Diagnostic evaluation of Pentraxin-3 and its complementary role to Procalcitonin for respiratory sepsis source identification using explainable machine learning

Sepsis management in intensive care units is often hindered by the inability to rapidly localise the primary infection source. While systemic biomarkers such as Procalcitonin (PCT) are widely utilised, their lack of site-specificity remains a significant clinical bottleneck. This study presents a secondary analysis of a prospectively collected, single-centre cohort of 555 patients and investigates the complementary diagnostic value of Pentraxin-3 (PTX3), a locally produced acute-phase reactant, in identifying respiratory-sourced sepsis. We developed an XGBoost-based machine learning framework to compare the diagnostic performance of PTX3, PCT, and C-Reactive Protein (CRP). Individual biomarkers offer limited discriminative power (AUC ≈ 0.62 ), whereas a multi-feature machine learning approach achieves an AUC of 0.865 (95% CI 0.812−0.918; mean cross-validated AUC $$= 0.851$$ , SD $$= 0.031$$ ). We further employ Explainable Machine Learning techniques specifically SHAP and LIME to provide clinical transparency, demonstrating that PTX3, when combined with soluble CD163 (sCD163) and selected vital signs, provides complementary discriminatory value for pulmonary infection source identification. Sensitivity analyses restricted to biomarker-only features (AUC $$= 0.791$$ ) confirm that this value is not entirely dependent on target-adjacent clinical variables. These findings suggest that a machine-learning-augmented biomarker panel has the potential to support antibiotic stewardship and precision medicine in septic cohorts, pending prospective multicentre validation.