Microsleep episodes are brief involuntary transitions into sleep-like states that occur during wakefulness and are strongly associated with fatigue, reduced vigilance, and cognitive overload. Although microsleeps may last only a few seconds, they can severely impair attention, reaction time, perception, and decision-making, posing substantial risks in safety-critical environments such as transportation, industrial manufacturing, healthcare monitoring, aviation, and military operations. Accurate prediction of microsleep onset before behavioural failure remains a major challenge due to the transient, heterogeneous, and low signal-to-noise nature of the underlying physiological processes. Existing approaches are further constrained by limited subject generalisation, heavy dependence on handcrafted features, and insufficient understanding of multimodal physiological interactions underlying vigilance decline. This thesis addresses these challenges through the development of robust EEG–EOG-based machine learning and deep learning frameworks for multimodal microsleep prediction, with a focus on physiologically meaningful representation learning, temporal modelling, multimodal interaction analysis, and practical real-world applicability. This thesis investigates microsleep prediction through a sequence of studies examining physiological feature representation, temporal forecasting, multimodal EEG–EOG integration, and label-efficient learning. The findings show that physiologically meaningful EEG features provide informative and discriminative representations for microsleep prediction, while temporal modelling substantially improves forecasting performance by capturing the progressive transition into microsleep states. The integration of EEG and EOG further improves predictive performance and provides insight into cross-modal physiological interactions preceding microsleep onset through information-theoretic measures including mutual information and transfer entropy. Finally, a multimodal disagreement-based active learning framework demonstrates that annotation requirements can be reduced considerably while maintaining competitive forecasting performance. Collectively, these results indicate that combining temporal modelling, multimodal physiological information, and label-efficient learning yields more robust and scalable microsleep prediction systems. Overall, this thesis contributes a comprehensive framework for multimodal microsleep prediction that combines physiological feature analysis, temporal forecasting, multimodal interaction modelling, and label-efficient learning. The findings advance understanding of microsleep-related neurophysiology while providing practical insights for the development of future vigilance-monitoring systems. More broadly, the methodologies developed in this work may be applicable to fatigue monitoring, neuroergonomics, biomedical signal analysis, and related brain–computer interface applications.