The study objective was classification of skill level based on the topographical features of the electroencephalogram(EEG) during the most complex Fundamentals of Laparoscopic Surgery(FLS) task.
We developed a novel microstate-based Common Spatial Pattern (CSP) analysis with linear discriminant analysis(LDA) classification that was compared with topography-preserving convolutional neural network(CNN) based approach to distinguish experts versus novices based on EEG. Ten expert surgeons and thirteen novice medical residents were recruited at the University at Buffalo. After informed consent, the subjects performed three trials of laparoscopic suturing and knot tying with rest periods in-between. 32-channel EEG during task performance was used to analyze spatial patterns of brain activity in 8 expert surgeons (2 dropouts due to data quality) and 13 novice medical residents. Besides conventional CSP analysis, microstate analysis was applied for preprocessing before CSP analysis for improved classification using LDA with 10-fold cross-validation. Also, a topography-preserving 3D CNN model (ESNet) was applied that considered both spatial and temporal information for the classification. Here, 5-fold cross-validation was repeated 10 times, and the results of each iteration of the testing data set were evaluated using indices, Accuracy, F1 score, Mathews Correlation Coefficient (MCC), sensitivity, and Specificity.
Microstate-based CSP analysis found that while novices had primarily the frontal cortex involved for a maximum of spatial pattern vectors, experts had the hotspot of the spatial pattern vectors over the frontal and parietal cortices where the discriminating parietal brain region was supported by the Gradient-weighted Class Activation Mapping (Grad-CAM) of our 3D CNN-based model. Here, LDA with 10-fold cross-validation achieved more than 90% classification accuracy with microstate-based CSP, while conventional regularized CSP could reach around 80% classification accuracy. Then, 3D CNN provided the highest sensitivity of 99.30%, the highest specificity of 99.70%, the highest F1 score of 98.51%, and the highest MCC of 97.56%.
Microstate-based CSP analysis improved the LDA classification (~90%) of experts versus novices based on EEG topography during a complex FLS task; however, combining the spatial and temporal information in the EEG topography preserving 3D CNN model significantly improved the classifier accuracy (>98%) in addition to providing mechanistic insights based on Grad-CAM analysis.