A machine learning-based risk prediction tool for early identification of invasive pulmonary aspergillosis in immunocompetent patients: a systematic review and meta-analysis

Background: Invasive pulmonary aspergillosis (IPA) increasingly affects non-neutropenic patients, posing diagnostic challenges due to nonspecific clinical features and limited sensitivity of conventional tests. Early identification is critical for improving outcomes. This study aimed to develop and validate a machine learning-based tool for early risk stratification of IPA in this population. Methods: A systematic meta-analysis identified independent IPA risk factors in non-neutropenic patients. A retrospective cohort of non-neutropenic patients was analyzed. Least absolute shrinkage and selection operator (LASSO) regression selected predictive features, and multiple machine learning (ML) algorithms were evaluated. The optimal model (random forest) was externally validated on an independent cohort. SHapley Additive exPlanations (SHAP) values interpreted feature importance. Results: A retrospective cohort of 524 hospitalized non-neutropenic patients (422 training, 102 validation) adhering to established diagnostic criteria was analyzed. Meta-analysis confirmed diabetes [odds ratio (OR) =1.43], respiratory disease (OR =1.76), corticosteroid exposure (OR =1.48), and smoking history (OR =1.64) as key risk factors. The random forest model incorporated nine predictors (including antibiotic use, viral pneumonia, intensive care unit (ICU) admission, low protein levels, and bacterial infections) and achieved high accuracy [area under the curve (AUC) =0.950, sensitivity =0.857, specificity =0.905]. External validation maintained robust discrimination (AUC =0.856; sensitivity and specificity =0.733). SHAP analysis revealed critical synergistic interactions, particularly between prolonged antibiotic use and viral pneumonia in elevating IPA risk. Conclusions: This validated ML tool integrates meta-evidence with clinical data for early, accurate IPA risk stratification in immunocompetent patients. It effectively captures nonlinear predictor interactions and demonstrates strong generalizability, supporting clinical utility.