Synergistic impact of Gram-negative bacterial co-infection on invasive pulmonary aspergillosis in non-neutropenic patients: a machine learning-based risk stratification approach

BACKGROUND: Invasive pulmonary aspergillosis (IPA) is increasingly recognized in non-neutropenic patients, where coexisting bacterial infections, particularly with Gram-negative pathogens, may impair susceptibility. However, validated tools for early risk stratification in this population remain unavailable. METHODS: We retrospectively analyzed 437 non-neutropenic adults with bacterial co-infection (derivation N.=331; validation N.=106) admitted between 2019 and 2024. Independent predictors of IPA were identified through multivariable logistic regression and incorporated into both a weighted clinical risk score and an ensemble machine learning (ML) model. Model performance was assessed using discrimination, calibration, and decision curve analysis, with subgroup validation in Gram-negative infection, intensive care unit (ICU) admission, and diabetes. RESULTS: Seven independent predictors of IPA were identified: nodular shadow, chronic respiratory disease, Gram-negative infection, corticosteroid exposure, ICU admission, smoking history, and diabetes. Gram-negative pathogens accounted for nearly half of infections, with Pseudomonas aeruginosa predominating. The ensemble score achieved a strong performance (area under the curve [AUC] 0.922 derivation; 0.862 validation) with superior calibration compared to traditional approaches. Risk stratification at a threshold score of ≥7.5 significantly enriched 28-day IPA incidence (log-rank P