Machine Learning for Prediction of High-Risk Hospitalizations in Lymphoma Patients: A Danish Population-Based Study

Abstract

Objective: Infections are a leading cause of hospitalization in patients treated for lymphoma and can be life-threatening. This study developed a machine learning (ML)-based risk stratification method to classify infection-related hospitalizations (IRH) into serious-IRH (S-IRH) and non-serious-IRH (NS-IRH). Methods: S-IRH was defined based on death, positive blood culture, blood stream infection, and sepsis during hospitalization. Clinical data from health records and registries were used to construct the feature matrix for an XGBoost model. The study included 727 adult lymphoma patients diagnosed 2013–2023 and treated with first-line therapies including CHOP (or CHOP-like), ABVD, BEACOPP, CVP, and Bendamustine. Results: A total of 591 IRHs were identified, of which 119 were categorized as S-IRH. The developed model achieved a ROC-AUC of 71.0% for predicting S-IRH. At a prediction threshold of 0.3, the average sensitivity and specificity were 63.0% and 63.0%, respectively. The negative and positive predictive values were 87.5% and 30.3%, respectively. Conclusion: The developed model demonstrated acceptable clinical performance in predicting S-IRH. However, despite the wealth of data points entered in the model, performance was not sufficient for a stand-alone decision tool, but it should rather be seen as a decision support tool for clinicians.