| 摘要: | 自閉症譜系障礙（Autism Spectrum Disorder, ASD）是一種神經發展障礙，其主要特徵包括社交溝通困難、重複性行為以及異常的感官反應。面部情緒辨識（Facial Emotion Recognition, FER）是社交互動的關鍵能力，許多 ASD 個體在此方面表現困難，通常與前額葉皮質功能異常有關。然而，過去多數研究採用靜態圖片或影片作為評估工具，缺乏情境與動態線索，可能無法真實反映現實生活中情緒處理的複雜性。為了解決此問題，本研究運用虛擬實境（Virtual Reality, VR）設計沉浸式且具情境脈絡的情緒辨識任務，讓參與者在具備時間連續性的情境中體驗情緒事件。同時，記錄多模態生理訊號（包括腦電圖 EEG、眼動追蹤與頭部運動），以探討 ASD 與典型發展兒童（TD）在神經與行為層面的差異。此外，亦實施 VR 訓練介入，並應用機器學習模型進行分類分析。
本研究共有 30 名被診斷患有自閉症的兒童和 23 名正常兒童參與了這項研究。記錄受試者在進行測驗時的腦電圖訊號、眼動及頭轉訊號，並對EEG在時域及頻域進行分析，並使用雙樣本T檢定進行統計分析。
結果顯示，ASD組與TD組在FER的行為上並沒有顯著差異。然而在生理訊號分析卻有明顯的差異。ASD組在情緒表達處理過程中表現出較低的 EEG 振幅和更快的潛伏期。相對於錯誤答案，TD組對正確選項呈現延遲但精準的神經反應，反映其具備選擇性處理能力與衝突監測系統。在頻域中，ASD組表現出較高的 EEG 振幅，尤其是在 FCz 和 FC3 電極處，表明使用了額外的神經資源進行情緒識別。在接受 VR 訓練介入後，ASD 組於額葉與頂葉區域出現明顯的神經調節改善。通過機器學習分類，正確率也有達到84.21%。進一步將訓練後資料納入模型測試，結果顯示約有半數 ASD 受試者被分類為 TD，反映其神經反應特徵已朝向典型發展趨近，支持訓練介入之有效性，並驗證模型對神經變化的高度敏感性。
綜上所述，本研究透過虛擬實境系統整合情境式任務與多模態生理訊號記錄，能夠直觀且高時效地捕捉 ASD 兒童於情緒處理歷程中的異常反應模式，不僅有助於深化對 ASD 情緒加工特徵的理解，亦突顯出 VR 結合生理訊號分析於 ASD 評估與訓練之潛在應用價值，為後續個別化介入策略的制定與早期診斷提供重要依據與方向。
;Autism Spectrum Disorder (ASD) is a neurodevelopmental condition characterized by impairments in social communication, repetitive behaviors, and atypical sensory responses. Facial Emotion Recognition (FER) is essential for social interaction, and many individuals with ASD exhibit difficulties in this domain, often linked to atypical prefrontal cortical functioning. However, previous studies using static images or video-based assessments may fail to capture the complexity of real-life emotion processing due to their lack of contextual and dynamic cues. To address this, the present study employed Virtual Reality (VR) to design immersive, context-rich FER tasks, allowing participants to experience emotional events with temporal continuity. Simultaneously, multimodal physiological signals—including EEG, eye tracking, and head movement—were recorded to investigate neural and behavioral differences between ASD and typically developing (TD) children. A VR-based training intervention was implemented, and machine learning models were used for classification.
A total of 30 children with ASD and 23 TD children participated. Although no significant group differences were observed in behavioral accuracy, EEG analyses revealed distinct neural patterns. ASD participants showed lower amplitudes and shorter latencies, while TD children exhibited delayed but more selective responses, reflecting mature conflict monitoring. In the frequency domain, ASD children showed increased activity in frontal-central regions, suggesting compensatory activation. Post-training results indicated improved neural modulation in the ASD group, particularly in frontal and parietal areas. CNN models achieved up to 84.21% classification accuracy using ERP features. When post-training data were tested, approximately half of the ASD participants were classified as TD, suggesting training-related normalization of neural responses.
In summary, this study highlights the potential of VR-integrated FER tasks and physiological signal analysis for revealing neural atypicalities in ASD, supporting their application in assessment, intervention, and early diagnosis. |
