A Photoplethysmography Based Mental Workload Evaluation Using Ensembled CatBoost Approach

• Published in: Journal of the Institution of Engineers (India). Series B, Electrical Engineering, Electronics and telecommunication engineering, Computer engineering Vol. 106; no. 1; pp. 165 - 180
• Main Authors: Pemmada, Suresh Kumar, Nayak, Janmenjoy, Routray, Ashanta Ranjan
• Format: Journal Article
• Language: English
• Published: New Delhi Springer India 01.02.2025; Springer Nature B.V
• Subjects: Accuracy; Communications Engineering; Electrocardiography; Engineering; Galvanic skin response; Networks; Original Contribution; Performance evaluation; Psychological stress; Wearable technology; Working conditions; Workload; Workloads; N-back; Ensemble learning; Mental workload; Wearable sensors
• ISSN: 2250-2106; 2250-2114
• DOI: 10.1007/s40031-024-01092-1

The increasing presence of cognitive activities needs the use of increased brain resources. This stress, called mental workload, may result in poor task performance. As a result, assessing mental workload will provide an appropriate working environment for a subject to boost work productivity or increase safety in high-risk work environments. The current research evaluates the dependability and capabilities of consumer wearable devices against laboratory equipment in differentiating certain mental states. Fingertip photoplethysmography (PPG), Galvanic Skin Response (GSR), and Electrocardiography (ECG) signals were obtained from 22 individuals with distinct working scenarios were considered in order to produce varying mental stress, strain, and emotional state. These signals were used to train CatBoost, which used a variety of classification metrics to classify the various mental workload levels. Results indicated that the CatBoost exhibited significant accuracy in accessing mental workload, as evidenced by an ECG with an accuracy of 0.83, TPR of 0.79, FPR of 0.13, precision of 0.86, TNR of 0.87, F1-Score of 0.82, and AUC-ROC of 0.83. Similarly, signals generated by PPG from the fingertip and wristband had accuracy levels of 0.84 and 0.75, respectively, along with comparable performance measures. Additionally, seven-fold cross-validation has been employed to verify the robustness of our method and the accuracy of the results.