Survival Prediction Using Transformer-Based Categorical Feature Representation in the Treatment of Diffuse Large B-Cell Lymphoma

Abstract

Diffuse large B-cell lymphoma (DLBCL) is a common and aggressive subtype of lymphoma, and accurate survival prediction is crucial for treatment decisions. This study aims to develop a robust survival prediction strategy to integrate various risk factors effectively, including clinical risk factors and Deauville scores in positron-emission tomography/computed tomography at different treatment stages using a deep-learning-based approach. We conduct a multi-institutional study on 604 DLBCL patients’ clinical data and validate the model on 220 patients from an independent institution. We propose a survival prediction model using transformer architecture and a categorical-feature-embedding technique that can handle high-dimensional and categorical data. Comparison with deep-learning survival models such as DeepSurv, CoxTime, and CoxCC based on the concordance index (C-index) and the mean absolute error (MAE) demonstrates that the categorical features obtained using transformers improved the MAE and the C-index. The proposed model outperforms the best-performing existing method by approximately 185 days in terms of the MAE for survival time estimation on the testing set. Using the Deauville score obtained during treatment resulted in a 0.02 improvement in the C-index and a 53.71-day improvement in the MAE, highlighting its prognostic importance. Our deep-learning model could improve survival prediction accuracy and treatment personalization for DLBCL patients.