應用於自適應深腦電刺激回饋控制策略之帕金森氏症局部場電位特徵擷取與生成對抗網路神經訊號轉譯之研究

Abstract

本研究建構於先前已開發之基底核-丘腦（Basal Ganglia-Thalamus, BGT）模擬模型與用於閉環深腦電刺激（Deep Brain Stimulation, DBS）參數控制的強化學習（Reinforcement Learning, RL）框架。該RL控制器基於動作電位（Action Potential, AP）特徵進行學習與決策，然而臨床中所能取得的訊號為局部場電位（Local Field Potential, LFP），兩者間存在明顯的模態不匹配問題。為彌補模擬系統與臨床應用之間的模態落差，本研究提出一套特徵轉譯流程，作為臨床可行性的橋樑。 本研究首先自丘腦下核（Subthalamic Nucleus, STN）LFP訊號中擷取三項具生理意義的特徵：Beta 頻段功率比（Beta Band Power Ratio）、Beta突波持續時間（Beta Burst Duration）與相位-振幅耦合（Phase-Amplitude Coupling），其在區分DBS on與off狀態中表現良好，對應的ROC-AUC分別為 0.764、0.944 與 0.891。 為實現跨模態的特徵對齊，我們提出一套基於Wasserstein生成式對抗網路（Wasserstein GAN, WGAN）之架構，採用單一生成器配合三個判別器（1G3C）設計。三個判別器分別對應健康（Healthy）、帕金森病（PD）、以及 DBS刺激下的生理狀態，用以學習AP特徵的分佈；而泛化的共享生成器則學習將不同狀態下LFP 特徵轉譯至六維的類AP特徵空間，以供RL控制器使用。結果顯示，生成的特徵能夠與對應狀態之模擬AP特徵的分佈對齊，在健康、PD與DBS狀態下，其餘弦相似度（Cosine Similarity）分別達到 0.86、0.86 與 0.65，特別在基準狀態中展現出良好的模態一致性與生理特異性。 綜合上述結果，本研究建立了一套能將臨床可得的LFP特徵有效轉譯為模擬環境中所需的AP特徵的流程，成功銜接臨床訊號與BGT模擬模型間的模態落差。此轉譯機制可作為既有RL控制器的前端模組，提供其可直接操作的輸入特徵，進而提升自適應閉環DBS策略從模擬實驗走向實際臨床驗證的可行性。

This study builds upon a previously developed Basal Ganglia-Thalamus (BGT) simulation model and a reinforcement learning (RL) framework designed for closed-loop deep brain stimulation (DBS) parameter control. While the RL controller is trained on action potential (AP)–level features, real-world clinical recordings typically provide only local field potentials (LFPs), resulting in a significant modality mismatch. To bridge this gap between simulation-based systems and clinical applications, we propose a feature translation pipeline that serves as a translational interface for clinical deployment. We begin by extracting three physiologically meaningful features from subthalamic nucleus (STN) LFP signals—Beta Band Power Ratio, Beta Burst Duration, and Phase-Amplitude Coupling—which effectively distinguish between DBS ON and OFF states, achieving ROC-AUC scores of 0.764, 0.944, and 0.891, respectively. To achieve cross-modal feature alignment, we propose a Wasserstein Generative Adversarial Network (WGAN) architecture, incorporating a single generator and three critics (1G3C). Each critic is trained to model AP feature distributions under a specific physiological state—Healthy, Parkinsonian, or DBS-modulated. The shared generator learns to translate LFP-derived features from various conditions into a six-dimensional AP-like feature space compatible with the RL controller. Experimental results show that the generated features closely match the simulated AP feature distributions, achieving cosine similarity scores of 0.86 (Healthy), 0.86 (PD), and 0.65 (DBS), demonstrating strong modality alignment and physiological specificity, particularly under baseline conditions. In summary, this study establishes a feature translation mechanism capable of converting clinically accessible LFP features into AP-like representations required by the simulation environment. This mechanism effectively bridges the modality gap between clinical recordings and the BGT simulation model. By serving as a front-end module to the existing RL controller, the proposed system facilitates the potential transition of adaptive closed-loop DBS strategies from simulation to clinical validation.