A causal machine learning approach for investigating learners’ mental states through electroencephalography (EEG)

Abstract The non-invasive technique of electroencephalography (EEG) has become increasingly popular in recent years as a way to use computers to understand the complex psychological behavior of the human mind. Further fascinating insights into human psychology can be gained from the study of brain-computer interfaces using EEG. However, the extremely low band frequency and high sensitivity of the signals to noise and muscle activity make it challenging to use EEG in brain contact successfully, resulting in complex computation and machine learning tasks. For two distinct scenarios, this work introduces a novel combination of statistical causal inference and machine learning to improve brain–computer interaction performance. A publicly accessible dataset in the first case categorizes confused learners' answers, revealing their degree of confusion with Massive Open Online Course stimuli, while the second case categorizes a learner's level of stress while watching videos with multiple stimuli. Weights of balance for non-EEG variables are determined using propensity score matching, and they are then integrated into the causal model using machine learning techniques for brain wave feature engineering. The suggested approach is validated using algorithms from SVM, Random Forest, Gradient Boosting for Additive Model, Bagged CART Model, and Bayesian Generalized Linear Model. LightGBM outperforms other models with SVM (≈ 96%), Gradient Boosting for Additive Model (≈ 91%), Bagged CART Model (≈ 97%), and Bayesian Generalized Linear Model (≈ 58%) with an accuracy of 98–4%, according to the results. By directing confounding factors that are usually overlooked in conventional methods, Propensity Score Matching, which incorporates causal inference, the suggested method dramatically increases classification accuracy.