Cognitive load detection in English language learning using wearable biosensors: A machine learning approach

Article ID: 892

Abstract

To increase the use of wearable biosensors in language learning environments, approaches for accurately extracting small signs of cognitive load are necessary. However, assessing subjective cognitive states, such as the load experienced during language acquisition, provides significant obstacles. This research uses data from physiological sensors worn on the wrist, such as skin conductance, skin temperature, heart rate, and R-R intervals, to organize a machine learning (ML) challenge to develop techniques for quantifying cognitive load in English learners. Participants used data from respondents who completed English language tasks of various difficulty levels. A robust evaluation of preprocessing approaches such as Z-score normalization, signal detrending, and moving average filtering, as well as feature extraction methods such as time-domain and frequency-domain analysis, demonstrated that robust models efficiently used biosensor data. Classical classifiers, such as Adaptive Random Forest (ARF), performed better when optimized with Barnacle Mating Optimization (BMO) for hyperparameter tuning. The proposed method of BMO-ARF has attained accuracy at 95.89%, F1-score in the cognitive load of low at 0.95, medium at 0.90 and high at 0.97, sensitivity in the cognitive load of low at 80.3%, medium at 88.5% and high at 93.0% and specificity in the cognitive load of low at 87.5%, medium at 91.8% and high at 95.1%. The results show that cognitive load classifications were more accurate for higher-difficulty tasks and particular learners, potentially impacted by model overfitting and the subjective nature of physiological responses. The research highlights the need for more sophisticated annotation techniques to improve cognitive load monitoring in language learning environments and handle student response variability.