AI Predicts Likelihood of Death from Chest Images Years in Advance

Chest radiography, such as x-ray imaging, is essential in diagnosing many conditions and is the most common imaging test in healthcare. In 2013, there were 1,039 outpatient chest radiographs done per 1,000 US Medicare Part B members, highlighting the frequency of this procedure in older patients. These images typically come back with no indication of major disease, but even these “normal” radiographs can indicate subtle issues like enlarged heart or arterial stiffening.

Being that the physician does not often see what their patients’ outcomes are decades after reviewing these radiographs, it’s hard to say which subtleties of these images have long-term value. To address this discrepancy, researchers hypothesized that a deep learning artificial intelligence (AI) model could be used to predict 12-year mortality from chest radiographs. Their findings were published on July 19 in JAMA Network Open.

Creating a Deep Learning Solution

To test this hypothesis, the team created a convolutional neural network called CXR-risk. This platform was created and tested via data from the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial (PLCO) and National Lung Screening Trial (NLST), two multicenter clinical trials of chest radiography screening. This data stemmed from both smokers and nonsmokers, with exclusion criteria including history of various cancers. The primary outcome was mortality of any cause.

The AI algorithm was trained to evaluate patient radiographs and predict the likelihood of long-term mortality based on this data. This prediction was generated in the form of a CXR-risk score, which identified patients at low and high risk for mortality.

Results of the CXR-Risk Study

The patients that the deep learning model predicted to be at a very high risk had a 53% mortality rate at 12 years in the PLCO dataset and 34% at 6 years in the NLST set. This risk score was found to be complementary to the diagnoses made by radiologists and standard risk factors such as age, diabetes, and sex. CXR-risk scores were independently associated with lung cancer deaths as well as cardiovascular and respiratory deaths in the PLCO and NLST datasets, respectively.

To the best of the authors’ knowledge, this was the first report of an AI model predicting long-term outcomes from chest radiographs. They note that these results support the use of this CNN in identifying patterns on chest radiographs that aren’t tied to one single disease, but rather are indicative of one’s overall health.

This deep learning model takes a fraction of a second to make a long-term prediction based on a single radiograph. The researchers mention that future studies with the CXR-risk system could be enhanced to detect specific disease outcomes, rather than just predict overall mortality risk. The clinical effects of this platform could help guide lifestyle, screening, and prevention decisions. The authors claim that further clinical trials are needed to determine the effect of this decision making on patient outcomes.

It is important to note that this CNN was only tested and developed in asymptomatic patients aged 55-74 years and that the results may not extend to symptomatic populations. These radiographs were all done in posterior-anterior fashion, and the relevance of this to other radiographic approaches is unknown.

“The results suggest that the CXR-risk CNN can stratify the risk of long-term mortality using chest radiographs,” the authors concluded. “Individuals at high risk may benefit from prevention, screening, and lifestyle interventions. Further research is necessary to determine how this can improve individual and population health.”

The leading authors of the study were Michael T. Lu, MD, MPH, Alexander Ivanov, BS, and Thomas Mayrhofer, PhD, all associated with the Department of Radiology at the Massachusetts General Hospital. Mayrhofer is also affiliated with the Stralsund University of Applied Sciences in Germany.