In a recent study published in the American Journal of Roentgenology, researchers used diagnostic chest radiographs as input to estimate the 30.0-day risk of death in patients with community-acquired pneumonia (CAP). developed a deep learning (DL) model to We validated model performance across patients from different centers and durations.
study: A deep learning model using chest radiographs for predicting 30-day mortality in patients with community-acquired pneumonia: development and external validation. Image credit: AndreySuslov/Shutterstock.com
Background
CAP, a common cause of pneumonia, is associated with significant mortality and medical resource utilization. Chest radiography is an essential tool for diagnosing and risk stratifying CAP.
However, the incorporation of chest radiographic findings into risk prediction tools has been limited by inter-reader variability and the difficulty of extracting objective biomarkers. The CURB-65 score and Pneumonia Severity Index are currently available tools to predict adverse outcomes in patients with CAP.
About research
In this retrospective study, investigators used primary chest radiographs to develop and externally validate a DL-based model that predicts death within 30.0 days in patients with CAP.
This model was developed to predict the 30.0-day all-cause mortality risk in CAP patients using the first chest radiograph.
This study included searching electronic medical records (EMRs) of a single tertiary referral agency for individuals with a CAP diagnosis during a medical practice between March 2013 and December 2019. rice field.
The team evaluated the deep learning model on individuals diagnosed with community-acquired pneumonia in the emergency department of the facility where the development group was diagnosed between January and December 2020 (temporary test group, 947 people). bottom.
They also modeled at two other facilities, Seoul National University Boramae Medical Center (external examination group A, 467 persons) and Chung-Ang University Hospital (external examination group B, 467 persons) from January to March 2020. evaluated. 381 people from March 2019 to October 2021).
The development cohort included patients diagnosed with CAP during any encounter, whereas the subsequent study cohort included only patients diagnosed with CAP during an emergency department encounter. The team compared area under the curve (AUC) values between the deep learning model and his CURB-65 tool and evaluated the results of the combined approach by logistic regression modeling.
The primary endpoint was mortality from any cause within 30.0 days after CAP diagnosis. A convolutional neural network (CNN) was developed to predict her 30.0-day mortality after his CAP diagnosis based on chest radiographic scans from patients in a developmental cohort.
The output of the model represented the conditional survival probabilities at various time intervals and an experienced thoracic radiologist performed a post hoc analysis of the class activation maps.
A deep learning model was devised with a participant distribution of 3.0:1.0:1.0 for the training, validation, and internal test groups, and for 30.0 days of all-cause mortality in CAP patients whose chest radiographic scans at the time were analyzed. We estimated the risk of death from As input for diagnostics.
Mortality data were confirmed using the Korean Ministry of Home Affairs and Security’s EMR, or death registration data.
result
This study analyzed 30-day mortality in 1,421 patients in a developmental cohort, including 1,421 patients in the internal test set.
AUC values for the estimated 30.0-day mortality risk were greater for the deep learning model compared to CURB-65 among temporal test group participants (0.80 vs. 0.70), but not among participants belonging to external test group A. It was not statistically significant between (0.8 vs 0.8) 0.7) and B (0.8 vs 0.7).
Compared to CURB-65, the DL model had similar sensitivity but higher specificity, with a positive predictive value (PPV) of 35% vs. 18% and a negative predictive value (NPV) of 95% vs. It was 94%. The DL model showed acceptable calibration in the interim test group, but significantly overestimated his 30-day risk of death in the external test cohorts A and B.
The DL model was a significant predictor of 30-day mortality, with an odds ratio of 1.08 for a 1.0% increase in predicted risk after adjusting for CURB-65 score.
In external test groups A and B, the DL model, the CURB-65 score, and the combined model showed qualitatively similar decision curves, indicating a net positive advantage of the deep learning and combined models compared to the CURB-65 score. improved slightly. .
Pneumonia images influenced the predictions of the DL model for high-risk prediction patients, whereas images for low-risk prediction patients were influenced by other regions of the image.
The model was unaffected by extraneous features such as radiographic markers or extraneous substances. Post-hoc evaluation of class activation maps showed that the predictions of the DL model were nearly accurate.
Conclusion
Overall, the study results demonstrate that a deep learning model can estimate mortality within 30.0 days after CAP diagnosis with better performance than the CURB-65 tool using chest radiographs obtained for diagnosis. showed.
This model yielded AUC 0.77–0.80 with higher specificity (range 61%–69%) compared to CURB-65 (range 44%–58%) with similar sensitivity.
This model can guide decision-making and improve CAP outcomes by identifying high-risk patients (those requiring hospitalization or intensive care, such as intravenous antibiotics or respiratory support). .
In contrast, early home discharge and conservative treatment for low-risk patients can reduce unnecessary medical resource utilization.
