Multi-Modal Deep Learning-Based Model to Predict Burkitt Lymphoma Recurrence.

Abstract

Burkitt Lymphoma (BL) is an aggressive B-cell non-Hodgkin lymphoma. Recurrent BL has extremely poor prognosis, and does not have a standard treatment option. Although the disease pathology is well characterized, there is currently a lack of research specifically on predictive modeling of BL recurrence. We developed a deep learning (DL) model, BLIMP (Burkitt Lymphoma multI-Modal recurrence Predictor) to predict BL recurrence by combining clinical, gene expression, and mutation data, utilizing a cohort of 184 patients from a publicly available dataset. Our approach achieved an AUC of 0.788 on a held-out testing set, outperforming traditional machine learning (ML) models. Explainability analysis was performed, and features which contributed to BLIMP's predictions are consistent with known patterns of BL pathophysiology and cancer progression. These results demonstrated the effectiveness of using genomic data in DL-based models for BL recurrence prediction, suggesting promise for future applications to BL research.