Chest radiograph-based artificial intelligence predictive model for mortality in community-acquired pneumonia

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pneumonia Research

Background:

• Chest radiography (CXR) is a fundamental diagnostic tool for community-acquired pneumonia (CAP).
• CXR possesses prognostic value in assessing CAP outcomes.
• There is a need for improved prognostic tools in CAP management.

Purpose of the Study:

• To develop and prospectively evaluate a CXR-based artificial intelligence (AI) model for predicting 30-day mortality in CAP patients.
• To assess the performance of the CAP AI predictive Engine (CAPE) against established pneumonia severity scores.

Main Methods:

• A deep-learning model (convolutional neural network - CNN) was trained on 2235 retrospective CXRs from adult CAP patients.
• Model performance was evaluated prospectively on a cohort of 315 inpatient episodes.
• The CAPE mortality risk score was derived from CNN analysis of initial CXRs.

Main Results:

• The CAPE model achieved an Area Under the Curve (AUC) of 0.79 for 30-day mortality prediction.
• CAPE demonstrated comparable performance to Pneumonia Severity Index (PSI) (AUC 0.80) and CURB-65 (AUC 0.76).
• Combining CAPE with PSI yielded the highest AUC of 0.84, indicating improved discrimination.

Conclusions:

• A CXR-based AI model (CAPE) can effectively predict 30-day mortality in CAP.
• The CAPE model's performance is comparable to traditional pneumonia severity scores.
• Integrating CAPE with existing scores, particularly PSI, enhances mortality prediction accuracy.