Feasibility of using Error-related potentials as an appropriate method for adaptation in a brain computer interface

Leerskov, Kasper Kunz ; Hadsund, Johannes Thorling

4. term

Biomedical Engineering and Informatics, Master

2017

2017-06-07

59

This thesis examines whether error-related potentials (ErrPs) can support adaptation of an EEG-based brain–computer interface (BCI) without knowing which mental command was misclassified. Continuous adaptation is crucial in BCIs, but naïve strategies risk reinforcing errors; ErrPs, elicited when users perceive system errors, offer an independent signal to guide adaptation. Twelve healthy participants performed three paradigms that elicited expected and unexpected feedback: overt hand movement, covert motor imagery, and controlling a snake in a game. EEG was recorded from 16 electrodes and segmented into 1‑second epochs following visual feedback. Temporal and spectral features were extracted and used in classifiers (Linear Discriminant Analysis, LDA, and Random Forest) trained either within the same paradigm or on data from the two other paradigms. The most informative features were located over central cortex 400–800 ms after feedback, notably correlation with a theta‑band energy template. LDA consistently outperformed Random Forest, and the area under the ROC curve was reported as highest for the covert motor paradigm (0.92), followed by overt motor (0.86), and lower for the game (0.68). Within‑paradigm accuracies were 81% (covert), 72% (overt), and 56% (game), and with cross‑paradigm training 77%, 74%, and 60%; game performance was significantly above chance (p = 0.05) but lower than for motor tasks. Overall, the results indicate that ErrPs can be detected with comparable accuracy across tasks and that key properties of ErrPs are relatively independent of the misclassified intention—supporting their feasibility for robust, intention‑independent BCI adaptation.

Denne afhandling undersøger, om fejlpotentialer (Error-related potentials, ErrP) kan bruges til at tilpasse et EEG-baseret brain-computer interface (BCI) uden at kende, hvilken mental kommando der blev fejlklassificeret. Baggrunden er, at løbende tilpasning er nødvendig i BCI, men naive strategier kan forværre fejl; ErrP’er, som opstår når brugeren oplever systemfejl, kan give et uafhængigt signal til at styre tilpasningen. Tolv raske deltagere gennemførte tre paradigmer, der fremkaldte forventede og uventede feedback: overte håndbevægelser, coverte (motorisk forestilling) og styring af en slange i et spil. EEG blev målt fra 16 elektroder og opdelt i 1-sekunds epoker efter visuel feedback. Tidslige og spektrale træk blev udtrukket og anvendt i klassifikatorer (Linear Discriminant Analysis, LDA, og Random Forest), som blev trænet enten på data fra samme paradigm eller på data fra de to øvrige paradigmer. De mest informative træk fandtes over central cortex 400–800 ms efter feedback, især korrelation til en skabelon for thetabåndsenergi. LDA overgik systematisk Random Forest, og arealet under ROC-kurven blev rapporteret som højest for det coverte motoriske paradigme (0,92), dernæst det overte (0,86) og lavest for spillet (0,68). Nøjagtigheden ved træning og test inden for samme paradigm var 81 % (covert), 72 % (overt) og 56 % (spil), og ved krydstræning på de to andre paradigmer 77 %, 74 % og 60 %; for spillet var ydeevnen signifikant over tilfældighed (p = 0,05), men lavere end for motorparadigmerne. Samlet peger resultaterne på, at ErrP’er kan detekteres med sammenlignelig nøjagtighed på tværs af opgaver, og at centrale egenskaber ved ErrP’er er relativt uafhængige af den misfortolkede intention—hvilket understøtter deres anvendelighed til robust, intention-uafhængig BCI-tilpasning.

[This apstract has been generated with the help of AI directly from the project full text]

BCI ; EEG ; ErrP ; Adaptation

Download PDF
View record in AAU Student Projects