Pneumonia remains a leading cause of morbidity and mortality globally, necessitating reliable clinical prediction tools to guide medical management decisions. The Pneumonia Severity Index (PSI), CURB-65, and CRB-65 are clinical scoring tools used to assess the severity of community-acquired pneumonia (CAP), aiding in risk stratification and guiding decisions on hospitalization, level of care and prognosis. Comparative data on their utility, specifically in immunocompetent patients hospitalized in internal medicine wards, are limited. This study aimed to evaluate the predictive capabilities of these scoring tools for mortality and intensive care unit (ICU) transfers in a large cohort of hospitalized patients. We conducted a retrospective, single-center cohort study including 12,670 immunocompetent patients hospitalized with pneumonia in the internal medicine division. PSI, CURB-65, and CRB-65 performance was compared across multiple outcomes, including in-hospital mortality, 30-, 60-, 90-day mortality, and ICU transfer from ward. Subgroup analyses were performed for key comorbidities (chronic obstructive pulmonary disease [COPD], congestive heart failure [CHF], diabetes, chronic kidney disease [CKD] and hypertension). PSI consistently demonstrated significantly superior discrimination between survivors and non-survivors across all mortality outcomes with AUC range of 0.73–0.75 (p < 0.001, FDR-corrected). In the subgroup analysis by comorbidities, PSI was significantly superior to the other scoring systems only in diabetes patients in 60 and 90-day mortality (AUC = 0.70–0.71). CURB-65 performed comparably to PSI in most of the cases and was superior to CRB-65 only in diabetes and hypertension patients. When predicting ICU transfer during hospitalization, there were no significant differences between the scoring tools, and all demonstrated low predictive capability. The PSI demonstrates superior discriminative ability in immunocompetent patients hospitalized with pneumonia. However, its greater complexity should be considered when evaluating its practicality for routine use. The online version contains supplementary material available at 10.1186/s41479-025-00191-x.

Background: Community-acquired pneumonia (CAP) is a leading cause of global mortality, yet established mortality risk scores often fail to balance clinical simplicity with adequate discrimination. Aims and objectives: We aimed to develop parsimonious machine-learning model based on commonly available clinical features for predicting 30-day mortality in hospitalized CAP patients. Methods: We conducted a retrospective cohort study of 12,670 immunocompetent adults hospitalized with CAP at a tertiary center (2007–2025). Twenty-four routinely available admission variables were used to train six ML models (including logistic regression, LASSO, elastic net, random forest, gradient boosting machine, and XGBoost). Recursive feature elimination was applied to derive a minimal feature set. Model performance was compared with PSI and CURB-65 and externally validated using the MIMIC-IV database. Results: The 24-feature model demonstrated the highest discrimination (AUC 0.819, 95% CI 0.80–0.84), significantly outperforming PSI (AUC 0.722) and CURB-65 (AUC 0.692; both p<0.001). To ensure clinical utility, we prioritized the highest-ranking features (Urea, Sodium, Albumin, Age—USA₂). The USA₂ model achieved an AUC of 0.778, retaining 95.0% of the full model’s performance while reducing complexity. External validation on the MIMIC-IV cohort confirmed generalizability (AUC 0.73–0.76), despite substantial missing albumin values. Conclusions: USA₂ is a parsimonious ML model using four routinely available variables that significantly outperforms established clinical scores for predicting 30-day mortality in CAP.

Background: Bronchoalveolar lavage (BAL) is frequently employed for diagnostic purposes in immunocompromised patients with pneumonia, yet its role in immunocompetent individuals remains debated. The study investigates whether BAL is associated with reduced mortality in immunocompetent patients hospitalized with pneumonia. Methods: A retrospective cohort study was conducted at SMC, including 13,180 immunocompetent patients hospitalized with pneumonia from 2007 to 2024. Patients who underwent BAL (n = 96) were matched 1: 4 to a control group (n = 384) using Propensity score matching based on age, gender, severity scores and comorbidities. Mortality was assessed at 30, 60, and 90 days using logistic regression models, adjusting for length of stay, inflammatory markers, albumin levels, smoking status, and BMI. Results: In the matched cohort, 30-day mortality did not differ significantly between groups (20.8% in BAL vs. 19.5% in controls, p = 0.886), and no significant association was observed after adjustment (OR = 0.95 [0.51–1.73]. However, both 60-day and 90-day mortality were significantly higher in the BAL group (39.6% vs. 24.7% and 45.8% vs. 26.8%, respectively; p < 0.01 for both), and these differences persisted after adjustment (60 days: OR = 1.74 [1.03–2.94]; 90 days: OR = 2.12 [1.27–3.54]). Conclusions: In this cohort of immunocompetent patients hospitalized with pneumonia, BAL was not associated with short-term survival. Clinicians should carefully weigh the potential risks and benefits of BAL in this population. Further studies are needed to identify patient subgroups that might benefit from BAL through enhanced diagnostic or therapeutic approaches.