Objective To analyze the pathogenic etiology of pulmonary infection after solid organ transplantation and construct a prognostic prediction model based on metagenomic next-generation sequencing (mNGS) technology, systematically identifying key predictors to provide evidence for clinical risk stratification and individualized interventions. Methods Clinical data were retrospectively collected from patients who developed pulmonary infection after liver or kidney transplantation at a single hospital between January 2020 and December 2023. All patients underwent mNGS detection of bronchoalveolar lavage fluid or sputum for pathogen identification. Collected data included demographic characteristics, transplant-related parameters, underlying diseases, laboratory test results, mNGS pathogen detection outcomes, and prognostic indicators. The dataset was randomly divided into a training set (n=262) and a test set (n=66). Within an AutoML framework, model hyperparameters were optimized using the Improved Dharma Optimization Algorithm (IDRA). Feature importance was validated bidimensionally via LASSO regression and SHAP interpretable models, with an interactive MATLAB-based decision support system developed. Results The overall positive detection rate of pathogens by mNGS significantly exceeded that of conventional methods (84.76% vs. 61.89%, P0.001). No statistically significant differences existed in baseline characteristics or laboratory indicators between the training and test sets (all P0.05), confirming randomized stratified sampling validity. Both cohorts showed highly consistent proportions of poor prognosis events (training set: 27.48% vs. test set: 28.79%, χ 2 =0.045, P = 0.832). The prediction model achieved a ROC-AUC of 0.9694 and PR-AUC of 0.9690 in the training set, and ROC-AUC of 0.9206 (95% CI: 0.854-0.987) with PR-AUC of 0.9273 (95% CI: 0.867-0.988) in the test set, outperforming comparative models. Fourteen key variables were ultimately selected: mNGS bacterial detection, mNGS fungal detection, procalcitonin (PCT), C-reactive protein (CRP), mNGS viral detection, white blood cell count, creatinine, post-transplantation time, neutrophil percentage, diabetes, age, total bilirubin, alanine aminotransferase (ALT), and lymphocyte percentage. The feature overlap rate with AutoML-screened variables was 78.6% (11/14). SHAP analysis revealed descending importance ranking: mNGS bacterial detection, mNGS fungal detection, PCT, etc. Conclusion Integrating multidimensional clinical data with explainable machine learning techniques, this study confirms the central role of pathogenic etiology characteristics in prognostic prediction for post-transplant pulmonary infection and demonstrates the potential for real-time risk assessment to inform clinical decisions. However, prospective validation across diverse care settings is required to establish its efficacy as an interventional guide. This work offers innovative tools and methodological frameworks to advance precision diagnosis and management, subject to ongoing refinement through multicenter collaboration.
Jiang et al. (Thu,) studied this question.