The Confounder in Plain Sight: A Retrospective Pilot Analysis on the Impact of Comorbidity on C-Reactive Protein Utility for Differentiating Bacterial vs. Viral Infections

Background: The antimicrobial resistance crisis is driven by antibiotic overuse, often due to the difficulty in distinguishing bacterial from viral infections. In the European Union, acute respiratory tract infections account for about 38% of all antibiotic prescriptions in community and emergency settings, and an estimated 30–50% of these prescriptions are potentially inappropriate. Point-of-care C-reactive protein (CRP) testing can support the distinction between bacterial and viral infections, but its diagnostic accuracy is often compromised by chronic inflammatory comorbidities that elevate baseline CRP levels. Objective: This exploratory, hypothesis-generating study evaluated the diagnostic utility of CRP in an Emergency Department (ED) cohort and proposed a novel “Comorbidity Confounder Score” (CCS) prototyped pilot tool as support to identify patient subgroups in whom CRP retains high diagnostic value. Methods: We conducted an exploratory, hypothesis-driven retrospective cohort study of 92 patients presenting to a tertiary ED with acute flu-like symptoms between 2023 and 2025. Microbiological diagnoses were confirmed using culture and PCR. ROC curve analysis and AUC comparisons were performed using the pROC package in R (v4.2.0; DeLong method). A post hoc power analysis confirmed 81% power at alpha = 0.05. The diagnostic performance of CRP (Area Under the Curve—AUC) was assessed in the total cohort and stratified into “Low-Utility” (high comorbidity, CCS ≥ 2) and “High-Utility” (low comorbidity, CCS < 2) subgroups. Results: In the unselected total cohort, CRP demonstrated suboptimal diagnostic performance (AUC = 0.61, 95% CI: 0.49–0.73). However, exploratory post hoc stratification revealed divergence. In the “Low-Utility” group, CRP had no diagnostic value (AUC = 0.52). In the “High-Utility” group, a preliminary signal of improved CRP discriminatory performance was observed (AUC = 0.84; DeLong test vs. total cohort, p = 0.004), subject to the optimistic bias inherent in derivation-cohort stratification. The AUC improvement was statistically significant (DeLong test, p = 0.004). The empirically derived optimal cutoff in the High-Utility group was 31.5 mg/L (Youden Index J = 0.54). Conclusions: These exploratory, post hoc findings are a first step into evaluation based on a pilot ML tool and require prospective multicenter validation before any conclusions regarding clinical utility can be drawn. The CCS represents a hypothesis-generating construct only and must not be used for clinical decision-making in its current form.