Machine Learning Models for the Diagnosis and Prognosis Prediction of High-Grade B-Cell Lymphoma

Abstract

High-grade B-cell lymphoma (HGBL) is a newly introduced category of rare and heterogeneous invasive B-cell lymphoma (BCL), which is diagnosed depending on fluorescence in situ hybridization (FISH), an expensive and laborious analysis. In order to identify HGBL with minimal workup and costs, a total of 187 newly diagnosed BCL patients were enrolled in a cohort study. As a result, the overall survival (OS) and progression-free survival (PFS) of the HGBL group were inferior to those of the non-HGBL group. HGBL (n = 35) was more likely to have a high-grade histomorphology appearance, extranodal involvement, bone marrow involvement, and whole-body maximum standardized uptake (SUVmax). The machine learning classification models indicated that histomorphology appearance, Ann Arbor stage, lactate dehydrogenase (LDH), and International Prognostic Index (IPI) risk group were independent risk factors for diagnosing HGBL. Patients in the high IPI risk group, who are CD10 positive, and who have extranodal involvement, high LDH, high white blood cell (WBC), bone marrow involvement, old age, advanced Ann Arbor stage, and high SUVmax had a higher risk of death within 1 year. In addition, these models prompt the clinical features with which the patients should be recommended to undergo a FISH test. Furthermore, this study supports that first-line treatment with R-CHOP has dismal efficacy in HGBL. A novel induction therapeutic regimen is still urgently needed to ameliorate the poor outcome of HGBL patients.

Keywords: classification models; clinical characteristics; diagnostic predictor; high-grade B-cell lymphoma; machine learning.