A machine learning model including pentraxin-3 as predictor of outcomes in community-acquired pneumonia

The clinical diagnosis, severity assessment, and outcome prognostication of community-acquired pneumonia (CAP) remain challenging due to the complex disease pathophysiology. Accurate outcome prediction is crucial for optimizing patient management, reducing mortality, and minimizing hospital and ICU admissions.

In this prospective observational cohort study, 228 CAP patients with varying degrees of disease severity were assessed. Clinical and demographic data, along with multiple biomarker measurements, including pentraxin-3 (PTX3), were analysed longitudinally. The primary outcome was clinical failure.

Among the single parameters evaluated, the oxygen saturation to fraction of inspired oxygen ratio (SpO₂/FiO₂), PTX3, and mid-regional pro-adrenomedullin (MRproADM) demonstrated the strongest predictive performance, with areas under the curve (AUC) of 0.799, 0.709, and 0.647, respectively. Machine learning (ML) experiments integrating multiple features identified the optimal algorithm for outcome prediction, combining these stand-alone markers at baseline and 72 h. The optimal ML model achieved an AUC of 0.950 (95% CI 0.83-0.96), recall of 92.6%, accuracy of 92.0%, and precision of 86.6%, representing a > 15% AUC improvement over any individual biomarker.

While SpO₂/FiO₂ remains the most reliable stand-alone prognostic marker, PTX3 demonstrated significant independent outcome predictive value. When integrated with other biomarkers using ML-based models, outcome prediction significantly improved, underscoring its potential for CAP patient management.

NCT06491004 (ClinicalTrials.gov).