The study drew on 4,966 adults admitted to 48 intensive care units across nine public tertiary hospitals in Central China between January 2023 and June 2025. Of these patients, 3,079 had influenza and 1,930 had COVID-19, and all required ICU stays of at least 48 hours. Patients with pre-existing pulmonary aspergillosis or missing infection-status documentation were excluded. Overall, 439 patients, or 8.8 percent, developed VAPA: 273 cases of influenza-associated pulmonary aspergillosis (8.9 percent of influenza patients) and 185 cases of COVID-19-associated pulmonary aspergillosis (9.6 percent of COVID-19 patients). Among the 43 patients co-infected with both viruses, 19 developed VAPA and were counted in both subgroups. A further 177 patients, 3.6 percent of the cohort, died from causes other than VAPA.

The methodological core of the work lies in how those deaths are handled statistically. Traditional Cox proportional hazards models treat patients who die from other causes as censored observations, implicitly assuming they remain at risk of developing aspergillosis in the future. In reality, death precludes the outcome entirely, so ignoring it as a competing event inflates the estimated absolute risk. The Fine-Gray subdistribution hazards framework addresses this by keeping deceased patients in the risk calculation as individuals who can no longer experience the event. The consequence was striking: in the highest-risk decile, the standard Cox model overestimated VAPA risk by 20 to 30 percent, while the Fine-Gray model overestimated it by only 1 to 2 percent. Both models slightly underestimated risk, by about 5 percent, in the lowest-risk decile.

Discrimination, the ability to separate patients who develop VAPA from those who do not, was statistically indistinguishable between the two models. In the derivation cohort, DeLong’s tests found no significant difference at either 7 days (D = -0.26, P = 0.798) or 14 days (D = -0.03, P = 0.974) after ICU admission. Calibration, however, clearly favored the competing-risk approach, and the gap widened with longer prediction horizons, because extended windows give patients more opportunity to die from competing causes. The authors argue this matters most precisely where clinical decisions are hardest: high predicted risk often crosses the threshold for starting empirical antifungal therapy, and a 20 to 30 percent overestimate in that stratum could expose patients who are not truly at high risk to drug toxicities, added costs, and the selective pressure that drives antifungal resistance.

To build the models, the team collected 54 candidate predictor variables at or within 24 hours of ICU admission from hospital and laboratory information systems, then used least absolute shrinkage and selection operator (LASSO) regularization to select predictors. Under the more stringent minimum-mean-squared-error-plus-one-standard-error criterion, six predictors survived: SOFA score, haematological malignancy, neutrophil-to-lymphocyte ratio (NLR), the radiological halo sign, use of priority antibiotic combinations, and cumulative corticosteroid dose administered before ICU admission. Restricted cubic splines with four knots were applied to the continuous variables, and likelihood-ratio tests confirmed significant non-linear relationships for SOFA score (LR = 164.93, P < 0.01), NLR (LR = 22.00, P = 0.009), and pre-ICU corticosteroid dose (LR = 364.56, P < 0.01). All three displayed a J-shaped association with VAPA risk in both models.

The effect sizes of the retained predictors were consistent across the two modeling frameworks. Haematological malignancy carried a subdistribution hazard ratio of 1.82 (95 percent CI 1.42-2.33), the halo sign 1.51 (95 percent CI 1.18-1.93), and priority antibiotic combinations, defined as concurrent use of two or more agents from a national priority list including carbapenems, glycopeptides, linezolid, tigecycline, polymyxins, and beta-lactam/beta-lactamase inhibitor combinations, a subdistribution hazard ratio of 1.90 (95 percent CI 1.55-2.32). The J-shaped corticosteroid finding is presented by the authors as the first such characterization for VAPA, and it carried a virus-specific twist: a significant interaction between viral subtype and pre-ICU corticosteroid dose (P = 0.006) showed a stronger association in COVID-19 patients, with a subdistribution hazard ratio of 1.17 per 10 milligram increase in cumulative dexamethasone-equivalent dose (95 percent CI 1.09-1.26).

Validation was temporal as well as internal. The cohort was split by calendar period, with patients admitted from January 2023 to June 2024 forming the derivation cohort for training and internal validation, and 1,948 patients admitted from July 2024 to June 2025 forming a prospective validation cohort. In that external cohort, the Fine-Gray model’s discrimination showed no statistically significant decline compared with the derivation cohort at either 7 days (D = 0.08, P = 0.937) or 14 days (D = 0.19, P = 0.853), and calibration remained excellent, with Brier scores at or below 0.05. A prespecified subgroup analysis confirmed that performance was consistent across viral etiologies, with no significant discrimination differences between influenza and COVID-19 patients at either time point (D = 1.01, P = 0.313 at 7 days; D = 1.22, P = 0.224 at 14 days) and good calibration in both subgroups.

The authors interpret the corticosteroid interaction as biologically informative without undermining predictive accuracy. Although the interaction term was statistically significant, overall model performance did not differ meaningfully between viral subtypes, a dissociation they highlight as a lesson for predictive modeling in heterogeneous populations: effect modification in an underlying risk mechanism does not necessarily compromise a multivariable model’s clinical applicability. They also point to NLR as pathophysiologically plausible, since an elevated ratio captures both hyperinflammation through neutrophilia and immunosuppression through lymphopenia, an immune dysregulation sometimes described as an immunological paradox that combines cytokine storm conditions with impaired cellular immunity, creating favorable ground for invasive fungal infection.

On treatment implications, the study’s observational data add to a contested debate about antifungal prophylaxis. No significant protective effect of prophylaxis was observed in the cohort. The authors note that influenza-associated aspergillosis often develops very early after ICU admission, leaving prophylaxis initiated on admission too late to help, and they cite their group’s recent target trial emulation finding that early prophylaxis did not reduce IAPA incidence and was associated with hepatotoxicity. COVID-19-associated aspergillosis typically occurs later in the ICU stay, but the authors argue that small retrospective studies suggesting benefit are limited by bias and have not consistently translated into survival benefit. Weighing the high cost, toxicity, and drug-drug interactions of first-line azoles such as voriconazole, they conclude that widespread prophylaxis for VAPA is currently unjustified and that risk-stratified targeting is the more defensible strategy.

The study has limits the authors state explicitly. The cohort comes from Central China, so generalizability to other regions, ethnicities, and healthcare systems requires independent multinational validation. The model covers influenza and COVID-19 only, leaving performance after pneumonia from other respiratory viruses such as respiratory syncytial virus or adenovirus unresolved. And because the investigation was non-interventional, it validated predictive performance rather than clinical utility; the authors propose randomized trials comparing model-guided antifungal initiation against routine care as the necessary next step. Within those boundaries, the work delivers a competing-risk prediction tool that matches the discrimination of conventional models while calibrating absolute risk far more accurately, deployed at the point of care through a publicly accessible web application designed for clinicians without statistical expertise.

Subject of Research: Development and validation of a competing-risk prediction model for pulmonary aspergillosis in critically ill patients with influenza or COVID-19 pneumonia

Article Title: A novel predictive tool for pulmonary aspergillosis in critically ill patients with viral pneumonia: a multi-center, ambispective cohort study

Article References: Zhang, J., Xu, S., Guo, L., Qiao, B., Li, Y., Zhou, Y., Hu, X., Guo, Y., Hu, B., Zheng, S., Chen, C., Li, Y., Liu, J., Zhang, J., & Li, P. (2026). A novel predictive tool for pulmonary aspergillosis in critically ill patients with viral pneumonia: a multi-center, ambispective cohort study. Frontiers in Cellular and Infection Microbiology, 16, 1956822. https://doi.org/10.3389/fcimb.2026.1956822

Image Credits: AI Generated

DOI: 10.3389/fcimb.2026.1956822

Keywords: pulmonary aspergillosis, viral pneumonia, competing risk, Fine-Gray model, ICU, influenza, COVID-19, corticosteroids, prediction model, LASSO, neutrophil-to-lymphocyte ratio, antifungal prophylaxis