Predicting survival of Hodgkin lymphoma using machine learning-an analysis based on the SEER database

Area of Science:

• Oncology
• Medical Informatics
• Biostatistics

Background:

• Hodgkin lymphoma (HL) prognosis prediction is crucial for effective clinical decision-making.
• Developing accurate prognostic models can significantly improve patient management strategies.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting Hodgkin lymphoma (HL) prognosis.
• To identify key prognostic factors influencing overall survival (OS) in HL patients.
• To assist clinicians in optimizing treatment decisions for HL.

Main Methods:

• Utilized the Surveillance, Epidemiology, and End Results (SEER) database (2000-2021) for patient data.
• Employed the Boruta algorithm for feature selection and developed four ML models.
• Evaluated model reliability using AUC, decision curve analysis, and Brier score; ranked feature importance with SHAP.

Main Results:

• The eXtreme Gradient Boosting (XGBoost) model demonstrated superior predictive performance.
• Key predictors of HL overall survival (OS) included age, Ann Arbor stage, B symptoms, marital status, and radiation.
• XGBoost, Coxph, and RSF models showed significantly higher net benefit for 1-year OS prediction compared to LightGBM.

Conclusions:

• The XGBoost model offers excellent predictive accuracy for Hodgkin lymphoma prognosis.
• This prognostic tool can empower clinicians to select more appropriate and personalized treatment strategies.
• Accurate prognosis prediction is vital for improving patient outcomes in Hodgkin lymphoma care.