Multimodal artificial intelligence for diagnosis and prognosis in head and neck cancer: A systematic review
TL;DR
Evidence remains limited by retrospective designs, heterogeneous validation strategies, and limited external validation, however, evidence remains limited by retrospective designs, heterogeneous validation strategies, and limited external validation.
Abstract
Multimodal artificial intelligence (AI) is increasingly applied in head and neck oncology to integrate radiological, histopathological, clinical, and molecular data. While individual studies report encouraging performance, further synthesis is needed to characterise the data modalities, fusion strategies, model architectures, evaluation metrics, and validation pathways used in this field. This systematic review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, and the protocol was registered in the International Prospective Register of Systematic Reviews (PROSPERO; CRD420251244308). A systematic literature search was conducted in PubMed, Web of Science, Scopus, and IEEE Xplore for studies published between 2020 and 2025. Study screening and data extraction were performed independently by two reviewers, with disagreements resolved by consensus. Eligible studies employed artificial intelligence or deep learning models that integrated at least two distinct biomedical data modalities for diagnostic or prognostic tasks in head and neck oncology. Methodological quality was assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) framework. Twenty-three studies met the inclusion criteria. Data modalities included computed tomography (CT), positron emission tomography (PET), magnetic resonance imaging (MRI), histopathology, clinical variables, photographic or endoscopic images, and multi-omics data. Multimodal models generally outperformed unimodal approaches, achieving diagnostic area under the receiver operating characteristic curve (AUC) values up to 0.982 and prognostic concordance indices (C-index) up to 0.966, although performance varied substantially across tasks, datasets, endpoints, and validation strategies. However, evidence remains limited by retrospective designs, heterogeneous validation strategies, and limited external validation.