Researchers have developed a convolutional neural network (CNN) that predicts long-term mortality from a single chest x-ray, according to new findings published in JAMA Network Open. The full assessment takes less than half a second.
“Even normal radiographs manifest additional minor abnormalities, such as aortic calcification or an enlarged heart, that may provide a new window into prognosis and longevity with the potential to inform decisions about lifestyle, screening, and prevention,” wrote Michael T. Lu, MD, MPH, department of radiology at Massachusetts General Hospital in Boston, and colleagues. “Whereas physicians may interpret thousands of chest radiographs during a career, they rarely know the outcomes in these patients a decade later. Therefore, it is difficult to develop an intuition to articulate which features have long-term prognostic value.”
The team’s CNN, known as CXR-risk, was developed and tested using data from the Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial (PLCO). External testing was carried out using data from the National Lung Screening Trial (NLST). The study’s primary outcome was all-cause mortality, and patients participating in the trials were followed up for up to 13 years (PLCO) or eight years (NLST).
The CNN’s output was a continuous CXR-risk probability, which was “converted to an ordinal CDR-risk score based on quantile thresholds set in the PLCO development data set and then applied to the PLCO and LST test data sets.” The top 75th-95th percentile was viewed as “high risk” and the top 95th percentile was viewed as “very high risk.”
Out of the PLCO data set, mortality rates were 24.9% among “high risk” patients and 53% among “very high risk” patients. In the NLST data set, mortality rates were 9.8% among “high risk” patients and 33.9% among “very high risk” patients. The authors noted a definite “graded increase in mortality with increasing CXR-risk score.”
Lu et al. also observed that the CNN’s output was “complementary” to the diagnostic findings of radiologists and “standard risk factors.” Also, there were associations between the CXR-risk score and lung cancer death, noncancer cardiovascular death and respiratory death. The researchers found that most patient deaths were from causes other than lung cancer.
“These observations suggest that this CNN should not be considered as a lung cancer detector,” they wrote. “Instead, we speculate that it identified patterns on the chest radiograph not tied to a single diagnosis or disease but as a summary measure of underlying prognosis and health. This concept of shared risk factors has been established for other biomarkers.”
The authors concluded by looking at the potential of this research. Their risk score only used a single x-ray from a patient, but there is an opportunity to take the deep learning model to a whole other level.
Michael has more than 16 years of experience as a professional writer and editor. He has written at length about cardiology, radiology, artificial intelligence and other key healthcare topics.