基於機器學習與休息腦波訊號之認知評估研究

Abstract

隨著全球高齡化人口的比例逐漸上升，許多國家已經進入高齡化社會。失智症（Dementia），這種由腦部異常退化所引起且常見於老年人的症候群，已成為社會中的重大議題。阿茲海默症（Alzheimer’s Disease, AD）佔據所有失智症成因的七成左右，其為神經退化性疾病的一種。輕度認知障礙（Mild Cognitive Impairment, MCI）通常被視為阿茲海默症的前兆，處於正常衰老與失智症之間的過渡期。如果能夠早期檢測，將有助於進行即時干預，不僅能改善患者的生活品質，還可能有效延緩病情的進展。本研究從桃園社區收集了40位受測者，基於樣本數量與共病情況的考量，我們使用臺灣版的蒙特利爾認知評估量表（Montreal Cognitive Assessment, MoCA）將其分為認知功能障礙組（20位，Cognitive Impairment, CI）與無認知功能障礙組（20位，Cognitive Normal, CN）。本研究將使用受測者在休息狀態下所收集的腦電圖（Electroencephalography, EEG）數據，訓練一個有效區分認知功能障礙組與無認知功能障礙組的檢測系統。 本研究從腦電圖中擷取五種特徵，分別是Katz碎形維度（Katz Fractal Dimension, KFD）、Higuchi碎形維度（Higuchi Fractal Dimension, HFD）、樣本熵（Sample Entropy, SE）、相對功率（Relative Power, RP）與相位延遲指數（Phase Lag Index, PLI）。將此五種特徵利用費雪準則進行排序，支持向量機（Support Vector Machine, SVM）調參過程所使用之驗證方法為留一個體驗證（Leave One Participant Out Cross Validation, LOPO-CV），接下來使用支持向量機搭配加一特徵法（Add One Feature In, AOFI）篩選出最好的特徵集。Katz碎形維度、Higuchi碎形維度與樣本熵表現最差，其平均分類率落於六成到七成之間。相對功率為表現第二好之特徵，其平均分類率落於七成到八成之間。表現最好之特徵為相位延遲指數，其最好的平均分類率達到88.88%。接著，將各種特徵表現最好的特徵集作為輸入，並將其餵入支持向量回歸（Support Vector Regression, SVR），以建立與臺灣版蒙特利爾認知評估量表分數之間的回歸模型。所訓練之回歸模型可能因樣本數不足或數據之間不夠平均，導致其表現不慎理想。總結來說，本研究結果顯示，利用基於休息狀態腦電圖的相位延遲指數特徵，對於區分認知功能障礙組與無認知功能障礙組的檢測效果相當理想；然而，回歸模型仍需進一步的探討與改進。

As the global population continues to age, many countries have entered aging societies. Dementia, a syndrome caused by abnormal changes in the brain and commonly observed in the older adults, has become a major public health issue. Alzheimer's Disease (AD), a neurodegenerative condition, is responsible for around 70% of dementia diagnoses. Mild Cognitive Impairment (MCI) is viewed as an early stage of Alzheimer's Disease (AD), serving as a bridge between typical aging and dementia. Early detection of MCI can enable timely intervention and treatment, which not only improves the patient’s quality of life but may also effectively delay disease progression. This study recruited 40 participants from communities in the Taoyuan area. Considering the sample size and potential comorbidities, participants were categorized using the Taiwanese version of the Montreal Cognitive Assessment (MoCA) into two groups: 20 individuals with cognitive impairment (Cognitive Impairment, CI) and 20 individuals without cognitive impairment (Cognitive Normal, CN). This study utilizes Electroencephalography (EEG) data collected from participants in a resting state to train a detection system capable of effectively distinguishing between the CI group and the CN group. Five types of EEG features were extracted in this study: Katz Fractal Dimension (KFD), Higuchi Fractal Dimension (HFD), Sample Entropy (SE), Relative Power (RP), and Phase Lag Index (PLI). These features were ranked using the Fisher Criterion. The Support Vector Machine (SVM) model was validated using the Leave One Participant Out Cross Validation (LOPO-CV) approach. Subsequently, the Add One Feature In (AOFI) method was applied in conjunction with SVM to identify the best feature subset. Among the features, KFD, HFD, and SE yielded the lowest classification performance, with average classification accuracies ranging from 60% to 70%. RP ranked second in performance, with accuracies between 70% and 80%. The best-performing feature was the PLI, achieving a maximum average classification accuracy of 88.88%. Then, the best feature subsets of each feature type were used as inputs to the Support Vector Regression (SVR) model to explore their regression relationships with the Taiwan version of the MoCA scores. However, due to the small sample size and potential data imbalance, the regression model’s performance was not ideal. In summary, this study demonstrates that PLI features derived from resting-state EEG data are effective in distinguishing between cognitively impaired and healthy individuals. Nevertheless, the regression model requires further investigation to improve its reliability.