A Predictive Framework for Optimizing Cognition during Human–Wearable Robot Interactions in the Construction Industry

Abstract Despite advancements in safety protocols, the construction industry continues to record high cases of musculoskeletal disorders. In response, the industry has adopted emerging technologies such as exoskeletons to mitigate musculoskeletal disorders. While exoskeletons have shown effectiveness in mitigating musculoskeletal disorders, recent studies indicate that human–wearable robot interaction during physically intensive tasks may introduce additional cognitive demand. This study presents a machine-learning-based predictive framework for mitigating cognitive risks during wearable-robot-assisted masonry tasks. Wearable electrodermal activity (EDA) and electroencephalography (EEG) sensors were used to collect physiological data from 19 participants during masonry tasks assisted by an exoskeleton. This data set was subsequently utilized to develop a predictive framework comprising both EEG and EDA models. The NASA Task Load Index (TLX) mental load subscale was employed to categorize cognitive risk states (low, medium, high), providing validated labels for machine learning classification. Results from EDA-based classification showed that the ensemble (bagged tree) model achieved the highest validation accuracy (74.39%). Other classifiers, such as <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="k" display="inline" overflow="scroll"> <mml:mi>k</mml:mi> </mml:math> -nearest neighbor KNN (medium) and tree (fine), also demonstrated competitive recall and precision performance, with faster computational times. EEG-based models significantly outperformed EDA models, with the KNN (fine) classifier achieving an accuracy of 96.41% and consistently high precision and recall across all cognitive risk levels, demonstrating the strong discriminative power of EEG data. Comparative analysis revealed that while EEG offers superior predictive performance and robustness, EDA models demonstrated faster prediction speeds ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="7,700 en dash 35,000 o b s divided by normal s" display="inline" overflow="scroll"> <mml:mn>7,700</mml:mn> <mml:mi>–</mml:mi> <mml:mn>35,000</mml:mn> <mml:mtext> </mml:mtext> <mml:mi>obs</mml:mi> <mml:mo stretchy="false">/</mml:mo> <mml:mi mathvariant="normal">s</mml:mi> </mml:math> ) and shorter training times (1–6 s) compared to EEG ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="350 en dash 720 o b s divided by normal s" display="inline" overflow="scroll"> <mml:mn>350</mml:mn> <mml:mi>–</mml:mi> <mml:mn>720</mml:mn> <mml:mtext> </mml:mtext> <mml:mi>obs</mml:mi> <mml:mo stretchy="false">/</mml:mo> <mml:mi mathvariant="normal">s</mml:mi> </mml:math> ; <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" alttext="tilde 540" display="inline" overflow="scroll"> <mml:mo form="prefix">∼</mml:mo> <mml:mn>540</mml:mn> </mml:math> to 1,588 s). This suggests that EDA can be suitable for real-time deployment where rapid feedback is essential. The findings advance adaptive safety systems designed to enhance workers’ well-being and performance in human–wearable robot interactions.