Image by freepik

Introduction

Pulmonary function testing is a critical component in diagnosing and managing obstructive or restrictive respiratory impairments. Forced vital capacity (FVC) and forced expiratory volume in one second (FEV1) are key metrics obtained through spirometry, a method that has been in clinical use since 1846. Despite its importance, spirometry has limitations, particularly for patients who cannot perform the test due to age, physical, or cognitive constraints. The COVID-19 pandemic further highlighted the need for alternative methods due to the risk of infection during spirometry. This study explores the potential of a deep learning AI model to estimate FVC and FEV1 from chest x-rays, offering a less invasive and more accessible alternative to traditional spirometry.

Methods

Study Design

This retrospective study aimed to develop and validate a deep learning model capable of estimating FVC and FEV1 from chest x-rays. Data were collected from five institutions in Japan, including Osaka Metropolitan University Hospital, Habikino Medical Center, MedCity21, Higashisumiyoshi Morimoto Hospital, and Kashiwara Municipal Hospital. The study period spanned from July 1, 2003, to December 31, 2021. Chest x-rays taken within 14 days of spirometry assessments were included to ensure temporal alignment of data.

Model Testing and Performance Evaluation

The model's performance was assessed on both internal and external test datasets. The prediction accuracy for FVC and FEV1 was evaluated by comparing the model's estimates with actual spirometry results. Statistical measures such as Pearson's correlation coefficient (r), intraclass correlation coefficient (ICC), mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) were calculated. Receiver operating curve analysis was used to compare the model's predictions with spirometry values for FVC and FEV1 thresholds.

Results

The AI model demonstrated high agreement with spirometry results, indicating its potential as an alternative method for assessing pulmonary function. Saliency maps were generated to visualize the regions of chest x-rays that the AI model deemed important for predicting FVC and FEV1. These maps were created using SHapley Additive exPlanations (SHAP), a method based on cooperative game theory that enhances the interpretability of machine learning models.

Discussion

Clinical Implications

The AI model offers several clinical advantages. It can serve as an alternative for patients who cannot undergo spirometry, such as young children, older adults, and individuals with physical or cognitive disabilities. Chest x-rays are less time-consuming and more reproducible than spirometry, making them a viable option for estimating respiratory function in these populations. Additionally, the model's ability to estimate pulmonary function from chest x-rays could improve diagnostic accuracy by allowing for the customization of subsequent imaging protocols, such as CT scans.

Comparison with Previous Studies

Previous studies have explored the correlation between dynamic digital radiography and pulmonary function. However, these studies were limited by small sample sizes and single-centre designs. In contrast, this study utilized a multicentre approach and focused on static chest x-rays, which are more widely available and require less radiation exposure. The AI model's performance was comparable to or better than previous methods, demonstrating its potential for broader clinical application.

Limitations and Future Directions

The study has several limitations. As a retrospective study, it lacks the rigor of a prospective design. The model was developed and validated using data from Japanese institutions, which may limit its generalizability to other ethnic and racial groups. Future studies should include diverse populations to validate the model's performance across different settings. Additionally, the model's predictions exhibited variability when assessing subgroups of patients with specific lung diseases, such as COPD and asthma. Ongoing research is needed to improve the accuracy and reliability of AI-based pulmonary function estimations.

Conclusion

This study demonstrates the potential of a deep learning AI model to estimate FVC and FEV1 from chest x-rays, offering a promising alternative to traditional spirometry. The model's high agreement with spirometry results and its applicability to general posterior-anterior view chest x-rays make it a valuable tool for assessing pulmonary function, particularly in populations where spirometry is challenging. Future research should focus on validating the model in diverse populations and integrating clinical information to enhance its accuracy and utility in clinical practice.

Read Paper │The Lancet Digital Health