Development and temporal validation of an explainable machine-learning model for prediagnostic risk stratification of secondary pulmonary fungal infection in severe pneumonia

Secondary pulmonary fungal infection (SPFI) is difficult to recognise early in patients with severe pneumonia, because persistent inflammation, progressive respiratory failure, and radiographic deterioration often overlap with bacterial progression, ventilator-associated complications, and non-infectious lung injury. We aimed to develop and temporally validate an explainable model for prediagnostic SPFI risk stratification using clinical information available before diagnostic confirmation. We performed a retrospective cohort study including a development cohort of 659 patients with severe pneumonia and a temporally independent validation cohort of 150 consecutive patients admitted between June and December 2025. Twenty-three routinely available prediagnostic predictors were considered. Recursive feature elimination with cross-validation (RFECV) and the one-standard-error rule were used for feature selection within the training data. Nine machine-learning algorithms were trained and compared using the same selected predictors. The model was developed for prediagnostic risk stratification before diagnostic confirmation rather than fixed-horizon prediction of incident SPFI. The final one-standard-error predictor set included 14 variables: age, body mass index, type 2 diabetes, glucocorticoid use, exposure to three or more antibiotic classes, prolonged antibiotic use, surgery before prediction, white blood cell count on admission, neutrophil percentage, PaO2, PaCO2, blood glucose, PaO₂/FiO₂ ratio, and C-reactive protein. The Support Vector Classifier was selected as the final model. In temporal validation, the model achieved an AUROC of 0.766, AUPRC of 0.659, sensitivity of 0.780, specificity of 0.600, and Brier score of 0.186. The calibration intercept was − 0.405 and the calibration slope was 0.630. SHAP and permutation analyses showed clinically plausible contributions from antibiotic exposure, age, surgery before prediction, prolonged antibiotic use, type 2 diabetes, and glucocorticoid use. An explainable model showed moderate and clinically plausible performance for SPFI risk stratification in severe pneumonia. The model should be regarded as an adjunct for structured clinical reassessment and targeted fungal investigation rather than as a fixed-horizon prognostic, standalone diagnostic, or antifungal-treatment tool. Relevant temporal miscalibration indicates that recalibration and multicentre evaluation are required before implementation.