To evaluate machine learning algorithms' mental workload estimation using NASA-TLX subjective evaluations in undergraduate students during multitasking. The study includes 74 students taking the Labin the Wild multitasking tests and NASA-TLX self-assessment. Participants rated six workload factors. KNN, Naive Bayes, and Extra Trees were tested using stratified k-fold cross-validation. Significant workload dimensions affecting categorization were identified using statistical analysis, specifically ANOVA. The K-Nearest Neighbours (KNN) algorithm estimated mental workload with the highest classification accuracy (88.67%) and recall (90.67%). With over 80% accuracy, Naive Bayes and Extra Trees classifiers performed well. ANOVA showed that Mental Demand, Effort, Performance, and Frustration differed across workload levels, emphasizing their importance in workload evaluation. In contrast, Physical and Temporal Demand were less correlated. Cognitive and emotional criteria appear to classify mental workload more than physical or time-based measurements. The KNN learning curve showed robust generalization without overfitting. The classification report showed great precision and recall for most classes, making KNN dependable. The study supports NASA-TLX's integration with machine learning to improve job allocation and user experience design, and it validates NASA-TLX as a reliable instrument for cognitive workload evaluation. This study uniquely integrates NASA-TLX with machine learning to discover KNN as a superior classifier and identify cognitive characteristics important for mental workload estimation in multitasking.
