Scholarlink

<p>Rising occupational and academic cognitive demands make real-time monitoring of mental workload a critical priority for occupational health and educational technology. Consumer-grade wearable devices offer a scalable and non-intrusive pathway to operationalize cognitive load (CL) measurement outside controlled laboratory settings. This study evaluates whether multimodal physiological features obtained from wearable devices can discriminate a controlled Stroop task from seated rest and examines the limitations of interpreting this experimental-condition classification as a measure of cognitive load. Secondary analysis was performed on the publicly available CogWear dataset (PhysioNet; DOI: <a href="https://doi.org/10.13026/5f6t-b637">https://doi.org/10.13026/5f6t-b637</a>), comprising physiological recordings from 24 participants collected using three concurrent wearable devices the Empatica E4 wristband, Samsung Galaxy Watch4, and Muse S EEG headband during a Stroop cognitive challenge and a resting baseline condition. Features including heart rate variability (HRV), electrodermal activity (EDA), skin temperature (TEMP), photoplethysmography (PPG/BVP) amplitude, and EEG theta and alpha band power were extracted following standardized preprocessing. Four supervised classifiers—Random Forest, SVM with an RBF kernel, SVM with a linear kernel, and L2-regularized logistic regression—were evaluated using window-level five-fold stratified cross-validation and leave-one-subject-out cross-validation. The Random Forest classifier achieved the highest performance among the evaluated models under window-level five-fold cross-validation and remained the strongest evaluated model under LOSO validation. EDA tonic-level measures and PPG/BVP-derived RMSSD were the highest-ranked features for differentiating the Stroop and seated-rest conditions. Cross-device comparison under matched five-fold cross-validation indicated that the Samsung Galaxy Watch4 PPG-derived RMSSD model achieved AUC-ROC = 0.831, compared with 0.864 for the Empatica E4 BVP-derived RMSSD model and 0.891 for the complete multimodal feature set. Under this controlled protocol, multimodal wearable features supported discrimination between the Stroop and seated-rest conditions. Because these conditions differ in cognitive demand, task engagement, motor responding, and autonomic arousal, the findings do not establish a task-independent cognitive-load detector. External validation against subjective workload measures, active low-load control tasks, and naturalistic settings is required before clinical, occupational, educational, or just-in-time intervention use.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AI-supported cognitive load detection: experimental insights using wearable workload and stress data for enhanced well-being interventions

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月4日

查看全文
  • Jing Yu,
  • Jian Wan

摘要

Rising occupational and academic cognitive demands make real-time monitoring of mental workload a critical priority for occupational health and educational technology. Consumer-grade wearable devices offer a scalable and non-intrusive pathway to operationalize cognitive load (CL) measurement outside controlled laboratory settings. This study evaluates whether multimodal physiological features obtained from wearable devices can discriminate a controlled Stroop task from seated rest and examines the limitations of interpreting this experimental-condition classification as a measure of cognitive load. Secondary analysis was performed on the publicly available CogWear dataset (PhysioNet; DOI: https://doi.org/10.13026/5f6t-b637), comprising physiological recordings from 24 participants collected using three concurrent wearable devices the Empatica E4 wristband, Samsung Galaxy Watch4, and Muse S EEG headband during a Stroop cognitive challenge and a resting baseline condition. Features including heart rate variability (HRV), electrodermal activity (EDA), skin temperature (TEMP), photoplethysmography (PPG/BVP) amplitude, and EEG theta and alpha band power were extracted following standardized preprocessing. Four supervised classifiers—Random Forest, SVM with an RBF kernel, SVM with a linear kernel, and L2-regularized logistic regression—were evaluated using window-level five-fold stratified cross-validation and leave-one-subject-out cross-validation. The Random Forest classifier achieved the highest performance among the evaluated models under window-level five-fold cross-validation and remained the strongest evaluated model under LOSO validation. EDA tonic-level measures and PPG/BVP-derived RMSSD were the highest-ranked features for differentiating the Stroop and seated-rest conditions. Cross-device comparison under matched five-fold cross-validation indicated that the Samsung Galaxy Watch4 PPG-derived RMSSD model achieved AUC-ROC = 0.831, compared with 0.864 for the Empatica E4 BVP-derived RMSSD model and 0.891 for the complete multimodal feature set. Under this controlled protocol, multimodal wearable features supported discrimination between the Stroop and seated-rest conditions. Because these conditions differ in cognitive demand, task engagement, motor responding, and autonomic arousal, the findings do not establish a task-independent cognitive-load detector. External validation against subjective workload measures, active low-load control tasks, and naturalistic settings is required before clinical, occupational, educational, or just-in-time intervention use.