Machine Learning Model Predicts Invasive Pulmonary Aspergillosis Risk in ICU Patients

Invasive pulmonary aspergillosis, a devastating fungal infection once considered almost exclusively a threat to patients with depleted neutrophil counts, is quietly becoming a major problem in intensive care units worldwide. Now, a team of researchers in China has harnessed the power of machine learning to predict which non-neutropenic ICU patients are most at risk of developing this deadly disease, and their findings challenge long-held assumptions about the role of corticosteroids in fungal infection risk.

The study, published in Frontiers in Cellular and Infection Microbiology, enrolled 837 non-neutropenic adults hospitalized with pulmonary infections across three tertiary hospitals in China between December 2020 and December 2025. The researchers developed a Random Forest classifier that achieved an impressive external validation area under the curve (AUC) of 0.865, demonstrating strong discriminatory power for identifying patients at risk of invasive pulmonary aspergillosis (IPA). But the real surprise came when the team examined which factors the algorithm deemed most important.

Using SHAP (SHapley Additive exPlanations) analysis, a technique that reveals how much each variable contributes to a model’s predictions, the researchers found that six predictors consistently dominated: isolation of Pseudomonas aeruginosa, hypoxemia defined as a PaO2/FiO2 ratio below 200 mmHg, diabetes mellitus, organ failure indicated by a Sequential Organ Failure Assessment score above 5, lymphopenia with lymphocyte counts below 0.8 × 10⁹/L, and ICU admission. Strikingly, corticosteroid exposure—a factor long feared as a major driver of invasive aspergillosis—did not rank among the top six predictors.

This finding alone would be noteworthy, but the story took an even more intriguing turn when the researchers examined how the association between corticosteroids and IPA varied by bacterial context. In patients carrying Pseudomonas aeruginosa, corticosteroid exposure was associated with an odds ratio below 1 (OR 0.52, 95% CI 0.214–1.283), suggesting a potentially protective effect. In contrast, among patients infected with other Gram-negative bacilli, the same exposure pointed toward an odds ratio above 1 (OR 2.17, 95% CI 0.722–6.535), indicating increased risk. Although the formal test for interaction did not reach statistical significance (P for interaction = 0.131), this directional pattern persisted after propensity score matching, suggesting it is not merely a statistical artifact.

The biological explanation for this divergence may lie in the complex, two-way antagonism between Pseudomonas aeruginosa and Aspergillus fumigatus. These two pathogens are not passive co-occupants of the airways; instead, they engage in an intricate molecular battle. Pseudomonas aeruginosa secretes redox-active phenazines, including pyocyanin, which suppress A. fumigatus growth by inducing oxidative stress and damaging mitochondrial ultrastructure. The bacterium also produces the filamentous bacteriophage Pf4, which binds to fungal hyphae and sequesters environmental iron, depriving the fungus of this essential nutrient. Additionally, the Pseudomonas siderophore pyoverdine has been identified as a central factor in inhibiting A. fumigatus biofilm formation through iron chelation.

The researchers speculate that corticosteroids may act directly on Pseudomonas aeruginosa, dampening the virulence programs that normally clear the path for Aspergillus. Recent work has shown that budesonide and fluticasone propionate can reshape Pseudomonas aeruginosa, reducing its invasiveness into host cells by 25–40% and cutting IL-8 release by 15–50%. However, the same drugs also nudged the organism toward biofilm growth, with a 1.6-fold increase at peak concentration after sufficient exposure time. This bacterial reshaping could alter the fungal risk landscape in ways that are only beginning to be understood.

In patients carrying other Gram-negative bacilli that lack such a potent anti-Aspergillus repertoire, this protective mechanism is absent. The corticosteroid effect in these patients may instead reflect the drug’s direct suppression of host immunity, with no bacterial counterbalance to offset the increased fungal vulnerability. This interpretation, while still speculative and awaiting direct experimental validation, offers a coherent framework that bridges the clinical observation with a large body of mechanistic evidence on interspecies antagonism in the airway.

Beyond the corticosteroid story, the researchers derived a simplified six-variable bedside score from the top-ranking predictors, assigning integer weights proportional to their mean absolute SHAP values: Pseudomonas aeruginosa isolation (3 points), PaO2/FiO2 below 200 mmHg (2 points), diabetes mellitus (1 point), SOFA above 5 (1 point), lymphocyte count below 0.8 × 10⁹/L (1 point), and ICU admission (1 point). The score ranges from 0 to 9, with a high-risk cutoff set at 4 or more points. This software-free tool achieved AUCs of 0.869 in training, 0.774 in internal validation, and 0.719 in external validation, offering a practical, pathogen-informed approach to individualized risk stratification at the bedside.

Survival analysis further reinforced the importance of bacterial context. In the discovery cohort, patients with non-pseudomonal Gram-negative infections had the highest mortality, with their survival probability separating from the Pseudomonas aeruginosa and non-Gram-negative groups by days 7–10 (log-rank P < 0.001). The external validation cohort reproduced this pattern, with the non-pseudomonal Gram-negative subgroup again showing the lowest survival. Notably, Pseudomonas aeruginosa isolation did not translate into extra short-term mortality, likely because it typically represents chronic colonization of damaged airways rather than an acute pathogenic insult.

The implications of this research extend far beyond the ICU. Current diagnostic frameworks for invasive pulmonary aspergillosis, including the 2025 German national guideline, lay out host factors, clinical features, and mycologic evidence as the diagnostic pillars—but bacterial co-pathogens are nowhere in that framework. This omission looks increasingly hard to defend in light of these findings. The researchers argue that future guidelines should incorporate bacterial context, and that working out which bacterial species turns the corticosteroid effect toward increased versus decreased risk would be a first step toward microbiome-informed prevention. As the population of non-neutropenic patients with invasive aspergillosis continues to grow, tools like this externally validated six-variable score may prove invaluable in helping clinicians navigate the complex interplay between bacteria, fungi, and the immune system in critically ill patients.

Subject of Research: Machine learning-based risk prediction for invasive pulmonary aspergillosis in non-neutropenic intensive care unit patients, with an exploratory analysis of pathogen-related risk factors

Article Title: Machine learning-based risk prediction for invasive pulmonary aspergillosis in non-neutropenic intensive care unit patients, with an exploratory analysis of pathogen-related risk factors

Article References: Bai, J., Feng, T., Liu, B., Wang, J., Wang, Z., Liu, F., & Zhu, J. (2026). Machine learning-based risk prediction for invasive pulmonary aspergillosis in non-neutropenic intensive care unit patients, with an exploratory analysis of pathogen-related risk factors. Frontiers in Cellular and Infection Microbiology, 16, Article 1967577. https://doi.org/10.3389/fcimb.2026.1967577

Image Credits: AI Generated

DOI: 10.3389/fcimb.2026.1967577

Keywords: Machine, learning-based, risk, prediction, invasive, pulmonary, aspergillosis, non-neutropenic, intensive, care, unit, patients