Predicting Respiratory Disease Mortality with Chest X-Ray
AI model tracks lung risk over time, validated in an Asian population
Identification of patients with chronic respiratory diseases, who are at high-risk for mortality, is crucial for improving patient outcomes. Worldwide, chronic respiratory diseases are ranked as the third leading cause of mortality. The ability of chest radiographs to uncover subtle anatomic changes makes them a useful risk stratification tool that can be quantified using AI models.
A study in Radiology: Artificial Intelligence assessed the prognostic value of a model that combines chest X-rays with AI to predict respiratory disease mortality in a cohort of Asian individuals.
Research has shown that one such model, CXR-Lung-Risk, can estimate the risk of mortality due to respiratory diseases from a single chest radiograph. As an open-source model, CXR-Lung-Risk has potential for integration into clinical practice for screening purposes.
While the model’s prognostic value has been studied on chest radiographs from a single time point, refining risk assessment with follow-up chest radiographs or further stratifying risk over time through an AI model represents a compelling avenue for research.
Feasibility of AI-Based Screening in Low-Risk Populations
For the new study, researchers studied this potential with trajectory analysis, a statistical method that helps show how a biomarker or variable changes over time.
“We used this analysis to identify trends such as increasing or decreasing patterns in the CXR-Lung-Risk score from serial chest radiographs,” said study senior author Hyungjin Kim, MD, PhD, from the Department of Radiology, Seoul National University Hospital and College of Medicine in South Korea. “This helps stratify individuals into different risk groups based on how their scores evolve longitudinally.”
Dr. Kim and colleagues evaluated the CXR-Lung-Risk model’s prognostic relevance for respiratory disease mortality within an Asian health screening population.
They analyzed chest radiographs from individuals who underwent health screenings between January 2004 and June 2018.
Of the 36,924 individuals in the study group, 264 (0.7%) died of respiratory illness over a median follow-up period of 11 years.
The model successfully identified individuals at high risk for respiratory mortality using a single chest radiograph. Trajectory analysis over a three-year follow-up period revealed that individuals with increasing risk scores had significantly worse outcomes.
“These findings support the feasibility of AI-based risk stratification,” Dr. Kim said.
Because the study population consisted of asymptomatic individuals at low risk for respiratory disease mortality, the findings show that risk stratification is feasible even in seemingly healthy individuals, underscoring the potential of the AI model as a screening tool.
The CXR-Lung-Risk model was originally developed using chest radiographs from a clinical trial conducted in the U.S., where most participants were white. The new study evaluated the model on a population that differed geographically, temporally and racially from the test datasets used in earlier research. This diversity enhances the model’s potential for generalizability, Dr. Kim said.
“Moreover, self-paid health check-ups that include chest radiography are widely practiced in East Asian countries, making this population a relevant and practical target for external testing of the model,” he said.
“An AI model may reveal differences in mortality risk not apparent through conventional means, prompting clinicians to adjust the frequency or intensity of follow-up accordingly. This could help allocate resources more efficiently and improve outcomes through tailored management.”
— EDUARDO MORENO JÚDICE DE MATTOS FARINA, MD
Potential of AI Hinges on Clinical Validation
The study results show the transformative potential of deep learning in medicine, according to Eduardo Moreno Júdice de Mattos Farina, MD, a neuroradiologist and pediatric radiology fellow at Universidade Federal de São Paulo, and Paulo Eduardo de Aguiar Kuriki, MD, neuroradiologist and assistant professor at UT Southwestern Medical Center in Dallas.
In a commentary accompanying the study, Drs. Farina and Kuriki praised the research for validating a publicly available AI tool in an external population and correlating it with clinical data and relevant outcomes, especially respiratory disease mortality and all-cause mortality.
AI models for prognosis like the one examined in the study can support more personalized health care planning, Dr. Farina noted. For instance, two patients with chronic obstructive pulmonary disease (COPD) of the same age and gender with similar lab results might traditionally receive similar care.
“However, an AI model may reveal differences in mortality risk not apparent through conventional means, prompting clinicians to adjust the frequency or intensity of follow-up accordingly,” he said. “This could help allocate resources more efficiently and improve outcomes through tailored management.”
The key challenge going forward, Dr. Farina said, lies in validating whether model-guided interventions genuinely improve patient outcomes without introducing new biases.
“We need randomized controlled trials where care decisions are based on AI predictions to assess their real-world impact,” Dr. Farina said. “Without this level of evidence, the model’s output risks becoming just another number without clinical consequence.”
Dr. Kim echoed Dr. Farina’s call for more research.
“This study focused on mortality, which is an important but relatively abstract outcome,” he said. “For clinical implementation, further research is needed to link the CXR-Lung-Risk score to more actionable endpoints, such as the incidence of specific respiratory diseases or the impact of targeted interventions.”
For More Information
Access the Radiology: Artificial Intelligence article, “Predicting Respiratory Disease Mortality Risk Using Open-Source AI on Chest Radiographs in an Asian Health Screening Population,” and the related commentary, “Predicting Mortality with Deep Learning: Are Metrics Alone Enough?”
Read previous RSNA News stories on chest imaging: