Deep Learning to Assess Long-term Mortality From Chest Radiographs

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Prognostic Modeling

Background:

• Chest radiography is a ubiquitous diagnostic tool with potential prognostic capabilities.
• Predicting long-term mortality from standard imaging is an area of active research.
• Existing risk assessment models often rely on clinical data, not imaging features.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) named CXR-risk.
• To predict long-term all-cause mortality, including noncancer deaths, using chest radiographs.
• To assess the prognostic value of the CXR-risk score against traditional risk factors and radiologist findings.

Main Methods:

• Development and testing of a CNN (CXR-risk) using data from the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial (PLCO).
• External validation using data from the National Lung Screening Trial (NLST).
• Analysis of all-cause and cause-specific mortality, adjusting for clinical risk factors and radiologist interpretations.

Main Results:

• A graded association was observed between the CXR-risk score and mortality across both PLCO and NLST cohorts.
• Very high-risk individuals exhibited significantly higher mortality rates compared to very low-risk individuals (e.g., PLCO unadjusted HR: 18.3).
• The CXR-risk score's prognostic value remained robust after adjusting for clinical risk factors and radiologist findings.

Conclusions:

• The deep learning CXR-risk score effectively stratifies long-term mortality risk from a single chest radiograph.
• This AI-driven tool offers a novel approach to identify individuals who may benefit from targeted interventions.
• Chest radiograph analysis using AI holds promise for enhancing prognostic assessments in clinical practice.