کد مقاله کد نشریه سال انتشار مقاله انگلیسی نسخه تمام متن
504799 864435 2016 9 صفحه PDF دانلود رایگان
عنوان انگلیسی مقاله ISI
Dynamic frequency feature selection based approach for classification of motor imageries
ترجمه فارسی عنوان
رویکرد مبتنی بر انتخاب ویژگی فرکانس پویا برای طبقه بندی تصویرسازی های حرکتی
کلمات کلیدی
موضوعات مرتبط
مهندسی و علوم پایه مهندسی کامپیوتر نرم افزارهای علوم کامپیوتر
چکیده انگلیسی

Electroencephalography (EEG) is one of the most popular techniques to record the brain activities such as motor imagery, which is of low signal-to-noise ratio and could lead to high classification error. Therefore, selection of the most discriminative features could be crucial to improve the classification performance. However, the traditional feature selection methods employed in brain-computer interface (BCI) field (e.g. Mutual Information-based Best Individual Feature (MIBIF), Mutual Information-based Rough Set Reduction (MIRSR) and cross-validation) mainly focus on the overall performance on all the trials in the training set, and thus may have very poor performance on some specific samples, which is not acceptable. To address this problem, a novel sequential forward feature selection approach called Dynamic Frequency Feature Selection (DFFS) is proposed in this paper. The DFFS method emphasized the importance of the samples that got misclassified while only pursuing high overall classification performance. In the DFFS based classification scheme, the EEG data was first transformed to frequency domain using Wavelet Packet Decomposition (WPD), which is then employed as the candidate set for further discriminatory feature selection. The features are selected one by one in a boosting manner. After one feature being selected, the importance of the correctly classified samples based on the feature will be decreased, which is equivalent to increasing the importance of the misclassified samples. Therefore, a complement feature to the current features could be selected in the next run. The selected features are then fed to a classifier trained by random forest algorithm. Finally, a time series voting-based method is utilized to improve the classification performance. Comparisons between the DFFS-based approach and state-of-art methods on BCI competition IV data set 2b have been conducted, which have shown the superiority of the proposed algorithm.

ناشر
Database: Elsevier - ScienceDirect (ساینس دایرکت)
Journal: Computers in Biology and Medicine - Volume 75, 1 August 2016, Pages 45–53
نویسندگان
, , , ,