Artificial Intelligence for Risk Stratification in Diffuse Large B-Cell Lymphoma: A Systematic Review of...

Background: Diffuse large B-cell lymphoma (DLBCL) is a biologically heterogeneous malignancy, with various outcomes despite significant advances in therapeutic options. Current conventional prognostic tools, e.g., the International Prognostic Index (IPI), lack sufficient precision at an individual patient level. However, artificial intelligence (AI), including machine learning (ML) and deep learning (DL), can enable specialists to navigate complex datasets, with the final aim of improving prognostic models for DLBCL. Objectives: This scoping review aims to systematically map the current literature regarding the use of AI/ML techniques in DLBCL outcome prediction and risk stratification. We categorized studies by data modality and computational approach to identify key trends, knowledge gaps, and opportunities for their translation into current practice. Methods: We conducted a structured search of the PubMed/MEDLINE, Scopus, and Cochrane Library databases through July 2025 using terms related to DLBCL, prognosis, and AI/ML. Eligible studies included original papers applying AI/ML to predict survival outcomes, classify risk groups, or identify prognostic subtypes. Studies were categorized based on input modality: clinical, positron emission tomography/computed tomography (PET/CT) imaging, histopathology, transcriptomics, genomics, circulating tumor DNA (ctDNA), and multi-omics data. Narrative synthesis was performed in line with PRISMA-ScR guidelines. Results: From the 215 records screened, 91 studies met the inclusion criteria. Group-wise we report the following categories: clinical risk features (n = 8), PET/CT imaging (n = 30), CT (n = 1), digital pathology (n = 3), conventional histopathology (n = 2), gene expression profiling (n = 19), specific mutational signatures (n = 18), ctDNA (n = 3), microRNA (n = 2), and multi-omics integration (n = 5). The most common techniques reported amongst the papers included ensemble learning, convolutional neural networks (CNNs), and LASSO-based Cox models. Several AI techniques demonstrated superior predictive performance over IPI, with area under the curve (AUC) values frequently exceeding 0.80. Multi-omics models and ctDNA-based predictors showed strong potential for clinical translation, a perspective worth considering in further studies. Conclusions: AI/ML methods are increasingly used in DLBCL to improve prognostic accuracy by leveraging data types with diverse inputs. These approaches allow an enhanced stratification, superior to traditional indices, and support the early identification of high-risk patients, earlier guidance for therapy tailoring, and early trial enrollment for flagged cases. Future investigations should focus on external validation and improvement of model interpretability, with tangible perspectives of integration into real-world workflows and translation from bench to bedside.