Główne pojęcia
EEG data can be effectively processed using established time series classification models, which can outperform specialized EEG classification approaches when incorporating subject-specific information through joint training.
Streszczenie
The paper explores the connection between EEG data and time series classification, challenging the prevailing view of EEG as a specialized domain requiring dedicated models. The key insights are:
- EEG data can be treated as time series data with static attributes (subject information), allowing the use of generic time series classification models.
- Three approaches are proposed to incorporate subject information into time series models: Constant Indicator Channels, Constant Embedding Channels, and Separate Embedding.
- Experiments on three EEG datasets show that time series models with subject-conditional training can match or outperform specialized EEG classification models, especially on the SSVEP and ERN datasets.
- The Inception architecture in particular demonstrates competitive performance and superior computational efficiency compared to the domain-specific MAtt model.
- The results suggest that integrating EEG classification into the broader time series analysis domain can lead to more efficient and better-understood learning on EEG data.
Statystyki
The paper reports the following key metrics:
Accuracy for the MI and SSVEP datasets
Area Under the Curve (AUC) for the ERN dataset
Cytaty
"EEG classification is especially hindered by the fact that EEG signals have an inherently low signal-to-noise ratio (SNR) [15] and are highly non-Gaussian, non-stationary, and have a non-linear nature [32]."
"Our results indicate that established time-series classification approaches with subject-conditional training can outperform dedicated state-of-the-art EEG classification models for the task of learning one classification model for all subjects."