Abstract: Migraine attack early warning captures dynamic changes in physiological, behavioral, and environmental data before an attack to reveal potential prodromal patterns, helping patients intervene early to alleviate suffering. In recent years, inspired by rapid advances in machine learning and wearable devices, many studies have integrated diverse data sources with artificial intelligence models and extended them to the field of migraine attack prediction, endowing early warning models with personalized, multimodal perception capabilities. However, systematic classification and discussion of machine learning-based migraine early warning algorithms remain limited. This paper first classifies relevant methods according to data sources and modeling strategies, elaborates on their technical approaches, and evaluates their advantages, disadvantages, and applicable scenarios, thereby providing methodological references for building migraine early warning systems. Second, it summarizes the strengths and weaknesses of multimodal fusion models, time-series models integrating physiological and diary data, and personalized models, noting that patients with episodic migraine generally derive greater benefits than those with chronic migraine. Finally, it identifies current challenges such as small sample sizes, lack of external validation, and inconsistent evaluation standards, and proposes future directions including the construction of standardized benchmark datasets, development of dedicated time-series architectures, cross-subtype adaptive modeling, and federated learning with edge deployment.