| 摘要: | 失智症會對大腦的不同區域造成損傷，進而影響患者的日常生活功能。失智症好發於老年族群，而隨著全球人口老化持續加劇，預估未來失智症患者人數將大幅增加，因此，針對失智症的早期篩檢與預防策略，已成為當今高齡社會中極需重視的議題。在傳統診斷中，醫院通常透過蒙特利爾認知評估 (MoCA)、簡短智能測驗 (MMSE) 等工具來量化患者的認知能力，以及使用磁振造影(MRI)或電腦斷層掃描(CT)來檢視腦部結構變化，以輔助診斷失智症的類型與嚴重程度。相較之下，腦電圖(EEG)作為一種非侵入性、成本相對低廉且時間解析度高的工具，能夠即時反映大腦神經活動狀態，對於早期偵測認知功能異常具有潛力。因此，若能利用腦電波 (EEG) 預測失智症患者的認知分數，將可以提供更客觀、更迅速且量化的結果。
本研究量測 102 失智症患者(正常:32 位、輕度認知障礙: 39 位、中度認知障礙:22 位與重度認知障礙:9 位)靜息狀態腦電訊號，並使用改良式圖注意力網路(Graph Attention Network v2, GATv2)對正常/輕度認知障礙與中重度失智症患者進行分類，以及正常/輕度認知障礙患者的 MoCA 分數回歸預測任務。研究結果顯示，於正常/輕度認知障礙與中重度失智症患者之測試分類的準確率達 83%；對正常/輕度失智症患者之 eMoCA 回歸分數預測結果中，平均絕對誤差(MAE±STD)為 3.37 ±1.74(r = 0.54，p < 0.01)。本次研究結合多頻帶功能連結性與深度學習，分析多通道 EEG 圖結構資料，達成對不同認知階段的辨識與回歸預測。;Dementia causes damage to different regions of the brain, thereby affecting patients’ daily functional abilities. It is more prevalent among the elderly population, and with the continued trend of global population aging, the number of individuals affected by dementia is projected to increase significantly. Therefore, early screening and prevention strategies for dementia have become critical issues in today’s aging society. In traditional clinical diagnosis, cognitive function is typically assessed using tools such as the Montreal Cognitive Assessment (MoCA) and the Mini-Mental State Examination (MMSE), along with brain imaging techniques like magnetic resonance imaging (MRI) or computed tomography (CT) to examine structural brain changes and assist in determining the type and severity of dementia. In contrast, electroencephalography (EEG), being non-invasive, relatively low-cost, and having high temporal resolution, can reflect neural activity in real time and holds great potential for the early detection of cognitive impairment. If EEG signals can be used to predict cognitive scores in dementia patients, it would provide a more objective, rapid, and quantifiable method for clinical support.
In this study, resting-state EEG was recorded from 102 dementia patients, including 32 with normal cognition, 39 with mild cognitive impairment (MCI), 22 with moderate impairment, and 9 with severe impairment. A modified Graph Attention Network (GATv2) was applied to perform classification between normal/mild and moderate/severe dementia groups, as well as MoCA score regression for patients with normal to mild impairment. The results showed that the classification accuracy on the test set reached 83% between normal/mild and moderate/severe dementia groups. For MoCA score regression among patients with normal to mild impairment, the model achieved a mean absolute error (MAE) of 3.37 ±1.74(r = 0.54，p < 0.01). This study integrates multi-band functional connectivity and deep learning to analyze multichannel EEG graph-structured data, achieving effective classification and regression prediction across different stages of cognitive impairment. |
