Construction and evaluation of a machine-learning-based prediction model for pneumonia in patients with acute leukemia.
TL;DR
An interpretable XGBoost model that accurately predicts pneumonia risk in AL patients based on routine admission data is developed and validated and provides actionable risk stratification to inform preemptive diagnostic and therapeutic strategies.
Abstract
Background
Pulmonary infection is a major cause of mortality in individuals diagnosed with acute leukemia (AL). However, conventional risk assessment tools often fail to adequately capture the complex, nonlinear interactions among clinical factors that contribute to pneumonia susceptibility. While machine learning (ML) offers potential advantages for such tasks, comparative evaluations of multiple ML algorithms for early pneumonia prediction in AL populations remain limited. This study sought to develop and rigorously validate an ML-based predictive model using readily accessible admission variables to enable early identification of high-risk patients and support proactive clinical management.
Method
We conducted a retrospective analysis of 2018 AL patients hospitalized at Zhangzhou Affiliated Hospital of Fujian Medical University between January 2019 and August 2025. Demographic information, comorbidities, vital signs, and laboratory parameters were retrieved from electronic medical records. Least absolute shrinkage and selection operator (LASSO) regression with cross-validation was applied to the training cohort for feature selection. Subsequently, seven ML algorithms were employed to construct predictive models. Model performance was assessed based on discrimination (area under the receiver operating characteristic curve [AUC]), calibration (Brier score and calibration plots), and clinical utility (decision curve analysis [DCA]). The optimal model was further interpreted using Shapley Additive Explanations (SHAP).
Results
The cohort was randomly split into training (n = 1413) and validation (n = 605) sets at a 7:3 ratio. LASSO regression identified 14 key predictors from the initial variables. Among the seven evaluated models, the extreme gradient boosting (XGBoost) algorithm demonstrated superior predictive performance, achieving the highest AUC (0.767) and accuracy (0.734) in the validation cohort. Calibration plots and DCA consistently indicated the superiority of XGBoost over comparator models. SHAP analysis revealed the top ten contributors to pneumonia risk: decreased levels of calcium, sodium, and platelets, alongside elevated age, body temperature, hemoglobin, red blood cells, heart rate, mean corpuscular hemoglobin concentration (MCHC), and white blood cells.
Conclusion
We developed and validated an interpretable XGBoost model that accurately predicts pneumonia risk in AL patients based on routine admission data. This tool provides actionable risk stratification to inform preemptive diagnostic and therapeutic strategies.