Detecting and Improving Human Cognitive State in Real-Time Using Data-Driven Adaptive Systems: A Systematic Review

Changes in human attention, workload, or alertness over time can affect task performance and may even increase the risk of injury. Detecting these changes in real time can be beneficial in improving system performance and safety. We reviewed 27 studies that developed models to sense physiological signals, classify one's cognitive state, and deliver automated intervention. Interventions included providing real-time feedback, adjusting the task's difficulty, or modifying automation levels across driving, education, rehabilitation, and human-robot collaboration applications. The findings showed that electroencephalography (EEG) sensors were used in 70% of studies, with attention (56%) and mental workload (26%) considered as the most targeted cognitive states. Within-subject classification reached 81.85-95.81% for multi-class tasks in laboratory settings. The most common interventions included neurofeedback display (30%) and task difficulty adjustment (19%), while automation adjustment was less frequent (11%). Only 33% of studies mentioned a latency of 15 milliseconds to 2.5 s, and all systems operated reactively by detecting cognitive states after their onset rather than anticipating them. The provided recommendations focus on the detection of multiple interacting cognitive states and predictive cognitive state trajectories. This review presents key directions for future research and provides a foundation for designing more effective cognitive state adaptive systems.