Abstract Background Multimodal neuroimaging enables integrated assessment of brain structure, function, metabolism, and connectivity, yet progress remains fragmented across methods, modalities, and clinical applications. Despite growing evidence that multimodal machine learning outperforms single-modality approaches, clinical translation has stalled due to unresolved challenges in harmonization, external validation, and deployment. Methods We searched PubMed, Scopus, and Web of Science from January 2005 through March 2025 and synthesized 127 peer-reviewed studies spanning Alzheimer's disease, Parkinson's disease, epilepsy, multiple sclerosis, and traumatic brain injury. Findings were organized using the Integrated Multimodal Fusion in Neuroimaging (IMFN) framework, a five-domain taxonomy covering Integration Architecture, Methodological Standardization, Fusion Algorithms, Neurological Applications, and Clinical Translation. Results Multimodal machine learning achieved pooled sensitivity of 94.6% (95% CI: 90.76-96.89%) for Alzheimer's disease versus healthy controls and 83.8% (95% CI: 78.87-87.71%) for mild cognitive impairment. External validation consistently reduced accuracy by 10-15 percentage points, exposing recurrent failure modes including site confounding, inconsistent preprocessing, and missing-modality handling. Clinical workflow vignettes demonstrated that multimodal integration produced the largest diagnostic gains in ambiguous cases where single modalities failed. Conclusions Current multimodal neuroimaging methods show strong discriminative performance under controlled conditions but face substantial barriers to clinical deployment. We propose Clinical Implementation Readiness criteria to evaluate translational maturity and identify prospective utility studies, standardized fusion pipelines, and equitable dataset development as priorities for closing the gap between research performance and reliable clinical use.
Lin et al. (2026) studied this question.