A deep-learning (DL) model that analyzes the initial chest x-rays of patients who have community-acquired pneumonia (CAP) may predict the risk of death by 30 days more accurately than an established risk-prediction tool, finds a new study published in the American Journal of Roentgenology.
For the study, a team led by Seoul National University Hospital researchers evaluated the ability of a DL model developed using 7,105 CAP patients at a single center from March 2013 to December 2019 to predict risk of all-cause death by 30 days.
The team tested the model using the initial x-rays of emergency department (ED) patients at the same center as the development cohort from January to December 2020 (temporal test cohort, 947 patients), a second center (external test cohort A, 467) over the same period, and a third center (external test cohort B, 381) from March 2019 to October 2021.
The results of the DL model and the established CURB-65 risk-prediction tool were compared, and the combination of the two tests was assessed using a logistic regression model.
Higher specificity at same sensitivity
The area under the curve (AUC, a measure of diagnostic accuracy) for risk of death by 30 days was higher for the DL model than for CURB-65 in the temporal test cohort (0.77 vs 0.67), but the result wasn't statistically significant in external test cohort A (0.80 vs 0.73) or B (0.80 vs 0.72).
The DL model showed higher specificity (range, 61% to 69% vs 44% to 58%) at the same sensitivity as the CURB-65 score in all three groups. Combined, the DL model and CURB-65 scores increased the AUC in the temporal cohort (0.77) and external cohort B (0.80) but the increase was not significant in cohort A (0.80).
The deep learning (DL) model may guide clinical decision-making.
"The deep learning (DL) model may guide clinical decision-making in the management of patients with CAP by identifying high-risk patients who warrant hospitalization and intensive treatment," coauthor Eui Jin Hwang, MD, PhD, of Seoul National University, said in an American Roentgen Ray Society news release.