Automatic detection of cognitive events using machine learning and understanding models' interpretations of human cognition.

The pupillary response is a valuable indicator of cognitive workload, capturing fluctuations in attention and arousal governed by the autonomic nervous system. Cognitive events, defined as the initiation of mental processes, are closely linked to cognitive workload as they trigger cognitive responses. In this study, we detect cognitive events for the task-evoked pupillary response across four domains (vigilance, emotion processing, numerical reasoning, and short-term memory). The problem is framed as a binary classification. We train one generalized model and four task-specific models on 1-s pupil diameter and gaze position segments. Five models achieve MCC between 0.43 and 0.75. We report three key findings: (1) the generalized model reduces the specificity to enhance the sensitivity, illustrating the trade-off from specialization to generalization; (2) the permutation feature importance analyses show that both pupil dilation and gaze position contribute to model predictions, with task-specific models focusing on task-specific structure patterns to predict while the generalized model is using human cognitive responses; and (3) in an online simulation environment, models performance decreases by approximately 0.05 on MCC. The findings highlight the potential of machine learning applied to pupillary signals for rapid, individualized detection of cognitive events.