Ongoing brain activity can be recorded as electroencephalograph (EEG) to discover the links between emotional states and brain activity. This study applied machine-learning algorithms to categorize EEG dynamics according to subject self-reported emotional states during music listening. A framework was proposed to optimize EEG-based emotion recognition by systematically 1) seeking emotion-specific EEG features and 2) exploring the efficacy of the classifiers. Support vector machine was employed to classify four emotional states (joy, anger, sadness, and pleasure) and obtained an averaged classification accuracy of 82.29% +/- 3.06% across 26 subjects. Further, this study identified 30 subject-independent features that were most relevant to emotional processing across subjects and explored the feasibility of using fewer electrodes to characterize the EEG dynamics during music listening. The identified features were primarily derived from electrodes placed near the frontal and the parietal lobes, consistent with many of the findings in the literature. This study might lead to a practical system for noninvasive assessment of the emotional states in practical or clinical applications.

Inter- and intra-subject variability pose a major challenge to decoding human brain activity in brain-computer interfaces (BCIs) based on non-invasive electroencephalogram (EEG). Conventionally, a time-consuming and laborious training procedure is performed on each new user to collect sufficient individualized data, hindering the applications of BCIs on monitoring brain states (e.g. drowsiness) in real-world settings. This study proposes applying hierarchical clustering to assess the inter- and intra-subject variability within a large-scale dataset of EEG collected in a simulated driving task, and validates the feasibility of transferring EEG-based drowsiness-detection models across subjects. A subject-transfer framework is thus developed for detecting drowsiness based on a large-scale model pool from other subjects and a small amount of alert baseline calibration data from a new user. The model pool ensures the availability of positive model transferring, whereas the alert baseline data serve as a selector of decoding models in the pool. Compared with the conventional within-subject approach, the proposed framework remarkably reduced the required calibration time for a new user by 90% (18.00 min-1.72 ± 0.36 min) without compromising performance (p = 0.0910) when sufficient existing data are available. These findings suggest a practical pathway toward plug-and-play drowsiness detection and can ignite numerous real-world BCI applications.

No abstract

To overcome the individual differences, an accurate electroencephalogram (EEG)-based emotion-classification system requires a considerable amount of ecological calibration data for each individual, which is labor-intensive and time-consuming. Transfer learning (TL) has drawn increasing attention in the field of EEG signal mining in recent years. The TL leverages existing data collected from other people to build a model for a new individual with little calibration data. However, brute-force transfer to an individual (i.e., blindly leveraged the labeled data from others) may lead to a negative transfer that degrades performance rather than improving it. This study thus proposed a conditional TL (cTL) framework to facilitate a positive transfer (improving subject-specific performance without increasing the labeled data) for each individual. The cTL first assesses an individual’s transferability for positive transfer and then selectively leverages the data from others with comparable feature spaces. The empirical results showed that among 26 individuals, the proposed cTL framework identified 16 and 14 transferable individuals who could benefit from the data from others for emotion valence and arousal classification, respectively. These transferable individuals could then leverage the data from 18 and 12 individuals who had similar EEG signatures to attain maximal TL improvements in valence- and arousal-classification accuracy. The cTL improved the overall classification performance of 26 individuals by ~15% for valence categorization and ~12% for arousal counterpart, as compared to their default performance based solely on the subject-specific data. This study evidently demonstrated the feasibility of the proposed cTL framework for improving an individual’s default emotion-classification performance given a data repository. The cTL framework may shed light on the development of a robust emotion-classification model using fewer labeled subject-specific data toward a real-life affective brain-computer interface (ABCI).