A Unified Machine Learning Model for Relapse Prediction in Clinical Stage I Testicular Cancer

Background Approximately one‐fourth of patients with clinical stage I testicular cancer relapse. For decades, risk stratification has been based on different tumor characteristics for seminomas and non‐seminomas. Previous studies primarily used Cox proportional‐hazards models and included only a limited number of variables. Machine learning techniques can integrate large datasets and may uncover novel combinations of risk factors. Objectives To develop and validate a unified machine learning‐based relapse prediction model for clinical stage I testicular cancer, regardless of histologic subtype, using nationwide histopathological and clinical data. Materials and Methods A population‐based cohort study of 1377 patients diagnosed with clinical stage I testicular cancer in Denmark from 2013 to 2018. Histopathological and clinical data were obtained through centralized pathology assessment and systematic medical record review. Two tree‐based binary classifiers (CatBoost and LightGBM) were trained to predict relapse, and a random survival forest model was used to estimate time‐to‐relapse. Data were split into training (80%, 5‐fold cross‐validation) and a test set (20%), balanced by seminoma/non‐seminoma subtypes and outcome. Subgroup analyses were performed for seminoma and non‐seminoma. Binary models were evaluated using receiver operating characteristic area under the curve, precision‐recall area under the curve, and Matthew's correlation coefficient; random survival forest performance was assessed using concordance index and Integrated Brier Score. Results CatBoost outperformed LightGBM (receiver operating characteristic area under the curve = 0.74) and demonstrated a high negative predictive value (0.86). The random survival forest achieved a concordance index of 0.71. Predictive performance was stronger in the non‐seminoma than in the seminoma subgroups. Top‐ranked predictive features included lymphovascular invasion, embryonal carcinoma, tumor necrosis, rete testis invasion, tumor size, and elevated lactate dehydrogenase and β‐human chorionic gonadotropin. Tumor necrosis and the anatomical location of lymphovascular invasion emerged as novel predictors. Discussion and Conclusion A unified machine learning‐based model for relapse prediction in clinical stage I testicular cancer is feasible and demonstrates moderate predictive accuracy. It is particularly useful for ruling out relapse and shows greater robustness in non‐seminoma. These findings provide a framework for validation in independent cohorts and highlight key predictive features for future research.