A machine learning-based model for assessing community-acquired pneumonia severity using routine blood tests

The choice of first-line therapy for community-acquired pneumonia (CAP) depends on disease severity. However, quickly and accurately differentiating mild from severe CAP patients remains challenging. This study aims to evaluate the performance of machine learning-based diagnostic models employing routine blood indicators to distinguish CAP severity.

A multicenter, retrospective, case-control study conducted at Xuhui Central Hospital (Discovery cohort), and Putuo People's Hospital (Validation cohort), from January 2016 to January 2024. Patients were further classified into mild or severe CAP according to the IDSA/ATS criteria. Routine blood tests were performed with an automatic blood cell analyzer. Twelve machine learning-based diagnostic models were developed from routine blood indicators for differentiating between mild and severe CAP.

A total of 3,127 (1,612 mild, 1,615 severe) and 2,087 participants (1,072 mild, 1,015 severe) were included in the discovery and validation cohorts, respectively. Of the 12 models developed, the random forest (RF) model showed the best performance with 9 routine blood indicators. In the discovery cohort, the model achieved an AUC of 0.95, an AUPRC of 0.94, a positive predictive value of 0.89, a negative predictive value of 0.88, an accuracy of 0.89, and an F1 score of 0.89, while in the validation cohort, it demonstrated similar performance, with values of 0.95, 0.94, 0.88, 0.87, 0.88, and 0.87, respectively. Decision curve analysis confirmed consistent net benefits from the model across all threshold probabilities. The RF model was integrated into a web application for clinical use.

We successfully developed a nine-feature RF model with promising value for differentiating mild from severe CAP patients.