Predicting recovery after urosepsis: a multicentre machine learning study using SOFA-based organ dysfunction measures and microbiology

Background: Urosepsis accounts for up to 40% of adult sepsis cases, yet prognostic tools have focused on mortality rather than recovery. Among survivors, recovery trajectories vary substantially and cannot be anticipated from composite severity scores alone. No validated framework exists for predicting recovery failure after the acute episode. Methods: Using prospectively collected data from 354 adults with confirmed urosepsis enrolled across 34 European centres in the SERPENS cohort, interpretable machine learning models were developed to predict failure to achieve full recovery at Day 3 and Day 30. Clinical, microbiological, and Sequential Organ Failure Assessment (SOFA) data were analysed using logistic regression and XGBoost models with total versus organ-specific SOFA representations. Results: Using baseline total SOFA, models achieved strong discrimination for short-term recovery prediction (AUROC 0.87) and good discrimination for long-term prediction (AUROC 0.79). Incorporating longitudinal SOFA measurements improved long-term prediction to AUROC 0.87. Models using organ-specific SOFA representations improved prediction at both horizons, achieving AUROCs of 0.90 (Day 3) and 0.83 (Day 30), with renal, CNS, and respiratory dysfunction emerging as dominant predictors. Microbiological features provided additional prognostic information, and performance remained broadly consistent during cross-centre validation. Conclusions: This study presents the first recovery-focused prognostic framework for confirmed urosepsis, demonstrating that routinely collected organ dysfunction and microbiological data can predict recovery trajectories beyond the acute phase.