Machine Learning Strategies to Improve Generalization in EEG-based Emotion Assessment: a Systematic Review

Overview

• This paper provides a systematic review of machine learning strategies for improving the generalizability of electroencephalography (EEG)-based emotion classification across different subjects and recording sessions.
• EEG signals are non-stationary, which can lead to the Dataset Shift problem and affect the performance of emotion classification models.
• The review examines 418 papers focused on modern machine learning techniques for improving generalization in EEG-based emotion assessment, with 75 studies selected for detailed analysis.

Plain English Explanation

Electroencephalography (EEG) is a technique that measures the electrical activity in the brain. Researchers have been trying to use EEG to classify and detect human emotions, which could be useful for applications like emotion-based detection of deepfakes, cross-corpus feature alignment, and Parkinson's disease diagnosis. However, a key challenge is that EEG signals can be quite variable and change over time, even for the same person. This "non-stationarity" of EEG data can make it difficult to create emotion classification models that work well across different people and recording sessions.

This review paper looked at 418 studies that used modern machine learning techniques to try to address this problem of poor generalization in EEG-based emotion recognition. After careful screening, the authors selected 75 relevant studies for a detailed analysis. They found that transfer learning methods, which leverage knowledge from one task or dataset to improve performance on a related task or dataset, seem to perform better than other approaches for improving cross-subject and cross-session generalization in EEG emotion classification.

Technical Explanation

The authors conducted a systematic review to examine the use of machine learning strategies for improving the generalizability of EEG-based emotion classification across different subjects and recording sessions. EEG signals are known to be non-stationary, which can lead to the Dataset Shift problem and negatively impact the performance of emotion classification models.

The researchers searched the Scopus, IEEE Xplore, and PubMed databases for relevant papers, ultimately retrieving 418 studies focused on modern machine learning techniques for enhancing generalization in EEG-based emotion assessment. After applying eligibility criteria (e.g., requiring specific cross-subject and cross-session validation strategies, excluding studies using other biosignals), 75 papers were selected for in-depth analysis.

Based on the review of these 75 studies, the authors proposed a taxonomy of the machine learning methods employed, including transfer learning, domain adaptation, and meta-learning approaches. They identified the studies reporting the best average classification accuracies, finding that transfer learning methods generally outperformed other techniques for improving cross-subject and cross-session generalization in EEG emotion recognition.

The review also discusses the potential impact of the emotion theory models and psychological screening of the experimental samples on the performance of the emotion classification models. This suggests that careful consideration of the underlying emotional constructs and the characteristics of the study participants may be important factors in developing robust and generalizable EEG-based emotion recognition systems.

Critical Analysis

The systematic review provides a comprehensive overview of the current state of research on machine learning strategies for enhancing the generalizability of EEG-based emotion classification. The authors' focus on cross-subject and cross-session validation is particularly relevant, as it addresses a key challenge in translating these techniques to real-world applications.

While the review highlights the potential benefits of transfer learning methods, it would be valuable to understand more about the specific transfer learning approaches that have been most effective. Additionally, the paper does not delve deeply into the limitations or potential shortcomings of the reviewed studies, such as the reliance on relatively small or homogeneous datasets, which could constrain the generalizability of the findings.

Further research is needed to explore the impact of factors like emotion theory models and participant characteristics in greater detail, as suggested by the authors. Incorporating time-series analysis and contrastive learning techniques may also help advance the field of EEG-based emotion recognition.

Conclusion

This systematic review provides a valuable synthesis of the current research on machine learning strategies for improving the generalizability of EEG-based emotion classification across subjects and recording sessions. The findings suggest that transfer learning approaches may be particularly promising for addressing the non-stationarity of EEG signals and the resulting Dataset Shift problem.

By highlighting the importance of considering emotion theory models and participant characteristics, the review also points to the need for a more holistic approach to developing robust and generalizable EEG-based emotion recognition systems. Further research in this area could have significant implications for a wide range of applications, from emotion-based deepfake detection to Parkinson's disease diagnosis and beyond.