| 摘要: | 帕金森氏症（Parkinson′s disease，簡稱PD）是一種影響中樞神經系統的慢性神經退化性疾病，肇因於黑質緻密部的多巴胺神經元大量死亡，導致和多巴胺神經元聯繫的基底神經節區域與其支配之功能出現異常，主要影響運動神經系統，包括早期最明顯的症狀為靜止性震顫、肢體僵硬、運動功能減退和步態障礙，也可能有認知和行為問題，失智症在病情嚴重的患者中也相當常見。在神經科學領域中，對於PD患者在運動系統中皮質之間的有效性連結仍然缺乏，我們想知道的是患者在執行自主運動任務的期間，不同區域之間在不同頻率之間發生了哪些變異？另外，由於PD患者早期的症狀較不明顯，這也增加了診斷上的困難。因此，在本研究中我們通過EEG紀錄了正常受試者和PD患者在執行自定義節奏的自主運動期間的大腦皮質活動。我們將利用動態因果模型（DCM），探討PD患者與正常人在運動時誘發響應的有效性連結，釐清PD患者運動皮質區連結機制的改變，並應用深度學習的方法建立出診斷系統，並期望能提供臨床上病症的輔助診斷以及治療。
本研究的研究對象為83位帕金森氏症患者與19位正常受試者，重複進行自行數6秒做手腕伸展（Wrist extension）的動作，並收集運動時的EEG訊號。由於運動輔助區（SMA）的功能與運動的計畫有關，包括處理運動前的想像以及引導動作。與前運動皮質區（PM）不同，SMA主要參與自身產生和控制的運動，而不是在外界刺激下所產生的運動。首先，我們將對於運動相關區域並針對alpha與beta頻帶進行ERD/ERS分析，之後利用誘發響應的動態因果模型（DCM）分析患者與正常人由輔助運動區（SMA）、雙側前運動皮質區（PM）和雙側初級運動皮質區（M1）組成的皮質網路內振盪耦合的變化。最後應用Convolutional Neural Network（CNN）、Long Short-Term Memory（LSTM）以及Multilayer Perceptron（MLP）對於PD患者與正常人的分類問題進行預測。
研究結果顯示，在PD患者執行自主運動期間，MRCP中BP與MP的功率都明顯低於正常受試者，反映了SMA和M1的活化不足，說明PD患者由基底神經節內部迴路所產生的運動程序受到損害，而導致運動準備的困難以及後續執行運動時的障礙。然而在ERD/ERS的分析結果當中卻發現，雖然在我們的PD患者中ERD的起始時間發生延遲，可能證明PD患者對於運動程序的啟動會受到影響，從而部分的解釋了運動障礙，但在PD患者中的alpha和beta ERD強度都較正常受試者高，ERD的明顯增加可能反映了運動皮質活動的相對提升，說明PD患者試圖補償基底神經節輸入的減少，而廣泛的招募其他區域並更強力的活化運動皮質以成功的執行運動任務。在時頻分析的結果中發現，SMA與對側M1在運動執行期間都有明顯增強的震盪活動，特別是在low-beta（13～20Hz）而不是high-beta（21～30Hz），有證據表明SMA與M1之間過度增強的low-beta震盪可能是由於慢性多巴胺耗竭後而產生的神經可塑性變化；對側PM與雙側M1在運動開始前到運動過程中都有明顯增強的高頻（>20Hz）震盪活動，可能是一項神經補償行為；SMA與同側PM和M1在運動過程中都有明顯減弱的low-theta（4～6Hz）震盪活動，可能說明PD患者的認知控制能力受到損害。另外我們也觀察到alpha和beta頻帶彼此之間存在相干性，說明這兩個頻帶在功能上的相互作用可能參與PD疾病的發展。從有效性連結的分析結果中我們表明SMA的確與PD患者的異常密切相關，整個運動迴路中由於Bottom-up的連結產生異常，再加上投射到SMA的內部迴路（BG）受到損害，導致Top-down的連結出現異常，無法將運動訊息有效的下傳至脊髓，最終造成運動的障礙。最後，我們試圖應用深度學習演算法的架構，分析PD患者的異常特徵，並建立出一套可靠且穩定的診斷系統。;Parkinson′s disease (PD) is a progressive neurodegenerative disease that affects the central nervous system (CNS). It is caused by the death of a large number of dopaminergic neurons in substantia nigra pars compacta (SNpc), resulting in abnormalities in the basal ganglia and associated motor circuitry components. The classical motor symptoms include resting tremors, bradykinesia, rigidity and postural instability, as well as possible cognitive impairments and dementia. In the field of neuroscience, the effective connectivity of the motor system in PD patients has not been well clarified as yet, and we would like to know what alteration occurs between different areas and frequencies during the voluntary movements. In addition, the early symptoms of PD patients are less obvious, which makes the diagnosis more difficult. In this study, we recorded cortical activity by electroencephalography (EEG) in normal subjects and PD patients during the self-paced palm-lifting task. We will use dynamic causal modeling (DCM) of induced responses to investigate the altered mechanisms of the interaction within motor cortical networks in PD patients, and then develop a diagnostic system by applying deep learning algorithms.
In this study, 83 patients with Parkinson′s disease and 19 normal subjects were recruited. During the self-paced palm-lifting task, EEG was recorded with 26 electrodes according to the international 10-20 system for EEG recording and the surface EMG signals from the extensor digitorum communis were recorded simultaneously. Each participant was instructed to lift the right palm approximately once every 6 seconds. The event-related (de)synchronizations (ERD/ERS) of alpha and beta-band activity were evaluated, and then we apply a DCM model of induced response to infer effective connectivity within a distributed network of 5 brain regions consisting of the supplementary motor area (SMA), bilateral premotor cortical areas (PM) and bilateral primary motor cortical areas (M1). In addition, several deep learning algorithums including convolutional neural network (CNN), long short-term memory (LSTM) and multilayer perceptron (MLP) were applied for EEG classification.
The results of this study showed that the power of Bereitschaftspotential (BP) and motor potential (MP) in MRCP was significantly diminished in PD patients during voluntary movement, reflecting insufficient activation of SMA and M1, suggesting that the motor program generated by the internal circuitry of the basal ganglia is impaired in PD patients, resulting in difficulties in motor preparation and the impairment of the postsynaptic potentials on the pyramidal cell apical dendrites. The analysis of ERD/ERS revealed that the delayed onset of ERD in our PD patients may suggest that the initiation of the voluntary movement is affected in PD patients, thus partially explaining bradykinesia. However, the increased ERD was found in PD patients, and this phenomenon indicated the relatively increase activation in motor cortical activity, suggesting that PD patients attempt to compensate for the reduced basal ganglia input by extensively recruiting other regions and more strongly activating motor cortices to successfully perform motor tasks. After analyzing the spectral density of induced response between this two groups, we found that both SMA and contralateral M1 showed significantly enhanced oscillatory activity during motor execution, especially in the low-beta (13-20 Hz) but not high-beta (21-30 Hz), indicating that the exaggeratedly enhanced low-beta oscillation might be a compensatory mechanism in PD patients but not the motor impairment. The significantly enhanced high-frequency (>20Hz) oscillation in both contralateral PM and bilateral M1 before movement onset may be a neurocompensatory behavior. The significantly reduced low-theta (4-6Hz) oscillation in both SMA and ipsilateral PM and M1 during motor execution may indicate impaired cognitive control in PD patients. In addition, we also observed a coherence between the alpha and beta band, suggesting that the functional interactions between these two bands may be involved in the development of PD. In our results of the effective connectivity, we showed that the SMA is significantly involved in the abnormalities in PD patients. The entire motor circuit due to the bottom-up connections with the reduced inputs to SMA, leading to abnormalities in the top-down connections, which cannot effectively transmit motor information down to the spinal cord and eventually cause motor impairments. Finally, we tried to apply deep learning frameworks to analyze the abnormal features of PD patients and establish a reliable and stable diagnostic system. |
