This study conducts a systematic literature review to investigate the role of artificial intelligence (AI) in Multimodal Learning Analytics (MMLA). Analyzing 87 publications (2016-2025) from Web of Science, it identifies four core functions of AI in MMLA:Automated analysis of learning states (cognitive load, affective states, etc.);Personalized feedback/interventions (real-time scaffolding, intelligent tutoring systems);Multimodal data fusion (integrating physiological, behavioral, and log data);System architecture development (MMLA platform frameworks).Supervised learning and deep learning emerge as dominant methods, demonstrating high efficacy in collaborative learning and special education (e.g., 98% behavior prediction accuracy). Critical challenges include limited model generalizability, "black-box" interpretability issues, and ethical risks (FATE: Fairness, Accountability, Transparency, Ethics). The review proposes seven future directions—enhancing cross-context robustness, developing explainable AI (XAI), establishing ethical guidelines—providing a roadmap for AI-enabled educational innovation.