Efficacy of personalized models in discriminating high cognitive demand conditions using text-based interactions.

Abstract

Although high cognitive demand conditions can impair psychological, physical, and behavioral processes without appropriate management, current measurement methods are too cumbersome for continuous monitoring of cognitive demand, and do not account for individual differences. This research uses keystroke and linguistic markers of typed text to construct individualized models of cognitive demand response to discriminate high and low cognitive demand conditions, the results of which can have implications for design of cognitive demand monitoring systems for personalized health management. We constructed within-subject models of cognitive demand response for nine participants and one between-subjects model based on 20 participants. The AUCs for personalized models ranged from 0.679 to 0.953 (Mean=0.826, SD=0.085), significantly higher than chance (p < 0.0001) and the 0.714 AUC for the generic model (p=0.002). Although the features in each model were different, the most common features across models are rate of negative emotion, lexical diversity, rate of words over six letters, and word count. These results confirm significant individual differences in cognitive demand response and suggest that those developing measurement methods used in a monitoring system should consider adaptation to individual characteristics. Our research operationalizes the effects of cognitive demand on HCI and contributes a unique combination of text and keystroke features used to detect high cognitive demand situations.