Interpretable machine learning model to predict 90-day radiographically confirmed pneumonia after chemotherapy initiation in non-Hodgkin lymphoma: development and internal validation of a single-center cohort.

Abstract

Background: Radiographically confirmed pneumonia within 90 days of chemotherapy initiation is a frequent and clinically important complication in patients with non-Hodgkin lymphoma, yet interpretable tools for early individualized risk estimation are limited.

Objective: To develop and internally validate an interpretable machine-learning model that predicts the 90-day risk of radiographically confirmed pneumonia after chemotherapy initiation in non-Hodgkin lymphoma.

Methods: We retrospectively analyzed 205 chemotherapy-treated NHL patients. A two-step feature selection (LASSO followed by random-forest-based recursive feature elimination) identified four predictors: high-grade malignancy, drinking (alcohol use), estimated glomerular filtration rate (eGFR), and smoking. Five algorithms were trained and compared under a stratified 70/30 split (training n = 145; internal hold-out test set n = 60) with leakage-safe preprocessing (within-fold kNN imputation, SMOTE, and scaling). The gradient boosting machine (GBM) performed best and was interpreted using SHAP. A web-based prototype was implemented for research use only.

Results: On the internal hold-out test set (n = 60), the GBM achieved an AUC of 0.855 (95% CI 0.746-0.964), an F1 score of 0.679, and a Brier score of 0.155. SHAP identified reduced eGFR, smoking, drinking, and high-grade malignancy as influential contributors; case-level waterfall and force plots enhanced transparency. These estimates reflect internal validation only and were obtained without systematic microbiological confirmation or standardized radiologic rescoring. Accordingly, performance may be optimistic, and real-world use is not advised pending temporal and multicenter external validation (with potential recalibration) and prospective evaluation.

Conclusion: The interpretable GBM model demonstrated promising discrimination and calibration on an internal hold-out test set; however, clinical deployment requires temporal and multicenter external validation (as well as prospective assessment with potential recalibration). The accompanying web calculator is a research-only prototype and is not intended for clinical decision-making until such validation is completed.