Enhancing mental workload recognition: a comparison of complexity-based eye movement metrics and conventional features

Conventional eye-movement metrics often produce inconsistent results in mental workload (MWL) recognition, due to their inability capturing dynamic time-series patterns. Emerging evidence suggests that complexity-based features may be better indicators. This study evaluates the effectiveness of complexity-based eye-movement features incorporating intrinsic mode functions (IMFs) as MWL indicators compared to conventional metrics. Participants solved mathematical problems of varying MWL while eye movements were recorded, followed by NASA-RTLX assessment. Eye-movement data were decomposed via empirical mode decomposition, and multiscale entropy was computed. Machine learning models were trained on conventional and complexity-based feature sets, respectively. Results showed that complexity-based features captured MWL effects more consistently and achieved higher classification accuracy in both subjective MWL recognition (68% vs. 53%) and task-based MWL recognition (73% vs. 57%) compared to conventional features. These findings demonstrate a 15-16% improvement in accuracy, reinforcing the potential of complexity-based metrics for enhancing human-computer interaction systems.