Evaluating the reliability of physiological and kinematic features for tracking cognitive load changes during repetitive task execution

• Attempt-aligned segmentation replaces fixed windows for cognitive-load tracking. • EDA and HRV show robust rest-task shifts under visual and code instructions. • Wrist kinematics capture early practice effects; short-window HRV is limited. • Attempt duration dropped sharply after the first repetition, consistent with rapid learning. • A compact multimodal feature set supports attempt-wise monitoring. Industry 5.0 calls for human-centric monitoring of workers’ cognitive states. This study focuses on attempt-by-attempt monitoring, testing whether physiological (electrodermal activity, EDA; heart rate variability, HRV) and wrist-kinematic (acceleration) features can track changes in cognitive load during a repeated assembly-like task under two instruction formats (visual and code-based). We used a dataset from a previous experiment in which 30 participants performed multiple repetitions per session. Rather than using fixed time windows or session averages, we segmented the signals at true attempt boundaries and used linear mixed models (LMMs) to (1) compare task attempts to baseline and (2) quantify within-session changes across attempts. Task execution produced clear rest-to-task shifts: EDA increased, heart rate rose, and RR intervals shortened under both formats, confirming robust autonomic engagement. Within sessions, attempt duration dropped sharply after the first trial (after the second in the code session), indicating rapid learning. Attempt-wise sensitivity differed by modality. EDA showed meaningful changes, most consistently a progressive increase in SC area during the visual session, whereas short-window HRV showed modest and inconsistent attempt-level effects. Kinematic features were most responsive to early practice; many acceleration features changed between the first two attempts, then plateaued as motor execution stabilized. Overall, a compact, multimodal feature set, especially wrist kinematics complemented by EDA, can index learning and adaptation within a single session at the level of individual attempts, while HRV reliably distinguishes rest from task. These findings support attempt-aligned multimodal monitoring and motivate validation in richer industrial settings with balanced repetitions for real-time cognitive load tracking.