TL;DR
This paper presents a hybrid Brain-Computer Interface (BCI) combining EEG and NIRS that achieves higher information transfer rates (ITRs) by using shorter trial lengths and intuitive mental tasks. This approach significantly improves BCI performance compared to unimodal systems and previous hybrid systems, making BCIs more practical for real-world applications.
What Problem Does This Paper Address?
Traditional Brain-Computer Interfaces (BCIs) often rely on electroencephalography (EEG) or near-infrared spectroscopy (NIRS) individually. While EEG offers high temporal resolution, it's susceptible to motion artifacts. NIRS, on the other hand, is more robust to artifacts but has a lower temporal resolution due to the slow hemodynamic response of the brain. Hybrid BCIs combining EEG and NIRS aim to leverage the strengths of both, but often require long trial lengths (e.g., 10 seconds or more) to adequately capture the NIRS signal, which in turn degrades the Information Transfer Rate (ITR). Furthermore, many existing BCI paradigms, like motor imagery, require extensive training, limiting their practicality. This paper addresses the challenge of improving ITR in hybrid BCIs by reducing trial length and employing intuitive mental tasks that require minimal training.
Methods
The researchers developed a hybrid EEG-NIRS BCI system. Key methodological aspects include:
- Intuitive Mental Tasks: Instead of motor imagery, they used mental arithmetic (MA) and word chain (WC) tasks. These tasks are considered more intuitive and require less preliminary training.
- Shortened Trial Length: The task period was reduced to 5 seconds, significantly shorter than the typical 10+ seconds used in previous NIRS-based BCIs. This was done to boost the ITR, despite the potential challenge of capturing the full hemodynamic response.
- Hybrid Data Fusion: EEG and NIRS data were simultaneously recorded. For NIRS, data was analyzed across three sub-time windows (0-2s, 2-4s, 4-5s) to capture the hemodynamic response more effectively within the short trial. EEG features were extracted using Filter Bank Common Spatial Patterns (FBCSP).
- Classification: A shrinkage linear discriminant analysis (sLDA) was used as the classifier. A meta-classification approach was employed, where individual classifiers for EEG, HbR, and HbO were combined.
- Evaluation: Both offline and pseudo-online analyses were conducted to assess classification accuracy and Information Transfer Rate (ITR).
Datasets
The study utilized a dataset collected from 10 participants. This dataset was collected specifically for this study and is referred to as an "open access dataset for EEG+NIRS single-trial classification" in related work (Shin et al., 2017). The paper also references and compares its results to data from previous studies [31-33].
Key Insights and Ideas
- Short Trials Can Work: Contrary to the expectation that short trials would severely degrade performance, the hybrid approach with EEG effectively compensated for the limited NIRS signal capture, maintaining high classification accuracy.
- Intuitive Tasks are Key: Using MA and WC tasks, which are easily understood and require minimal training, makes the BCI system more accessible and practical for a wider range of users.
- WC as a Viable BCI Task: The study validates the suitability of the word chain task for BCI applications, a task that has been underutilized compared to MA.
- Synergy of EEG and NIRS: The results strongly support the complementary nature of EEG and NIRS, demonstrating that their fusion significantly outperforms either modality alone, especially under time constraints.
Key Results
-
Offline Classification Accuracy:
- MA vs. BL: 90.0 ± 7.1%
- WC vs. BL: 85.5 ± 8.1% These accuracies were significantly higher than unimodal EEG or NIRS BCIs.
-
Pseudo-Online Classification Accuracy:
- MA vs. BL: 85.8 ± 8.6%
- WC vs. BL: 79.5 ± 13.4% Again, these were superior to unimodal systems.
-
Information Transfer Rate (ITR) Improvement:
- For MA vs. BL, the average ITR was 96.6% and 127.8% higher than previous studies using longer trial lengths.
- For WC vs. BL, the average ITR was 87.1% higher than a previous study. These improvements are attributed to the combination of higher classification accuracy and shorter trial lengths.
Limitations and Future Work
- Pseudo-Online vs. True Online: While the pseudo-online analysis suggests feasibility for online application, it's not a true online implementation. Further work is needed to optimize the calibration phase for real-time BCI systems.
- Task Difficulty: The MA task used a specific range of subtrahends (6-9) to maintain difficulty. Further investigation into how varying task difficulty affects BCI performance could be beneficial.
- WC Task Nuances: While WC is validated, the difficulty level can be influenced by the initial letter, which was managed by varying it. More controlled studies on WC difficulty could be explored.
- Individual Differences: The study acknowledges that optimal BCI task pairs might vary between individuals, suggesting future work to examine this.
- Verbal Fluency Comparison: The paper notes the underutilization of WC and suggests future research to compare its performance against more traditional verbal fluency tasks.
Summary updated