Machine-Learning Model for Mortality Prediction in Patients With Community-Acquired Pneumonia: Development and...

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Biostatistics

Background:

• Machine learning (ML) offers potential to enhance clinical prediction tools, but its application in community-acquired pneumonia (CAP) mortality prediction is understudied.
• Existing prognostic scores for CAP have limitations in predictive accuracy.

Purpose of the Study:

• To apply and validate a causal probabilistic network (CPN) model, SeF-ML, for predicting 30-day mortality in CAP patients.
• To compare the predictive performance of SeF-ML against established scoring systems like PSI, SOFA, qSOFA, and CURB-65.

Main Methods:

• A retrospective derivation-validation study was conducted across two Spanish university hospitals.
• A CPN model (SeF-ML) was developed and validated for CAP mortality prediction.
• Performance was evaluated using receiver operating characteristic (ROC) curves and areas under the curve (AUCs), with comparisons made using the DeLong method.

Main Results:

• In the validation cohort, SeF-ML achieved an AUC of 0.826, consistent with its performance in the derivation cohort (AUC 0.801).
• SeF-ML demonstrated significantly higher predictive accuracy than CURB-65 (AUC 0.764) and qSOFA (AUC 0.729).
• No significant difference was found between SeF-ML and PSI (AUC 0.830) or SOFA (AUC 0.771) in predicting 30-day mortality.

Conclusions:

• The SeF-ML model shows potential for improving mortality prediction in CAP patients utilizing structured health data.
• External validation studies are necessary to confirm the generalizability of SeF-ML across diverse patient populations and healthcare settings.