Deep Learning Model Accurately Predicts Mortality in Pneumonia Patients

The American Journal for Roentgenology published original medical research last month that described how an artificial intelligence tool accurately predicted the severity of community-acquired pneumonia (CAP) and estimated the 30-day mortality in patients better than traditional methods.

Researchers from Seoul National University in South Korea trained their model using 7,000+ X-ray images from CAP patients. In testing, the model was asked to offer predictions for the risk of death among 2,000 medical images collected over three years. Results showed that the deep learning (DL) model improved diagnostic accuracy compared to the established CURB-65 risk prediction tool, which is used to stratify patients by expected mortality.

Using test cohorts, the DL model achieved a higher specificity range when predicting 30-day mortality rates. Results showed that the DL model generated a specificity of 61–69% compared to CURB-65, which achieved a range of 44–58%. Additionally, after averaging the DL model and CURB scores, researchers improved diagnostic accuracy for two of the three cohorts included in the study.

"The deep learning (DL) model may guide clinical decision-making in managing patients with CAP by identifying high-risk patients who warrant hospitalization and intensive treatment," said coauthor Eui Jin Hwang, MD, PhD, of Seoul National University.

Last year, Harvard Medical School researchers successfully used an artificial intelligence model to detect pathologies in chest X-ray images. That tool, called Chexzero, could detect pathologies as effectively as human radiologists without disease-specific information or manual annotations.

Earlier integrations of AI into the medical imaging process often involved improving the images generated by a radiologist and their machine. An example is GE Healthcare’s AIR Recon DL tool which uses a deep learning algorithm to de-noise images and speed up imaging. As researchers progress in the field, they seek to assist radiologists in more aspects of the workflow, including reading images and making predictions. 

“An advantage of using chest radiographs as an input to the DL model compared to the use of clinical variables is that chest radiographs may allow automated processing that is not influenced by subjective evaluation by physicians,” wrote the Korean study’s authors.

A better forecast of mortality can do more than save lives; it can also help physicians manage CAP, which is one of the most costly infectious diseases faced in healthcare. According to one analysis of CAP patients in US clinical settings, care for hospitalized CAP patients cost $40,000 on average and can rise to more than $80,000 for patients in long-term settings.