Development and validation of an interpretable machine learning model for the early diagnosis of Pneumocystis jirovecii pneumonia in immunosuppressed patients

Background Pneumocystis jirovecii pneumonia (PJP) is a common, potentially fatal infection in immunosuppressed patients. Rapid and accurate diagnosis of PJP remains challenging, as delays increase mortality. This study aimed to develop a machine-learning model using non-invasive, easily accessible data to quickly assess PJP risk in these patients. Methods A multicenter retrospective study was conducted among immunosuppressed patients with pneumonia who underwent bronchoalveolar lavage evaluation between January 1, 2018, and December 31, 2021. Clinical baseline data from patients diagnosed with PJP or non-PJP across 10 hospitals were included. After data preprocessing, patients were divided into training and validation cohorts according to the admission time. After feature selection, a model was developed. The model’s performance was assessed using the imputed datasets from the validation cohort. The optimized model was interpreted using SHapley Additive exPlanations. Results The 359 enrolled patients included 189 with PJP and 170 with non-PJP pneumonia. Of these, 271 were assigned to the training set for model development, while 88 formed the validation set for performance assessment, respectively. Based on the ranking of feature importance, 15 features were selected for model development. The CatBoost model, selected after tuning, achieved an accuracy of 81.8% (95% CI: 73.5%–90.1%), sensitivity of 82.9% (95% CI, 70.0–95.9%), and specificity of 81.1% (95% CI: 70.2%–91.9%). Conclusions We constructed a high-performing diagnostic model for the rapid, early diagnosis of PJP in immunosuppressed patients using readily available clinical data within 24 hours of hospital admission. This tool may assist clinicians in evaluating PJP probability in immunocompromised pneumonia patients.