Machine learning-based early prediction of severe pneumonia risk in hospitalized leukemia patients: a dual-center retrospective cohort study.

Abstract

Background: Severe pneumonia is a major infectious complication in patients with leukemia and may rapidly progress to respiratory failure or septic shock. Disease- and treatment-related immunosuppression increases susceptibility, yet population-specific early risk-stratification tools are limited.

Objectives: To develop and externally validate a machine learning model for predicting severe pneumonia in hospitalized patients with leukemia.

Methods: Data from 2,978 leukemia patients at two Chinese hospitals were retrospectively analyzed. The derivation cohort was divided into training and internal test sets; SMOTE was applied only to the training set. LASSO and SVM-RFE were used for predictive feature selection, eight machine learning algorithms were compared, and RCS was used for exploratory association analysis. The final model was presented as a nomogram.

Results: Seven predictors were retained: age, history of myelosuppression, acute lymphoblastic leukemia (ALL) subtype, cardiac insufficiency, atrial fibrillation, red blood cell distribution width-standard deviation (RDW-SD), and prothrombin time activity. Logistic regression showed the best generalizability, with an area under the curve (AUC) of 0.801 in external validation. Sensitivity analyses showed similar discrimination with multiple imputation and complete-case analysis.

Conclusion: We developed and externally validated an interpretable model for severe pneumonia in hospitalized patients with leukemia. Because direct programmed-cell-death biomarkers were not measured and SMOTE substantially altered training prevalence, mechanistic interpretation and absolute-risk calibration require caution. Prospective recalibration and clinical-impact evaluation are needed before routine use.