Physiological Pattern Recognition Of Cognitive Load

Cognitive load is a kind of psychological load imposed on the cognitive system in the state of learning task.Moderate cognitive load can keep learners' learning efficiency in a relatively optimal state,while too high cognitive load will significantly reduce learners' efficiency,and long-term cognitive load mismatching will make learners give up learning.Therefore,cognitive load detection is of great significance for improving the quality of education in classroom teaching and especially in distance education.As a non-invasive neurophysiological measurement method,ECG and EEG are widely used in cognitive load detection,which makes it possible to objectively measure cognitive load and learning state in distance education.This thesis proposes cognitive load pattern recognition models based on ECG and EEG signals.The models can distinguish the resting baseline(BL)state from the cognitive load(CL)state,and recognize cognitive load matching(CLM)state or cognitive load mismatching(CLMM)state.The aim of these models is to help teachers detect the cognitive load and learning state in classroom teaching,so as to adjust teaching strategies in time.The specific research contents and results are divided into the following two parts.(1)The physiological pattern recognition model of cognitive load was established based on the data samples of public database.Firstly,six features were extracted from each channel of the 20-channel EEG signal,and 27 features were extracted from the ECG signal.For the imbalanced data set,the Borderline-SMOTE1 algorithm was used to oversample the data,so that the number of samples in each category was equal.Secondly,the feature subsets that were crucial to the above binary classification problems were selected from the original feature subsets by using Sequential Backward Selection(SBS)algorithm and Particle Swarm Optimization(PSO)algorithm.Finally,several conventional classifiers were used to solve the above two classification problems.All the classification models were verified by leave-one-subject-out cross validation.The results showed that:by combining EEG and ECG,the S VM(Support Vector Machine)classifier achieved the highest accuracy of 94.4%(sensitivity=96.1%,specificity=92.9%)in distinguishing resting baseline(BL)state and cognitive load(CL)state,and the highest accuracy of 96.3%(sensitivity=100%,specificity=90.0%)in distinguishing cognitive load matching(CLM)and mismatching state(CLMM).(2)A new physiological data set of cognitive load was constructed to test the effectiveness of the above-mentioned physiological modeling method of cognitive load.Firstly,the ECG and EEG signals of 30 college freshmen and juniors were collected,and the data samples were labeled as CL,BL,CLMM or CLM according to the operational definition.Secondly,six features were extracted from each channel of 128-channel EEG signals,and 27 features was extracted from the ECG signal.The Borderline-SMOTE1 algorithm was also used to obtain balanced data set of the binary classification problems.The data set was divided into two parts:training/test set and validation set.Then,starting from a high-dimension original feature set,particle swarm optimization and sequential backward selection were used to find a critical low-dimension feature subset for the CL vs.BL and CLM vs.CLMM classification problems.Finally,the selected feature subset and the trained classifier were verified by using the validation data set.The classification results showed that:by combining the features of EEG and ECG and using SVM classifier to distinguish resting baseline(BL)state and cognitive load(CL)state,the highest validation accuracy is 80.9%(sensitivity=71.8%,specificity=90.1%),and the highest accuracy is 66.7%(sensitivity=56.7%,specificity=76.7%)to distinguish cognitive load matching(CLM)state and mismatching(CLMM)state.The validation set is completely independent of the classifier training and feature selection process,which reflects that the EEG and ECG based physiological recognition models of cognitive load have much better generalization accuracy than random guess.The findings of this thesis are as follows.Cognitive load state and resting baseline state have distinguishable neurophysiological patterns,and cognitive load matching and mismatching states also have distinguishable neurophysiological patterns,but the pattern recognition of CLM and CLMM is far from practical application.Using machine learning methods to automatically determine whether a learner is in learning state is more difficult than determining whether the difficulty of the learning task is suitable for the learner.