Construction of a machine learning based prognostic model for Pneumocystis pneumonia in HIV-negative patients

Objective Pneumocystis pneumonia (PCP) is increasingly prevalent among HIV-negative patients, and its mortality rate is high in severe cases. The aim of this study is to analyze the clinical characteristics and prognostic factors of severe PCP in HIV-negative patients. Methods An innovative dual-center design with a relatively large sample size was adopted. A total of 82 HIV-negative patients between January 2021 and October 2023 at the Affiliated Hospital of Xuzhou Medical University and Suining County People's Hospital were included, and they were diagnosed with severe PCP infection by metagenomic next-generation sequencing (mNGS). Mortality-associated risk factors were identified using univariate/multivariate analyses and machine learning algorithms. Results Eighty-two severe PCP patients were enrolled, with a mean age of 64 ± 17 years; 67.1% were male and 32.9% had a history of smoking. The most common comorbidities were cardiovascular disease (40.2%), malignancy (34.1%), and interstitial lung disease (ILD) (28.0%). Some patients had received immunosuppressive therapies, including high-dose corticosteroids (46.3%), chemotherapy (21.9%), or immunosuppressants (32.9%). Clinically, 89.0% presented with dyspnea, 82.9% with cough, and 73.2% with fever. Oxygen therapy was required in 96.3% of patients, with 43.9% receiving high-flow oxygen and 52.4% requiring mechanical ventilation. Imaging results revealed bilateral lung involvement in 97.6% and ground-glass opacities in 97.6% of cases. Furthermore, common complications included septic shock (13.4%) and hepatic/renal dysfunction (12.2%). Multiple pathogens were detected in severe PCP patients, and 59.8% with Epstein-Barr virus and 45.1% with cytomegalovirus. When stratified by outcome, the mortality group showed elevated neutrophil counts (NEU), reduced lymphocytes (LYM) and platelets (PLT), reduced oxygenation index, and increased rates of septic shock and invasive mechanical ventilation. The overall fatality rate was 41.5%, which is consistent with previous literature. Patients with ILD had higher mortality. Among the machine learning models, Random Forest (RF) and XGBoost achieved the best fitting effect on the test set. After decision curve analysis (DCA) and calibration curve evaluation, RF demonstrated the best overall performance. Conclusions Elevated NEU, decreased LYM and PLT, reduced oxygenation index, septic shock, and invasive mechanical ventilation are risk factors for mortality in severe PCP among HIV-negative patients. Patients with ILD have poorer outcomes. The RF model showed the best overall prognostic performance. This study, by integrating mNGS technology and machine learning algorithms, provides new insights and tools for early diagnosis, precision therapy, and individualized management of severe PCP in HIV-negative patients, with potential to reduce mortality and improve prognosis.