Multimodal and Hybrid Artificial Intelligence for Real-World Decision-Making: Methods, Evidence, and Applications

Abstract

Real-world decision-making rarely depends on a single data stream. Healthcare diagnosis, industrial fault detection, agricultural disease monitoring, business intelligence, cybersecurity threat response, and assistive technology all require AI systems capable of integrating heterogeneous evidence from images, text, physiological signals, sensors, graphs, tabular records, and their combinations. Multimodal and hybrid AI systems address this challenge by combining complementary data modalities, complementary architectures, fusion strategies, domain knowledge, and deployment infrastructures. This review identifies and critically examines four categories of directly multimodal or hybrid evidence, multimodal EEG analysis, vision-audio fusion, hybrid multimodal emotion recognition, and privacy-preserving multimodal cancer diagnosis, alongside a broader set of hybrid, ensemble, attention-based, graph-guided, and Bayesian architectures that advance multimodal integration. It situates these within seven application domains and examines the cross-cutting challenges of fusion design, modality alignment, interpretability, robustness, privacy, computational feasibility, and human oversight. Synthesis reveals that while fusion strategies have diversified from feature concatenation through tensor and attention-based fusion to knowledge-guided integration, evidence validation practices, including fusion ablation, modality-dropout testing, and calibrated uncertainty reporting, remain inconsistently applied. A structured research agenda addresses these gaps with eleven actionable future directions.