Clinical factors associated with CT-confirmed pneumonia in febrile patients: a retrospective study addressing verification bias with inverse probability weighting

Abstract Background Identifying pneumonia among febrile patients is clinically important, particularly in real-world settings where chest computed tomography (CT) is not performed uniformly. We aimed to identify clinical factors associated with CT-confirmed pneumonia in febrile patients and to examine the robustness of these associations after accounting for potential verification bias due to non-random CT completion. Methods This retrospective observational study included 902 febrile patients from Huadong Hospital between December 2022 and May 2023. The primary analysis was restricted to patients who underwent chest CT, among whom CT findings were categorized as no pneumonia ( n = 335) or pneumonia ( n = 210); 357 did not undergo CT. Baseline characteristics were compared using the Wilcoxon rank-sum test or chi-square test, as appropriate. Right-skewed biomarkers, including C-reactive protein (CRP), neutrophil-to-lymphocyte ratio (NLR), and lymphocyte-to-monocyte ratio (LMR), were transformed using log(x + 1). Candidate predictors were screened using univariable logistic regression and 10-fold cross-validated least absolute shrinkage and selection operator (LASSO) regression, followed by a non-penalized multivariable logistic regression model to estimate adjusted odds ratios (ORs) and 95% confidence intervals (CIs). Model performance was evaluated by discrimination, calibration, decision curve analysis, and bootstrap internal validation (B = 1000). To address potential verification bias arising from selective CT completion, stabilized inverse probability weighting (IPW) was applied as a sensitivity analysis. Supplementary analyses evaluated the incremental discriminative value of laboratory biomarkers beyond respiratory symptoms and assessed sensitivity to a stricter pneumonia outcome definition. Results Among patients who underwent CT, those with CT-confirmed pneumonia were older and had higher CRP levels than those without pneumonia. The final multivariable model included age, sex, body temperature, monocyte percentage, eosinophil percentage, log-transformed CRP, and log-transformed LMR. Older age (OR 1.02, 95% CI 1.01–1.03, p < 0.001) and higher log-transformed CRP (OR 1.42, 95% CI 1.18–1.70, p < 0.001) were robustly associated with CT-confirmed pneumonia. A positive association was also observed for higher log-transformed LMR (OR 1.86, 95% CI 1.10–3.15, p = 0.021), but this association was sensitive to parameterization and should be regarded as exploratory. The model showed modest discrimination, with an apparent area under the receiver operating characteristic curve (AUC) of 0.691 and a bootstrap-corrected AUC of 0.673 (95% CI 0.630–0.716); the Brier score was 0.211. The final model outperformed CRP, NLR, and white blood cell count alone in discrimination. Respiratory symptoms were nearly identically distributed between patients with and without pneumonia and did not improve discrimination beyond demographics, whereas adding laboratory biomarkers significantly improved the AUC from 0.644 to 0.697 ( p = 0.003). In IPW-adjusted analyses, the estimates for age and log-transformed CRP remained materially unchanged. Sensitivity analyses showed that the exploratory association involving LMR was dependent on parameterization, remaining directionally consistent in log-transformed and categorized models but not significant when modeled as a raw continuous variable.; a stricter outcome definition requiring compatible respiratory symptoms or hypoxemia yielded directionally consistent estimates for the main predictors. Conclusions Older age and higher inflammatory burden, particularly elevated CRP, were independently associated with CT-confirmed pneumonia in febrile patients. Adjustment for verification bias using IPW supported the robustness of the main findings. A multivariable model based on these factors showed modest discrimination and may assist with preliminary risk stratification, whereas the positive association involving LMR was parameterization-dependent and should be interpreted as an exploratory finding.