Detecting Racial/Ethnic Health Disparities Using Deep Learning From Frontal Chest Radiography.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Cardiovascular Disease Epidemiology
• Health Disparities Research

Background:

• Atherosclerotic vascular disease (ASVD) prevalence estimation is crucial for patient management.
• Administrative codes may not fully capture ASVD burden, potentially leading to disparities.
• Deep learning models offer novel approaches for disease detection from medical imaging.

Purpose of the Study:

• To assess racial/ethnic and socioeconomic disparities in ASVD prevalence.
• To compare ASVD prevalence from a deep learning model on chest radiographs (CXRs) versus administrative codes in COVID-19 patients.
• To investigate factors associated with discrepancies between model predictions and administrative data.

Main Methods:

• A convolutional neural network (CNN) model trained to predict ASVD from CXRs was validated on two COVID-19 patient cohorts.
• CNN predictions were compared against electronic health record administrative codes.
• Discrepancies (Δ_vasc) were analyzed for associations with demographic and socioeconomic factors using regression models.

Main Results:

• The CNN model showed good performance in predicting ASVD (AUCs ranging from 0.69 to 0.85).
• Discrepancies (Δ_vasc) were significantly associated with non-English language preference and Black/Hispanic race in ambulatory patients.
• Social deprivation index and age were associated with discrepancies in hospitalized patients; Δ_vasc independently predicted the presence of administrative codes.

Conclusions:

• A CNN model can predict ASVD from CXRs in COVID-19 patients.
• Discrepancies between CNN predictions and administrative codes (Δ_vasc) are linked to health disparities.
• Biomarkers from imaging, compared with EHR data, may help address healthcare inequities for underserved populations.