In a retrospective study of 2,018 hospitalized patients with acute leukemia, routine admission data alone were enough to predict pneumonia risk with an interpretable XGBoost model that reached an AUC of 0.767 and accuracy of 0.734 in the validation cohort. The study found 14 key predictors, with low calcium, sodium, and platelets among the strongest signals, suggesting earlier risk stratification may be feasible at admission.

Why It Matters To Your Practice

• Pulmonary infection is a major cause of death in acute leukemia, and standard risk tools may miss nonlinear interactions among common clinical variables.

• This model used information clinicians already collect on admission, which could make implementation more practical than approaches requiring specialized testing.

• Earlier identification of high-risk patients could support faster diagnostic workup, closer monitoring, and more proactive antimicrobial or supportive care decisions.

Clinical Implications

• The best-performing model was XGBoost, outperforming six other machine learning approaches on discrimination, calibration, and decision curve analysis.

• The dataset was split into training (n=1,413) and validation (n=605) cohorts, and the validation performance was AUC 0.767 with accuracy 0.734.

• Top contributors to predicted pneumonia risk included decreased calcium, sodium, and platelets, plus higher age, temperature, hemoglobin, red blood cells, heart rate, MCHC, and white blood cells.

Insights

• The model was built from a retrospective cohort at Zhangzhou Affiliated Hospital of Fujian Medical University spanning January 2019 to August 2025.

• Feature selection used LASSO with cross-validation before seven machine learning models were trained and compared.

• Interpretability matters for adoption: SHAP analysis helped show which variables were driving model predictions rather than treating the output as a black box.

The Bottom Line

• For clinicians interested in how AI may affect practice, this study is a practical example of machine learning adding value to routine inpatient triage rather than replacing judgment.

• The immediate opportunity is decision support: flagging acute leukemia patients at higher pneumonia risk early enough to guide surveillance and preemptive management.

• Because this was a single-center retrospective study, external validation is still needed before broad clinical deployment.