| 摘要: | 持續性心房顫動是全球最常見的心律不整之一，亦為導致心衰竭、中風與死亡率增加的主要病因。其病理機轉牽涉旋轉波、發散波前與心房組織電生理特性之空間異質性，導致現有驅動源定位與波動重建技術，在穩定性、解析度與臨床適用性方面皆面臨挑戰。
本研究首先針對既有轉子定位與向量場重建方法進行評估與模擬驗證。比較現有的心房顫動序列偵測演算法在不同導管設計下之偵測限制，並透過模擬轉子的旋轉波與擴散模擬，建立兩項幾何指標：「轉子指數」與「擴散源指數」。研究證實，這些指標能有效量化波前動態特性，並適用於不同幾何構型的導管，提供導管設計的幾何建議與準確識別驅動源的可行依據。
接著進一步導入過往臨床電燒的成功經驗於人工智能模型，建立以深度學習預測心房顫動源的決策系統DeePRISM。在前瞻性研究中，DeePRISM導引消融可顯著提升長期無心律不整發生率；在模擬的轉子自放電模型中亦顯示，DeePRISM模型所預測的心房顫動區域與轉子的核心區域高度吻合，顯示深度學習模型於電生理導引治療中的潛力。
最後我們提出一項創新技術：「主導方向相干性向量(Principal Directed Coherence Vector, PDCV)分析法」，此方法結合時序因果性與空間向量場重建，建構可視化的心房波前傳導地圖，並揭示驅動源與其下游區域之因果流向。與傳統Granger因果分析相比，PDCV不僅適用於非同步、低密度導管資料，更可於旋轉與放射波場中穩定辨識驅動源核心。臨床驗證亦顯示，心房顫動患者的PDCV呈現肺靜脈具「外傳導」特性，則其接受肺靜脈隔離術後具較高心房顫動終止率與長期穩定性，該特徵亦為復發風險的獨立預測因子。
綜合而言，本研究從過往的心房顫動序列分析方法出發，結合深度學習模型與因果場重建技術，建立具多層次分析能力的心房顫動驅動源定位架構。此方法不僅可作為個人化消融策略的輔助依據，亦展現優於現行商用系統之穩定性與靈敏度，未來可望應用於臨床電生理導航、基質修飾策略優化與心律不整機轉研究等多元場域。
;Persistent atrial fibrillation (AF) is one of the most common arrhythmias worldwide and a leading cause of heart failure, stroke, and increased mortality. Its underlying mechanisms involve complex interactions among rotors, divergent wavefronts, and spatial heterogeneity of atrial tissue electrophysiology. These factors pose significant challenges to current rotor localization and wavefront reconstruction techniques in terms of stability, spatial resolution, and clinical applicability.
This study first evaluates and simulates existing rotor detection and vector field reconstruction methods. By comparing common AF sequence-based detection algorithms under different catheter configurations, we identified critical limitations in accuracy and robustness. To address these gaps, we developed two novel geometric indices the Rotor Index (RI) and the Focal Index (FI) using simulated rotor and divergent source models. These indices effectively quantify wavefront dynamics and provide practical guidance for catheter design and rotor localization across varying electrode geometries.
Building upon prior clinical ablation experiences, we then developed a deep learning-based decision support system, DeePRISM, designed to predict atrial fibrillation driver. In a prospective clinical study, DeePRISM-guided ablation significantly improved long-term arrhythmia-free survival. Moreover, in simulated spontaneous rotor models, the regions predicted by DeePRISM closely matched the rotor core, highlighting its potential as an intelligent tool for electrophysiological guidance.
Finally, we proposed a novel analysis method, the Principal Directed Coherence Vector (PDCV) approach, which integrates temporal causality with spatial vector field reconstruction to visualize the directionality and influence of atrial wavefront propagation. Compared to traditional Granger causality, PDCV accommodates asynchronous and low-density catheter recordings and demonstrates superior stability in identifying rotor cores in both rotational and focal wavefront fields. Clinical validation further showed that patients exhibiting “outward propagation” from the pulmonary veins in PDCV maps experienced higher termination rates and greater long-term rhythm stability after pulmonary vein isolation, with this feature serving as an independent predictor of AF recurrence.
In conclusion, this study combines traditional AF signal analysis with deep learning and causal vector field techniques to establish a multi-layered framework for rotor localization and personalized ablation planning. The proposed methods not only outperform existing commercial systems in stability and sensitivity but also hold promise for future applications in clinical EP navigation, substrate modification strategies, and arrhythmia mechanism research. |
