Investigation of features reduction methods for improving EMG and EEG pattern recognition robustness

Reducering af dimensioner til forbedring af robustheden for EMG og EEG mønstergenkendelse

Nørgaard, Astrid Clausen

4. term

Biomedical Engineering and Informatics, Master

2015

2015-06-03

54

Forskning i feature-reduktion overser ofte robusthed, selv om biomedicinske signaler kan variere fra dag til dag. Vi undersøgte, hvor robuste otte metoder til feature-reduktion er, når data indsamles over flere dage. Metoderne blev testet på to datasæt: elektromyografi (EMG), som måler elektrisk aktivitet i muskler, indsamlet over tre dage, hvor otte deltagere udførte syv forskellige håndbevægelser; og elektroencefalografi (EEG), som måler elektrisk aktivitet i hjernen, indsamlet over syv dage, hvor syv deltagere udførte to forskellige dorsifleksioner (opadbøjende bevægelser). Feature-reduktion havde stor betydning for både nøjagtighed og stabilitet i EMG- og EEG-klassifikation. For EMG gav nonparametrisk diskriminantanalyse (NDA) høj nøjagtighed og var den mest robuste metode. For EEG gav kernel principal component analysis (KPCA) den højeste nøjagtighed og var blandt de mest robuste. For at bygge klassifikationssystemer, der holder sig pålidelige over tid, bør man inkludere feature-reduktion og afprøve flere metoder for at finde det bedste match til data.

Research on feature reduction often overlooks robustness, even though biomedical signals can change from day to day. We examined how robust eight feature reduction methods are when data are recorded across multiple days. We tested them on two datasets: electromyography (EMG), which measures muscle electrical activity, collected over three days as eight participants performed seven different hand movements; and electroencephalography (EEG), which measures brain electrical activity, collected over seven days as seven participants performed two different dorsiflexions (upward flexing movements). Feature reduction had a large impact on both accuracy and stability of EMG and EEG classification. For EMG, nonparametric discriminant analysis (NDA) achieved high accuracy and was the most robust. For EEG, kernel principal component analysis (KPCA) achieved the highest accuracy and was among the most robust. To build classification systems that remain reliable over time, include feature reduction and compare methods to find the best match for the data.

[This abstract was generated with the help of AI]

Mønstergenkendelse ; Robusthed ; EMG ; EEG ; Reducering af dimensioner

Download
View record in AAU Student Projects