Multi-population generalizability of a deep learning-based chest radiograph severity score for COVID-19

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Medical Diagnostics
• Infectious Disease Research

Background:

• Deep learning models show promise for assessing COVID-19 lung disease severity.
• Generalizability of these models across diverse patient populations remains a key challenge.
• Chest radiographs (CXRs) are crucial for evaluating COVID-19 related pulmonary complications.

Purpose of the Study:

• To tune and test the generalizability of a deep learning model for COVID-19 lung disease severity assessment on chest radiographs (CXRs).
• To evaluate the model's performance on diverse patient populations from different geographical locations.

Main Methods:

• A convolutional Siamese neural network was tuned using outpatient CXRs.
• The model generated a pulmonary x-ray severity (PXS) score.
• Performance was evaluated on four test sets (US academic, US community, US outpatient, Brazil emergency) using radiologist-assigned severity scores (Pearson R) and Uniform Manifold Approximation and Projection (UMAP) for visualization.

Main Results:

• Tuning with outpatient data improved model performance on US hospitalized datasets (R = 0.88, 0.90).
• Performance remained high on US outpatient (R = 0.86) and Brazil emergency department (R = 0.85) datasets.
• UMAP visualization confirmed that the model learned generalizable disease severity information across test sets.

Conclusions:

• A deep learning model can accurately quantify COVID-19 lung disease severity on CXRs.
• The model demonstrates generalizable performance across multiple patient populations and continents.
• This AI tool has potential for consistent COVID-19 severity assessment in varied clinical settings.