A convolutional neural network (ConvNeXt base) was trained to predict the continuous FEV₁/FVC ratio using 2,263 patients for training, 566 for validation and 708 for testing. By thresholding predictions at 0.70, examinations were also classified as obstructive or non obstructive. Performance was summarized overall and within age, sex and ethnicity strata.

Results:

On the held out test cohort (708 patients, 3,274 radiograph examinations), the model achieved a mean squared error of 0.07 for ratio prediction. For the binary obstruction task, sensitivity was 0.70, specificity 0.72, positive predictive value 0.71 and negative predictive value 0.71. These values exceeded those of a prevalence matched random classifier across all examined demographic groups. Subgroup analyses showed particularly large gains in specificity, and absolute accuracy improvements of 0.20–0.30 were observed in cohorts with higher obstruction prevalence. Fairness analyses revealed no clinically meaningful differences in performance across age, sex or ethnicity.

Conclusions:

This study extends earlier work on chest radiograph–based estimation of lung function by demonstrating comparable performance in a North American cohort and providing a comprehensive fairness assessment. Given the ubiquity of chest radiography and the under utilization of spirometry, such models could offer a practical screening tool for obstructive lung disease, especially in regions where access to spirometry is limited. Prospective validation is warranted to support clinical adoption. Clinical Trial: Not applicable.