New AI Solution Identifies High-Risk Patients From Chest X-Rays
Chest X-rays may hold more information that could identify patients at high risk for a range of conditions, including heart disease and lung cancer. That’s the consensus arrived at by researchers at Massachusetts General Hospital, who have developed a new AI tool called CXR-risk to determine which patients would benefit the most from screening and preventive medication. “We developed and tested this convolutional neural network in healthy outpatients. The hope is that it will help primary care physicians and patients make decisions about prevention, screening, and lifestyle,” Dr. Michael T. Lu, director of research of MGH Cardiovascular Imaging, told HCB News. The underlying idea is that we can use CNNs to extract information about health and longevity embedded in everyday medical images.”
Trained on analyses of more than 85,000 chest X-rays from 42,000 subjects, the solution is designed to identify combinations of features on a chest X-ray that best predict health and mortality. Each image used to train it included information on whether or not the person died at any point over a 12-year period. Applying CXR-risk to the chest X-rays of 16,000 patients from two earlier clinical trials, researchers found that 53 percent of people identified by the CNN as “very high risk” died over twelve years, compared to fewer than four percent of those that CXR-risk labeled as “very low risk”. In addition, the solution provided information for predicting long-term mortality, independent of radiologists’ readings of X-rays and other factors, such as age and smoking status. Lu believes the system can be applied to other types of scans and will be more accurate when combined with other risk factors, such as smoking status and genetics. He notes that earlier identification of high-risk patients could lead to greater preventive and treatment programs. “We focus on chest x-rays because they are so common, but I believe there will be similar prognostic information from other types of imaging,” he said. The findings were published in the journal, JAMA Network Open.