快速序列視覺呈現及穩態視覺複合式腦機介面開發與效能評估

Abstract

腦機介面泛指腦部與外部裝置溝通的技術，其中快速序列視覺呈現(rapid serial visual presentation, RSVP) 與穩態視覺誘發電位(steady-state visually evoked potential, SSVEP)均為具高效率的技術，近年來複合式的腦機介面研究開始受到重視，因為其在實際應用下的準確性與指令數量方面有相當的潛力，RSVP是一種高效率事件關聯電位技術，可偵測使用者是否識別到指定項目，目前RSVP實驗設計多為單一項目位置的單指令輸出，然而多項目位置才能在實際應用時提供更多的指令，本研究目的為開發複合RSVP與SSVEP技術腦機系統，藉此解決RSVP無法計算項目位置的問題，並以遊戲形式呈現，探討在持續使用下人的狀態變化對腦機介面使用成效的影響。在複合腦機介面系統中SSVEP用於識別項目位置，RSVP用於識別項目是否為目標項。本實驗採用商用腦電波收錄信號，並為此實驗開發了一個多層分類方法，第一層對SSVEP使用典型相關分析判定受試者注視的刺激頻率(4, 5, 6, 7Hz)，其準確度達97.90%，第二層則判定受試者是否在RSVP中識別到關鍵項目的事件關聯電位經測試Bagging Tree、SVM、LDA及BLDA四個分類器性能，其中Bagging Tree的分類準確度為83.45%，分層分類的整體準確度為81.70%。另外發現受試者在持續使用腦機系統的表現受生理狀態變化影響，前額葉腦區阿法波能量上升隨RSVP識別成效下降出現，同時枕葉腦區西塔波能量上升隨RSVP識別成效下降出現，均有顯著中度負相關。最後根據前述的分類模型開發一套即時的複合式腦機系統，在僅使用四個腦波通道(Cz, CPz, O1, O2)時的整體準確度可達78.10%。

Brain–computer interface (BCI) is a collaboration between the human brain and an external device. Rapid Serial Visual Presentation (RSVP) and Steady State Visually Evoked Potential (SSVEP) are the highly efficient BCI technologies using visual stimulation. Recently, hybrid BCIs started gaining attention due to its immense potential for practical applications in terms of accuracy and number of user commands. RSVP is a high-efficiency technique for examining the visual perception and frequently used to identify the target or nontarget items of human attention. However, the current existing RSVP-based BCI technologies are designing the target position in the same place such that there will be the only single command for the BCI control. Limited target position in the RSVP-based BCI becomes the major challenge for the practical BCI applications. In this study, we aim to develop a hybrid BCI system integrating with RSVP and SSVEP technologies for solving the limited position challenge and exploring the sustained usage of the BCI performance through the neurogaming experiment. Inside the hybrid BCI system, SSVEP is utilized to identify the multiple target position and RSVP is to recognize the target/nontarget. In the experiment setting, we adopted a commercial EEG system to collect the EEG data. We proposed a hierarchical classification process to evaluate the hybrid BCI performance. The first layer is using Canonical Correlation Analysis (CCA) to classify the SSVEP patterns for the target position detection. Its classification accuracy can be achieved 97.90% through four different stimuli (4, 5, 6 and 7Hz). The second layer is using machine-learning classifiers such as Bagging Tree, SVM, LDA, and BLDA to classify the event-related potentials (ERP) in RSVP for the target/nontarget detection. The classification performance can be achieved 83.45%. The overall hybrid system performance was estimated 81.70% by the hierarchical classification. We also observed that participants’ BCI performance significantly affected by their physiological states change due to the negative correlation between the accuracy of target detection in different sessions and frontal alpha and occipital theta powers, while in the sustained use of RSVP. Furthermore, we finally implemented the real-time hybrid BCI system via only using four EEG channels (Cz, CPz, O1, and O2) can be achieved 78.10% of the system performance.