Development and validation of an interpretable machine learning model for the early diagnosis of invasive pulmonary aspergillosis in patients with severe fever with thrombocytopenia syndrome: a retrospective cohort study.

OBJECTIVES

Invasive pulmonary aspergillosis (IPA) is a common complication in patients with severe fever with thrombocytopenia syndrome (SFTS); however, its diagnosis remains challenging due to non-specific clinical manifestations and limited diagnostic tools. This study aimed to develop and validate a machine learning (ML)-based early diagnostic model to identify IPA in SFTS patients.

METHODS

A total of 374 SFTS patients with suspected IPA were retrospectively enrolled at Qishan Hospital between March 2022 and June 2025. Both the Least Absolute Shrinkage and Selection Operator (LASSO) and the Boruta algorithms were employed for feature selection. Seven ML algorithms were subsequently developed and evaluated using multiple evaluation metrics to compare their predictive performance. Optimal cutoff values were determined by the Youden index. The AUCs of the models were compared using the DeLong test. Feature contributions in the optimal model were interpreted using SHapley Additive exPlanations (SHAP).

RESULTS

Among the seven algorithms, the Random Forest (RF) model demonstrated the best predictive performance. The final model incorporated five key features: sex, serum potassium, international normalized ratio (INR), galactomannan (GM) test positivity, and SFTSV RNA load. The RF model achieved an area under the receiver operating characteristic curve (AUC) of 0.871 (95% CI: 0.799-0.933), with a specificity of 0.829 (95% CI: 0.735-0.912), a sensitivity of 0.854 (95% CI: 0.738-0.953), and an F1 score of 0.795. Calibration was acceptable, with a Brier score of 0.140 (95% CI: 0.105 - 0.179), and clinical utility was confirmed through decision curve analysis (DCA).

CONCLUSIONS

Our ML model provides a non-invasive and cost-effective approach for the early identification of SFTS-associated IPA. It may serve as a valuable adjunct to existing diagnostic strategies and support timely diagnostic evaluation in high-risk patients. Further external validation is warranted before clinical implementation.