Development and validation of an interpretable machine learning model for predicting ICU prognosis in critically ill patients with invasive pulmonary aspergillosis: a multicenter study

Abstract

Invasive pulmonary aspergillosis (IPA) is associated with elevated mortality among intensive care unit (ICU) patients. Early prognostic risk stratification is of great importance; however, validated clinical tools remain limited. This study aimed to develop and validate an interpretable machine-learning (ML) model to predict ICU mortality in critically ill patients with IPA. This multicenter study included a retrospective derivation cohort ( n = 267, 2015–2025) and an independent prospective external validation cohort ( n = 129, 2022–2025) of critically ill patients with IPA diagnosed per modified AspICU (BM-AsperICU) algorithm. Using baseline clinical features acquired within 24 h of initial IPA diagnosis, we evaluated 11 individual ML algorithms and 9 ensemble models to predict ICU mortality. Performance was assessed via area under the receiver operating characteristic curve (AUROC) with 95% confidence intervals (CIs), calibration, and decision curve analysis (DCA). SHapley Additive exPlanations (SHAP) provided interpretability. Optimal cut-off-based risk stratification and subgroup validation were further conducted. Overall ICU mortality reached 59.1% and 53.5% in training and external validation cohorts, respectively. Eleven key features—age, interstitial lung disease (ILD), cumulative glucocorticoid dose (prednisone-equivalent, 7 days prior to IPA diagnosis), neutrophil count, platelet count, blood urea nitrogen, total bilirubin, D-dimer, PaO 2 /FiO 2 ratio, lactate, and serum galactomannan-to-lymphocyte ratio—were selected. The logistic regression (LR) model yielded the best overall performance, achieving an internal AUROC of 0.740 (95% CI 0.626–0.840) and an external AUROC of 0.798 (95% CI 0.714–0.868) and was well-calibrated. Using an optimal cutoff of 0.54, the model identified effective prognostic stratification; high-risk patients exhibited a significantly higher ICU mortality risk, independent of age and baseline SOFA score (adjusted hazard ratio [aHR] 2.00, 95% CI: 1.07–3.70, P = 0.030). Based on real-world multicenter data, our interpretable LR model and proof-of-concept web calculator enable early prognostic risk stratification upon IPA diagnosis. Further large prospective studies are required to evaluate implementation and clinical utility.