結合多元特徵與預處理策略之腦電訊號認知建模研究

Abstract

本研究旨在探討腦電訊號（EEG）作為認知功能評估之輔助工具的可行性。由社區量測開始，針對數十名50 至90歲受試者蒐集靜息狀態與任務狀態EEG資料，並搭配 MoCA-T 認知測驗進行分析，而後根據 MoCA-T 分數可以將受測者切分為兩個認知群體。本研究不僅進行傳統二元分類任務，也引入迴歸模型預測 MoCA-T 分數，再以此分數進行分類，作為一種新型態的認知評估方法。 本研究建構完整建模流程，使用隨機森林與梯度提升樹（XGBoost）為主要演算法，資料經三種預處理策略（原始訊號、獨立成分分析、重新參考以及獨立成分分析），並提取多種EEG特徵（如時域、頻域、複雜性、連接度、小波、希爾伯特-黃轉換及其組合）進行訓練與預測。結果顯示結合靜息狀態與任務狀態EEG之模有著平均最高的分類準確性，特別是在經過兩種預處理方法後。此外，預處理策略與特徵性質交互影響模型表現，獨立成分分析（ICA）能有效提升任務狀態在回歸模型中的表現，而部分非線性特徵在原始資料中反而保留更多資訊，例如希爾伯特-黃轉換（HHT）相關特徵可達到 80% 以上的分類準確率。 綜上所述，藉由大規模的特徵提取與模型訓練，本研究成功篩選出在靜息態 EEG 中，原始訊號結合非線性特徵（如 Connectivity HHT）可達到最高 92% 的分類準確率。在任務態 EEG 中，原始訊號搭配與連結性相關的特徵表現最佳，分類準確率可達 87% 至 88%。在整合後的模型中，若搭配 ICA 前處理與複雜度特徵，或是結合重參考與 ICA 前處理與連結性特徵，皆可達到 88% 至 89% 的分類準確率。雖然本研究模型預測效能仍有進步空間，但已建立一套可行的分析框架，未來經由資料規模擴充、特徵篩選、模型優化與模型間的集成學習，有潛力應用於失智症早期篩檢，並進一步發展為認知功能評估之輔助工具。

This study aimed to investigate the feasibility of using electroencephalography (EEG) as an auxiliary tool for cognitive function assessment. Beginning with community-based data collection, resting-state and task-state EEG recordings were obtained from dozens of participants aged 50 to 90, alongside administration of the MoCA-T cognitive assessment. Based on the MoCA-T scores, participants were categorized into two cognitive groups. In addition to traditional binary classification tasks, this study further introduced a regression-based approach to predict MoCA-T scores and subsequently classify cognitive status based on the predicted values, offering a novel framework for cognitive evaluation. A comprehensive modeling pipeline was established using Random Forest and XGBoost as the primary algorithms. EEG data underwent three preprocessing strategies: raw signals, Independent Component Analysis (ICA), and re-referencing combined with ICA. A wide range of EEG features were extracted for model training and prediction, including time-domain, frequency-domain, complexity, connectivity, wavelet, Hilbert-Huang Transform (HHT), and their combinations. The results showed that models utilizing integrated EEG features (i.e., normalized differences between task-state and resting-state EEG) achieved the highest average classification accuracy, especially under the two preprocessing methods. Furthermore, the interaction between preprocessing methods and feature types significantly influenced model performance: ICA preprocessing notably improved regression results in task-state EEG, while certain nonlinear features, e.g., HHT-related features, retained more informative content in raw EEG, achieving classification accuracies exceeding 80%. In summary, through large-scale feature extraction and systematic model training, this study identified high-performing combinations across EEG paradigms. In resting-state EEG, raw signals paired with nonlinear features (Connectivity HHT) achieved the highest classification accuracy of 92%. For task-state EEG, raw data combined with connectivity-based features yielded optimal results, reaching 87%–88% accuracy. In integrated EEG models, classification accuracies of 88%–89% were achieved when using complexity features with ICA preprocessing, or connectivity features with both re-referencing and ICA. Although there remains room for improvement in model performance, this study establishes a feasible analytical framework. With future enhancements through expanded datasets, feature optimization, model refinement, and ensemble learning, EEG has the potential to support early detection of dementia and serve as an assistive tool for cognitive function assessment.