Swin Transformer Based Depression Detection Model Learning Only Single Channel EEG Signal

• Item type: Symposium(1)
• 公開日: 2023-12-20
• タイトル: Swin Transformer Based Depression Detection Model Learning Only Single Channel EEG Signal
• タイトル / 言語: en
• タイトル: Swin Transformer Based Depression Detection Model Learning Only Single Channel EEG Signal
• 言語: eng
• 資源タイプ / 資源タイプ識別子: http://purl.org/coar/resource_type/c_5794
• 資源タイプ: conference paper
• 著者所属: Shibaura Institute of Technology
• 著者所属: Shibaura Institute of Technology
• 著者所属(英): en
• 著者所属(英): Shibaura Institute of Technology
• 著者所属(英): en
• 著者所属(英): Shibaura Institute of Technology
• 著者名: Kei, Suzuki; Midori, Sugaya
• 著者名(英): Kei, Suzuki; Midori, Sugaya
• 論文抄録 / 内容記述タイプ: Other
• 論文抄録 / 内容記述: We propose a method for constructing a highly accurate depression detection model that learns only a single channel EEG signal. In recent years, in order to support diagnosis of depression, objective evaluation methods for depression have been studied. One method is based on EEG signals, which are obtained by placing channels on the scalp. In this method, multiple channels are placed on the scalp, and information from different parts of the head is trained by machine learning model to achieve highly accurate depression detection. Therefore, the depression detection methods often assume multiple channels EEG signals. However, this method has a practical problem: using multiple channels may cause a lot of fatigue for the EEG device wearer and multiple channels EEG device could be more expensive. Therefore, in order to solve these issues, purpose of this study is to realize highly accurate depression detection by analyzing only a single channel EEG signals. Then, we propose construction model method to achieve this purpose. In order to evaluate proposed construction model method, we used a public dataset containing about 60 depressed patients and healthy controls EEG signals. The EEG signals were transformed into a time series of power variation in a constant frequency band by continuous wavelet transform. The power variations were transformed into images. The images were used to fine-tune the Swin Transformer, which has been shown to be highly accurate in image classification tasks. The accuracy of this model was evaluated using Stratified Group K Fold (K=5). As a result, more than 90% was achieved in ROC-AUC metric for the binary classification of depressed patients and healthy controls. This result suggests the validity of the method used in this study. We consider the results shows that the research could make the depression detection method more practical than multiple channels methods.
• 論文抄録(英) / 内容記述タイプ: Other
• 論文抄録(英) / 内容記述: We propose a method for constructing a highly accurate depression detection model that learns only a single channel EEG signal. In recent years, in order to support diagnosis of depression, objective evaluation methods for depression have been studied. One method is based on EEG signals, which are obtained by placing channels on the scalp. In this method, multiple channels are placed on the scalp, and information from different parts of the head is trained by machine learning model to achieve highly accurate depression detection. Therefore, the depression detection methods often assume multiple channels EEG signals. However, this method has a practical problem: using multiple channels may cause a lot of fatigue for the EEG device wearer and multiple channels EEG device could be more expensive. Therefore, in order to solve these issues, purpose of this study is to realize highly accurate depression detection by analyzing only a single channel EEG signals. Then, we propose construction model method to achieve this purpose. In order to evaluate proposed construction model method, we used a public dataset containing about 60 depressed patients and healthy controls EEG signals. The EEG signals were transformed into a time series of power variation in a constant frequency band by continuous wavelet transform. The power variations were transformed into images. The images were used to fine-tune the Swin Transformer, which has been shown to be highly accurate in image classification tasks. The accuracy of this model was evaluated using Stratified Group K Fold (K=5). As a result, more than 90% was achieved in ROC-AUC metric for the binary classification of depressed patients and healthy controls. This result suggests the validity of the method used in this study. We consider the results shows that the research could make the depression detection method more practical than multiple channels methods.
• 書誌情報: Proceedings of Asia Pacific Conference on Robot IoT System Development and Platform; 巻 2023, p. 56-58, 発行日 2023-12-20
• 出版者 / 言語: ja
• 出版者: 情報処理学会