Dynamically monitoring pneumonia severity scores to predict the prognosis in patients with community-acquired pneumonia: an international, multicenter cohort study.

Background: Traditional pneumonia severity scores are widely used at treatment initiation to stratify disease severity and predict prognosis. Although they can be reassessed during treatment to evaluate disease progression, their predictive accuracy remains suboptimal. To address this limitation, we introduce a novel machine learning algorithm that quantifies continuous disease progression, enhancing real-time evaluation of treatment efficacy and outcome prediction in community-acquired pneumonia (CAP).

Methods: We analyzed 2052 CAP patients from China and 1862 from Japan. CURB-65 and pneumonia severity index (PSI) scores were recorded at admission and 72 h post-treatment. To quantify disease progression, we developed a new algorithm, Mortality Vector Optimization (MVO), which measures the similarity between a patient's condition and critical states within severity scoring systems. Model discrimination was assessed using the C-index and AUC, while calibration was evaluated with Brier scores. External validation was conducted using the GOSSIS-1-eICU and SCRIPT CarpeDiem datasets.

Results: MVO demonstrated superior predictive performance compared to nine established machine learning models. In the derivation cohort, MVO achieved a C-index of 0.734 and AUC of 0.825 for PSI, while in the GOSSIS-1-eICU dataset, it reached a C-index of 0.850 and AUC of 0.870. The trend score derived from MVO effectively captured disease progression, with hazard ratios of 2.63 (for CURB-65) and 2.50 (for PSI) in predicting in-hospital mortality.

Conclusion: The trend score provides a dynamic and reliable method for monitoring early treatment effectiveness in CAP. MVO offers enhanced predictive accuracy and holds potential for broader applications in disease prognosis assessment.