Knee osteoarthritis (KOA) is a common chronic degenerative disease that causes pain, functional limitation, and reduced quality of life in middle-aged and elderly populations. Early KOA is characterized by relatively subtle symptomatic, structural, and functional changes, including knee pain or stiffness, cartilage matrix alterations, meniscal degeneration, bone marrow lesions, and mild abnormalities in load-related function, while radiographic findings may remain atypical. A single imaging modality or clinical indicator is therefore insufficient to characterize the early pathological status of KOA. Deep learning (DL)-based multimodal fusion can integrate X-ray, magnetic resonance imaging, clinical variables, gait/biomechanical data, and potential molecular biomarkers, supporting early identification and risk stratification from the perspectives of bony structure, soft-tissue change, symptom burden, and functional loading. This review summarizes multimodal data types, preprocessing and feature extraction methods, early/intermediate/late fusion strategies, representative DL architectures, and current application evidence for early KOA diagnosis. We further discuss key translational issues, including cost-effectiveness, label consistency, center effects, missing modalities, and multimodal interpretability. Current evidence suggests that multimodal fusion may provide incremental value over unimodal approaches, but its clinical utility requires further confirmation through unified early-stage definitions, standardized acquisition, multi-reader consensus annotation, and multicenter external validation.