Multimodal AI in hematology: a systematic review of fusion approaches for automated diagnosis of blood disorders

Abstract

Hematological disorders, including anemia and leukemia, pose major health risks globally, especially in developing areas. Around 1.92 billion people are affected by anemia, with about 475,000 new leukemia cases diagnosed yearly, resulting in approximately 310,000 deaths. The reliance on subjective conventional diagnostic methods highlights the necessity for automated diagnostics. The systematic review evaluates AI techniques for diagnosing hematological diseases, focusing on multimodal approaches that integrate blood smear images with hematological parameters to enhance diagnostic accuracy. It contrasts multimodal and single-modality systems on criteria like accuracy, sensitivity, and specificity, while addressing challenges such as data size and the absence of explainable AI. The review suggests strategies for improving clinical AI implementation, including data standardization, multicenter validation, and development of portable diagnostic systems. A literature search following PRISMA 2020 guidelines identified 3,000 records, with 220 removed prior to screening (100 duplicates, 120 non-English). From 2,780 screened, 2,600 were excluded, resulting in 180 reports, of which 40 were included in the final evidence set. Only peer-reviewed original studies meeting multimodal criteria were synthesized. Data extraction included study characteristics, datasets, AI models, fusion strategies, validation methods, and performance metrics. Risk of bias was assessed with PROBAST and QUADAS-2, while evidence certainty was evaluated using the GRADE framework. Multimodal and fusion systems in the reviewed studies demonstrated potential, particularly through hybrid fusion and deep learning architecture. However, variability in disease, datasets, and evaluation methods hindered quantitative comparisons with single-modality systems, resulting in inconsistent performance evidence. Limitations included small, imbalanced datasets and inadequate multicenter validation, yielding a low - GRADE quality assessment due to bias risks and validation shortcomings. Multimodal AI in hematology shows potential for enhancing automated diagnosis and classification by merging various data types. Nonetheless, existing evidence lacks consistency, hindering clear superiority over single-modality techniques. Future studies should prioritize developing explainable AI models, using standardized datasets, exploring privacy-preserving federated learning, and creating lightweight models for limited-resource settings.