廣泛性焦慮症（Generalized Anxiety Disorder， GAD）是一種常見的情緒障礙，患者常伴隨情緒調節與認知控制功能缺損。過去研究多聚焦於情緒性 Stroop 作業，而事件相關電位（Event-Related Potentials， ERP）研究指出，GAD 患者於認知衝突歷程中，早期感覺處理相關的 如N100、N200 成分常呈現振幅增強或異常，但研究結果仍存在不一致；P300 振幅多呈下降，反映注意力資源配置及工作記憶更新能力受損；N450 則常呈現振幅增加，顯示衝突監測及認知控制歷程異常。此外，時頻分析研究亦指出，Theta 振盪多呈現活化不足或延遲、Alpha 去同步化異常、Beta 功率增加，以及 Gamma 神經振盪異常增強，顯示 GAD 患者存在認知控制及神經網路同步化失衡。然而，目前對於非情感性認知衝突之神經動力學機制及腦區間有效連結（Effective Connectivity）的研究仍相當有限，尤其缺乏整合 ERP、皮質源空間時頻分析、動態因果模型（Dynamic Causal Modeling， DCM）與深度學習之研究。因此，本研究採用傳統色字 Stroop 作業（Stroop Task），結合腦電圖（Electroencephalography， EEG）、ERP、皮質源空間時頻分析、DCM 與深度學習，系統性探討 GAD 患者認知控制網路之神經動力學特徵及其臨床診斷潛力。 本研究共招募 48 名受試者，包括健康對照組（HC）25 名及 GAD 患者 23 名。行為學結果顯示，兩組皆呈現典型的速度與正確率權衡現象，而 GAD 患者反應時間整體較慢。ERP 分析發現，GAD 患者於 P100、N100、P200、P300 與 N450 成分波幅皆顯著高於 HC 組，且 HC 組僅於 N450 呈現條件特異性差異，GAD 患者則未出現一致與不一致條件差異，顯示其皮質持續處於高覺醒狀態。 DCM 分析建立三種認知控制網路模型，並透過貝葉斯模型選擇（Bayesian Model Selection， BMS）比較模型適配性。結果顯示，以雙側額眶葉（LFO、RFO）、背外側前額葉皮質（DLPFC）、前扣帶回（ACC）、雙側內側上頂葉（LMSPC、RMSPC）、右後頂葉皮質（RPC）、雙側外側枕葉（LOC、ROC）及顳頂交界處（TPJ）所建構之 Model 3 為最佳模型。進一步分析顯示，GAD 患者於前扣帶回（ACC）與頂葉認知控制網路之間存在異常的有效連結調控，高階回饋路徑亦由抑制性轉為興奮性調節，反映認知控制網路神經訊息傳遞失衡。 皮質源空間時頻分析進一步顯示，GAD 患者於刺激前即呈現預期性緊繃，Theta 與 Alpha 頻帶於晚期出現延遲活化，而 Beta 與 Gamma 頻帶則呈現相反極性的動態變化，反映前後腦網路通訊失調與訊息整合異常。最後，本研究整合 DCM 有效連結與皮質源時頻特徵作為深度學習模型輸入，分類準確率達 84%，顯示本研究所提出之神經生物標記具有 GAD 客觀輔助診斷之潛力。;Generalized Anxiety Disorder (GAD) is a common emotional disorder frequently accompanied by deficits in emotional regulation and cognitive control. While previous research has predominantly focused on emotional Stroop tasks， Event-Related Potential (ERP) studies have indicated that GAD patients exhibit altered early sensory processing—such as enhanced or abnormal amplitudes of the N100 and N200 components during cognitive conflict processing—though these empirical findings remain inconsistent. Additionally， reduced P300 amplitudes are commonly observed， reflecting impaired attentional resource allocation and working memory updating， whereas increased N450 amplitudes suggest dysregulated conflict monitoring and cognitive control processes. Time-frequency analyses further demonstrate that GAD patients exhibit hypoactivation or delayed responses in Theta oscillations， atypical Alpha desynchronization， elevated Beta power， and abnormally enhanced Gamma oscillations， collectively highlighting an imbalance in cognitive control and neural network synchronization. However， the neural dynamic mechanisms and inter-regional effective connectivity underlying non-emotional cognitive conflict remain under-researched， particularly regarding integrated approaches combining ERP， cortical source-space time-frequency analysis， Dynamic Causal Modeling (DCM)， and deep learning. Therefore， this study employed a classic color-word Stroop task coupled with Electroencephalography (EEG)， ERPs， cortical source-space time-frequency analysis， DCM， and deep learning to systematically investigate the neurodynamic characteristics of the cognitive control network in GAD patients and evaluate their potential for clinical diagnosis. A total of 48 participants were recruited， comprising 25 Healthy Controls (HC) and 23 GAD patients. Behavioral results demonstrated a classic speed-accuracy trade-off across both groups， with the GAD group exhibiting overall slower reaction times. ERP analysis revealed that GAD patients displayed significantly higher amplitudes than the HC group across the P100， N100， P200， P300， and N450 components. Furthermore， while the HC group exhibited condition-specific differences exclusively in the N450 component， the GAD group showed no significant differences between congruent and incongruent conditions， indicating a state of persistent cortical hyperarousal. DCM analysis established three cognitive control network architecture models， with model fit evaluated via Bayesian Model Selection (BMS). Results identified Model 3—constructed using the bilateral frontoorbital cortices (LFO， RFO)， dorsolateral prefrontal cortex (DLPFC)， anterior cingulate cortex (ACC)， bilateral medial superior parietal cortices (LMSPC， RMSPC)， right parietal cortex (RPC)， bilateral lateral occipital cortices (LOC， ROC)， and temporoparietal junction (TPJ)—as the optimal model. Further effective connectivity analysis demonstrated abnormal modulation within the fronto-parietal cognitive control network involving the ACC in GAD patients， alongside a functional shift in high-level feedback pathways from inhibitory to excitatory regulation， reflecting disrupted neural information transmission. Cortical source-space time-frequency analysis further revealed that GAD patients exhibited anticipatory tension prior to stimulus onset， delayed late-stage activation in the Theta and Alpha bands， and opposite-polarity dynamic shifts in the Beta and Gamma bands， suggesting fronto-occipital network communication breakdown and impaired information integration. Finally， integrating DCM effective connectivity parameters with cortical source-space time-frequency features as inputs into a deep learning architecture yielded a classification accuracy of 84%， highlighting the clinical utility of these neurobiological markers as potential objective diagnostic tools for GAD.
