Title
题目
Accurate and efficient cardiac digital twin from surface ECGs: Insights intoidentifiability of ventricular conduction system
从体表心电图构建精准高效的心脏数字孪生:心室传导系统可识别性的研究洞察
01
文献速递介绍
心脏数字孪生(CDT)技术旨在构建与个体心脏1:1的计算机仿真模型(Niederer等人,2021)。CDT基于先进的生物物理模拟,并通过持续或周期性的实时观测数据更新(Laubenbacher等人,2024),以精确追踪患者心脏的生理状态。基于其建模技术的机制本质,CDT范式的核心假设是虚拟心脏与实体心脏紧密关联,无论在真实还是虚拟空间中,任何刺激或扰动都会引发相同的应激反应。满足这一假设的CDT为个性化医疗提供了巨大的变革潜力,因其能以安全有效的方式评估患者心脏功能,支持疾病诊断、风险分层和治疗方案的优化规划(Corral-Acero等人,2020;Cluitmans等人,2024)。 大规模构建高保真CDT的理想愿景面临诸多技术挑战。复杂的建模流程需包含两个不同阶段——称为“解剖学孪生”和“功能学孪生”——前者专注于从医学影像构建解剖学精确的心脏模型,后者则致力于校准模型的大量参数,以实现模型与患者心脏电生理功能的匹配。在自动化解剖学孪生以大规模构建患者心脏解剖结构方面,已有显著进展(Crozier等人,2016)。而功能学孪生方法的发展相对滞后且更具挑战性。绝大多数标注为CDT的模拟研究并未涉及功能学孪生,而是对所有患者统一使用平均参数,并且避免与心电图(EGM)或体表心电图(ECG)等临床可观测信号进行逐例对比,因为这类对比会揭示虚拟模型与物理现实之间的显著差异(Sung等人,2022;Bishop和Plank,2025)。最近开发增强型功能学孪生技术的努力显示出突破当前模型校准限制的潜力。关键进展包括引入更高效的生物物理模型评估方法,现可支持器官尺度的心脏电生理(EP)模拟并实现实时性能(Pezzuto等人,2017;Neic等人,2017)。这些方法已用于概念验证研究,通过采样(Pezzuto等人,2021;Gillette等人,2021b;Camps等人,2024;Álvarez Barrientos等人,2025)和优化方法(Grandits等人,2021a,2024),将患者心脏模型校准至非侵入性临床ECG数据。尽管如此,功能学孪生的核心挑战依然存在——如何从有限的临床数据(理想情况下为非侵入性获取)中拟合模型的高维参数空间,以稳健、唯一且可扩展的方式精确复现临床记录。 本研究的主要目标是从患者体表ECG推断心室(心脏主要泵血腔室)的电激动序列,并在电生理波阵面传播的生理学精确模型中复现该序列及相关ECG(特别是QRS波群)(Pezzuto等人,2021;Grandits等人,2021a;Gillette等人,2022;Li等人,2022;Grandits等人,2024;Camps等人,2024)。心室激动序列及其在体表QRS波群中的反映由心室传导系统(即希氏-浦肯野系统,HPS)驱动,该系统启动心室激动。HPS包括通过房室结接收心房信号的希氏束、分为 fascicles 并连接浦肯野网络的左右束支。浦肯野网络是快速传导的纤维网络,遍布心内膜和心内膜下组织(Spach等人,1963;Myerburg等人,1972;Vigmond和Stuyvers,2016)。浦肯野网络在终末连接处(称为PMJ,浦肯野-心肌连接处)与心室心肌耦合,这些连接处是最早心室激动的起始部位,因此在决定心室激动模式和体表ECG方面起着关键作用。然而,HPS的结构及PMJ的分布具有高度个体特异性,且通常无法通过非侵入性方法获取(Pullan等人,2010),即使使用侵入性高级标测方法(Palamara等人,2014),也只能获得有限精度的观测。由于HPS作为控制心室激动的关键结构,以及作为传导障碍治疗的靶点(Sharma等人,2018),揭示其结构和特性的方法是心脏电生理学中最基本的挑战之一。然而,从ECG等有限、稀疏且含噪的数据中推断心室激动序列这样的高维对象,构成了一个不适定问题。因此,仅从标准12导联ECG能在多大程度和精度上推断心室激动序列仍不明确(Grandits等人,2022)。因此,必须考虑不同心室激动图谱可能在12导联ECG中产生相同QRS波群的可能性(Colli Franzone等人,2014;Li等人,2025)。 基于程函方程和反应-程函模型的快速EP模拟器,为解决推断心室激动序列(理想情况下唯一推断)的可识别性挑战提供了途径(Colli Franzone和Guerri,1993;Pezzuto等人,2017;Neic等人,2017)。这些模拟器允许快速探索参数空间,通过比较模拟ECG与记录ECG来识别候选PMJ集合(Gillette等人,2021b;Camps等人,2024)。基于采样的方法已成功从ECG数据中识别出局灶活动(Pezzuto等人,2022;Meisenzahl等人,2024)、束支传导阻滞(Pezzuto等人,2021;Álvarez Barrientos等人,2025)和正常窦性心律(Gillette等人,2021b;Camps等人,2024;Cardone-Noott等人,2016;Barber等人,2021)等条件下的心室激动模式。然而,鉴于所需样本数量庞大,这些方法仅适用于少量参数或与快速仿真器结合使用时(Pezzuto等人,2022;Salvador等人,2023;Camps等人,2024)。在此类研究中,HPS常通过少量PMJ结合快速心内膜层来近似(Pezzuto等人,2021)。但总体而言,基于采样的方法和仿真器随着参数数量增加而扩展性不佳(Larson等人,2019),因此在更一般情况下推断心室传导系统的能力有限。 在本研究中,我们利用Geodesic-BP——一种从ECG数据识别PMJ的快速梯度基方法(Grandits等人,2024)(Geodesic-BP概述见图1)。该方法可在单个高端图形处理单元(GPU)上于30分钟内实现对ECG的高精度拟合。需注意的是,尽管该方法也可在CPU上计算,但运行时间会增加至6小时以上。作为梯度基方法,它随参数数量高效扩展,并保证收敛至局部最优解,这使其成为研究可识别性问题的理想平台。一般而言,Geodesic-BP收敛至损失函数的局部最小值,该函数衡量记录ECG与模拟ECG之间的不匹配度。多个最小值的存在表明缺乏可识别性,而这些最小值在参数空间中的分布则为潜在正则化策略提供了洞见。 我们通过使用不同初始参数条件和改变ECG导联数量(包括高密度体表面电位图(BSPM)),对单个病例的ECG进行数百次前所未有的高精度拟合,以解决这一问题。随后,我们量化了识别过程中的不确定性。我们的结果通过使用不同EP模拟器生成的高保真真实数据(GT)进行了定量验证(Gillette等人,2022)。最后,我们提出了基于生理学的PMJ约束条件,显著提高了从12导联ECG识别心室传导系统的可识别性。研究未使用GT模型的精确约束,以阐明HPS的生理学假设如何影响ECG精确匹配的能力。 这是首项证明从标准临床12导联ECG以前所未有的精度推断HPS可行性的研究。我们的方法实现了高精度ECG匹配,且所识别的心室激动时间的不确定性范围极窄,远低于最先进侵入性标测技术的测量不确定性。我们认为,我们新颖的稳健且可扩展的优化和正向EP方法是创建具有可验证保真度的可信CDT的关键技术,从而推动CDT向临床应用转化。
Abatract
摘要
Digital twins for cardiac electrophysiology are an enabling technology for precision cardiology. Current forwardmodels are advanced enough to simulate the cardiac electric activity under different pathophysiologicalconditions and accurately replicate clinical signals like torso electrocardiograms (ECGs). In this work, weaddress the challenge of matching subject-specific QRS complexes using anatomically accurate, physiologicallygrounded cardiac digital twins. By fitting the initial conditions of a cardiac propagation model, our noninvasive method predicts activation patterns during sinus rhythm. For the first time, we demonstrate thatdistinct activation maps can generate identical surface ECGs. To address this non-uniqueness, we introduce aphysiological prior based on the distribution of Purkinje-muscle junctions. Additionally, we develop a digitaltwin ensemble for probabilistic inference of cardiac activation. Our approach marks a significant advancementin the calibration of cardiac digital twins and enhances their credibility for clinical application
心脏电生理数字孪生技术是精准心脏病学的赋能技术。当前的正向模型已足够先进,能够模拟不同病理生理条件下的心脏电活动,并准确复现躯干心电图(ECG)等临床信号。在这项工作中,我们致力于解决利用解剖学精确、生理学可靠的心脏数字孪生匹配个体特异性QRS波群的挑战。通过拟合心脏电传导模型的初始条件,我们的非侵入性方法可预测窦性心律期间的激动模式。我们首次证明,不同的激动图谱可生成相同的体表心电图。为解决这一非唯一性问题,我们引入了基于浦肯野纤维-心肌连接处分布的生理学先验知识。此外,我们开发了一种用于心脏激动概率推断的数字孪生集合方法。我们的方法标志着心脏数字孪生校准方面的重大进展,并增强了其在临床应用中的可信度。
Method
方法
2.1. Electrophysiology model
The forward model for the ECG is based on the eikonal equationwith the lead field method (Neic et al., 2017; Pezzuto et al., 2017;Gillette et al., 2021b). Considering a domain 𝛺 ⊂ R3 representingthe active ventricular myocardium, the eikonal equation provides thecardiac LAT 𝜏(𝒙) ∶ 𝛺 → R given a symmetric, positive-definiteconduction velocity tensor 𝐌(𝒙) ∶ 𝛺 → R3×3 and a set of boundaryconditions 𝑋0 :
{√ (𝐌(𝒙)∇𝜏(𝒙)) ⋅ ∇𝜏(𝒙) = 1, in 𝛺,𝜏(𝒙𝑖) = 𝑡 𝑖 , for (𝒙𝑖 , 𝑡𝑖 ) ∈ 𝑋0 ,(1)
where𝐌(𝒙) = 𝑣 2 𝑓 𝒇(𝒙) ⊗ 𝒇(𝒙) + 𝑣 2 𝑠 𝒔(𝒙) ⊗ 𝒔(𝒙) + 𝑣 2 𝑛 𝒏(𝒙) ⊗ 𝒏(𝒙), (2)**
𝑋0 = { (𝒙𝑖 , 𝑡𝑖 ) ∶ 𝒙𝑖 ∈ 𝛺, 𝑡 ̄ 𝑖 ∈ R, 𝑖 = 1,… , 𝑁} , (3)**
and 𝑣{𝑓,𝑠,𝑛} (𝒙) are the conduction velocities (CVs) in the fiber 𝒇, sheet 𝒔and normal 𝒏 direction, respectively. The set 𝑋0 contains the PMJs𝒙𝑖 and their activation time 𝑡 𝑖 , which should satisfy the followingcompatibility condition:
𝑡𝑖 ≤ 𝑡 𝑗 + 𝛿(𝒙𝑖 , 𝒙𝑗 ), for all (𝒙𝑖 , 𝑡𝑖 ), (𝒙𝑗 , 𝑡𝑗 ) ∈ 𝑋0 , (4)
where 𝛿(𝒙, 𝒚) is the travel time between 𝒙 and 𝒚 according to Eq. (1).The numerical approximation, previously described (Grandits et al.,2024), consists in a piecewise linear approximation of 𝜏(𝒙) on a simplicial mesh of 𝛺. The conduction velocity tensor is assumed piecewiseconstant in each element of the grid. The solution is based on a fixedpoint iteration scheme where 𝜏(𝒙) at each node is updated accordingto the neighbor values of 𝜏:𝜏*(𝑘+1)(𝒙𝑖 ) = min𝒚∈𝜕𝜔(𝒙𝑖 ){ 𝜏 (𝑘) (𝒙𝑖 ), 𝜏(𝑘) (𝒚) + 𝛿(𝒙𝑖 , 𝒚) } , (5)where 𝒙𝑖 is the position of the 𝑖th mesh node and 𝜔(𝒙𝑖 ) is the setof elements sharing 𝒙𝑖 . Eq. (5) is applied in parallel to all nodesuntil convergence. The geodesic distance is computed by solving alocal optimization problem on 𝜕𝜔(𝒙𝑖 ) with a fixed number of FISTAalgorithm (Beck and Teboulle, 2009) iterations. When a PMJ (𝒙𝑗 , 𝑡𝑗 ) ∈𝑋0 is not a mesh node, we solve a local eikonal problem in themesh element containing 𝒙𝑗 , and then assign the boundary condition𝜏(𝒙𝑘 ) = 𝑡 𝑗+𝛿(𝒙𝑗 , 𝒙𝑘 ) for all neighbor nodes 𝒙𝑘 . Finally, the compatibilitycondition (4) is automatically fulfilled by the algorithm, in the sensethat when violated, Eq. (1) overwrites the corresponding PMJ with thecorrect timing. In this way, the final number of active PMJs could belower than 𝑁. Overall, our vectorized approach is highly efficient forGPUs and it is suitable for automatic differentation (see Section 2.3).
心电图正向模型基于程函方程与导联场方法(Neic等人,2017;Pezzuto等人,2017;Gillette等人,2021b)。考虑表示活动心室心肌的区域Ω⊂R³,程函方程在给定对称正定传导速度张量𝐌(𝒙)∶Ω→R³×³和一组边界条件X₀时,可提供心脏局部激活时间τ(𝒙)∶Ω→R: {√(𝐌(𝒙)∇τ(𝒙))⋅∇τ(𝒙) = 1,在Ω内,τ(𝒙ᵢ) = 𝑡ᵢ,对于(𝒙ᵢ, 𝑡ᵢ)∈X₀,(1) 其中𝐌(𝒙) = v²f 𝒇(𝒙)⊗𝒇(𝒙) + v²s 𝒔(𝒙)⊗𝒔(𝒙) + v²n 𝒏(𝒙)⊗𝒏(𝒙),(2) X₀ = {(𝒙ᵢ, 𝑡ᵢ)∶𝒙ᵢ∈Ω, 𝑡̄ᵢ∈R, 𝑖=1,…,𝑁},(3) 且v{f,s,n}(𝒙)分别为纤维方向𝒇、片层方向𝒔和法线方向𝒏的传导速度(CV)。集合X₀包含浦肯野-心肌连接处(PMJ)𝒙ᵢ及其激活时间𝑡ᵢ,需满足以下相容性条件: 𝑡ᵢ ≤ 𝑡ⱼ + δ(𝒙ᵢ, 𝒙ⱼ),对所有(𝒙ᵢ, 𝑡ᵢ), (𝒙ⱼ, 𝑡ⱼ)∈X₀,(4) 其中δ(𝒙, 𝒚)是根据方程(1)计算的𝒙与𝒚之间的传导时间。如先前研究(Grandits等人,2024)所述,数值近似通过在Ω的单纯形网格上对τ(𝒙)进行分段线性近似实现。传导速度张量在网格的每个单元中假设为分段常数。解基于不动点迭代方案,其中每个节点的τ(𝒙)根据相邻节点的τ值更新: τ𝑘+1(𝒙ᵢ) = min𝒚∈∂ω(𝒙ᵢ) {τ𝑘(𝒙ᵢ), τ𝑘(𝒚) + δ(𝒙ᵢ, 𝒚)},(5) 其中𝒙ᵢ为第𝑖个网格节点的位置,ω(𝒙ᵢ)为共享𝒙ᵢ的单元集合。方程(5)并行应用于所有节点直至收敛。测地距离通过在∂ω(𝒙ᵢ)上使用固定次数的FISTA算法(Beck和Teboulle,2009)迭代求解局部优化问题计算。当PMJ(𝒙ⱼ, 𝑡ⱼ)∈X₀不是网格节点时,我们在包含𝒙ⱼ的网格单元中求解局部程函问题,然后为所有相邻节点𝒙ₖ分配边界条件τ(𝒙ₖ) = 𝑡ⱼ + δ(𝒙ⱼ, 𝒙ₖ)。最终,算法自动满足相容性条件(4)——若条件被违反,方程(1)会用正确时序覆盖相应PMJ,因此最终激活的PMJ数量可能低于𝑁。总体而言,我们的向量化方法对GPU高度高效,且适合自动微分(见2.3节)。
Results
结果
3.1. Identifiability in the general case
From a theoretical standpoint, there is no guarantee that the HPSgoverning ventricular activation can be uniquely identified from thesurface ECG. Optimization algorithms such as Geodesic-BP might converge towards different local minima of the loss function, dependingon the initial positions and timings of the PMJs. Here, we empiricallyquantify the variability of optimized solutions in the unrestricted case,without imposing any physiological constraints on the permissible positions of PMJs, and by randomly sampling the initial guess for the PMJsin the optimization algorithm (see Section 2.4).Optimization results using Geodesic-BP for fitting the ECG areshown in Fig. 5 (left panel) where the distribution obtained from20 optimized ECGs (𝜇 , 𝜎 ) is overlaid on the GT ECG we aim torecover. The tight envelope around the measured ECG demonstratesthat Geodesic-BP is able to fit the ECG with very high fidelity withoutscaling. The maximum absolute error between optimized and GT ECGwas <2.8 × 10−2 mV (relative error 5.098 930 131 663 843 %) accordingto (17) (average: 2.07 × 10−2 mV/4.066 672 407 563 584 %). Similarly, thePearson correlation coefficient of the ECGs of all samples w.r.t. theGT was >0.994. Convergence is also rapidly achieved with dist fallingbelow 1×10−1 mV within ≤100 iterations for most samples (see also Fig.11).However, the low errors observed in the ECG did not necessarilycorrespond to small errors in the ventricular activation map 𝜏. Thisis illustrated in Fig. 5 (right panel) where the two cases with themost extreme difference (dist𝜏 ) in the activation map are shown. Thevariability in reconstructed position of PMJs was significant betweenthese two samples, as readily evidenced by fundamentally differentinitial activation sites. For instance, a site of earliest activation in the LVmid-anterior endocardium in one run (top) is a site of latest activationin another run (bottom). Quantitatively, the absolute error in LAT was23.12 ms (average between the samples: 18.32 ms).While the 12-lead ECG was very
从理论角度来看,并无保证能从体表心电图唯一识别调控心室激动的希氏-浦肯野系统(HPS)。诸如Geodesic-BP的优化算法可能会因浦肯野-心肌连接处(PMJ)的初始位置和时序不同,收敛至损失函数的不同局部最小值。在此,我们通过在优化算法中对PMJ初始猜测值进行随机采样(见2.4节),在无约束情况下(即不对PMJ允许位置施加任何生理限制),从经验上量化优化解的可变性。 使用Geodesic-BP拟合心电图的优化结果如图5(左图)所示,其中20次优化得到的心电图分布(均值$\mu$,标准差$\sigma$)叠加在目标重建的真实心电图(GT)上。测量心电图周围的紧密包络线表明,Geodesic-BP能够在不进行缩放的情况下以极高保真度拟合心电图。根据(17)式,优化后与GT心电图的最大绝对误差<2.8×10⁻² mV(相对误差5.098 930 131 663 843%),平均值为2.07×10⁻² mV/4.066 672 407 563 584%。类似地,所有样本心电图与GT的皮尔逊相关系数>0.994。对于大多数样本,距离dist$\phi$在≤100次迭代内降至1×10⁻¹ mV以下,收敛速度较快(另见图11)。 然而,心电图中观察到的低误差并不一定对应心室激动图$\tau$的小误差。如图5(右图)所示,激动图差异最大的两个案例(dist$\tau$)中,这两个样本间重建的PMJ位置可变性显著,从根本不同的初始激动位点即可明显看出。例如,一次运行中左室前壁中部心内膜的最早激动位点(上图),在另一次运行中却成为最晚激动位点(下图)。定量而言,局部激活时间(LAT)的绝对误差为23.12 ms(样本间平均值为18.32 ms)。 尽管所有样本的12导联心电图极为相似,但体表面电位图(BSPM)中可观察到更明显的可变性,如图6所示。我们同样展示了根据(18)式差异最大的两个BSPM(最大绝对差异0.11 mV),以及GT BSPM。值得注意的是,图6中BSPM距离最大的这两个样本,与图5中激动图距离最大的样本并非同一组。 总体而言,所有样本均能较好捕捉偶极子形躯干电位,表明所有样本的整体激动模式相似。BSPM中相关性最低的区域位于胸骨区,其中重建的BSPM表现出比GT更复杂的模式。选取该区域的一个点,可见两个样本的细胞外电位$\phi_i$、$\phi_j$与GT的$\phi_{GT}$之间存在差异(图6底图)。该未观测点处的细胞外电位与GT电位的相关系数相对较低,分别仅为0.69和0.64。
Figure
图

Fig. 1. Geodesic-BP: fast definition of a CDT from the surface ECG. Geodesic-BP (Grandits et al., 2024) optimizes the distribution of PMJs until the mismatch between recorded and simulated ECG is minimized. Physiological constraints for the PMJs are automatically imposed during the optimization process
图1. Geodesic-BP:从体表心电图快速构建心脏数字孪生。Geodesic-BP(Grandits等人,2024)优化浦肯野-心肌连接处(PMJ)的分布,直至记录的心电图与模拟心电图之间的不匹配度降至最低。优化过程中会自动施加针对PMJ的生理学约束条件。

Fig. 2. Endocardial surface restrictions limits the domain of possible locations for PMJs. The PMJs are restricted to lie inside the subdomain 𝑆 which is spanned by a band,equidistant to the endocardium 𝑆?
图2. 心内膜表面约束限定了浦肯野-心肌连接处(PMJ)的可能位置范围。PMJ被限制位于子区域S内,该区域由与心内膜S等距的条带所构成。

Fig. 3. The in silico GT solution generated from the physiologically-detailed HPS. The LAT map 𝜏 is shown together with the HPS in blue (top). The bottom row shows thecorresponding, calculated 12-lead ECG
图3. 由生理细节丰富的希氏-浦肯野系统(HPS)生成的计算机仿真真实数据(GT)解决方案。顶部显示局部激活时间(LAT)图𝜏和蓝色的HPS结构,底行展示相应计算得到的12导联心电图。

Fig. 4. Electrode locations of the 12-lead ECG (left) compared to BSPM configurations comprising 32, 64 and 128 electrodes from an anterior (top) and posterior (bottom) view.For each BSPM electrode the uni-polar lead field was computed using the WCT as reference. Electrodes for computing the WCT are shown in red
图4. 12导联心电图的电极位置(左)与包含32、64和128个电极的体表面电位图(BSPM)配置对比(前视图在上,后视图在下)。每个BSPM电极的单极导联场以Wilson中央电端(WCT)为参考计算,用于计算WCT的电极以红色显示。

Fig. 5. Identifiability problems of inverse ECG methods. Left panel: Using Geodesic-BP, the identification of optimal PMJs locations and timings to fit a given ECG (GT) can beachieved with high fidelity. The ECG distribution obtained from 20 optimization runs with different initializations forms a tight envelope around the ECG to be reconstructed(left, 𝜇 ± 𝜎 ). Right panel: The two samples with the largest difference in LAT are shown that reveal significant variability in the solution space which is indicative of limitedidentifiability in the general unrestricted case. We highlight a few of the major differences in LAT (red dashed ellipses). This prompts for quantification of this variability and fora-priori constraints to reduce the non-uniqueness of the solution space.
图5. 心电图逆问题的可识别性挑战。左图:使用Geodesic-BP方法可高精度识别最优浦肯野-心肌连接处(PMJ)的位置和时序以匹配给定心电图(真实数据GT)。通过20次不同初始化的优化运行得到的心电图分布形成围绕目标重建心电图的紧密包络线(左图,均值±标准差)。右图:展示了局部激活时间(LAT)差异最大的两个样本,揭示了解空间中显著的可变性——这表明在无约束常规情况下可识别性有限。我们用红色虚线椭圆标注了LAT的几个主要差异区域。这提示需要量化这种可变性,并通过先验约束减少解空间的非唯一性。

Fig. 6. Shown are the two BSPMs of maximum distance among the 20 samples next to the sought-after GT solution at a single instant in time (𝑡 = 60 ms, top), after normalizationwith respect to the mean unipolar potential (see Section 2.6). Differences in potential fields are witnessed in space across the body surface, as well as over time, when comparedat a single unseen electrode location (bottom panel). At all ECG electrode locations used for optimization no differences are apparent, all potential traces are predicted withapproximately zero loss (see Fig. 5, left panel).
图6. 图中展示了20个样本中与目标真实解(GT)距离最大的两个体表面电位图(BSPM),对应单一时刻(t=60 ms,上图),并已对平均单极电位进行归一化处理(见2.6节)。对比可见,体表电位场在空间分布上存在差异;在单个未参与优化的电极位置(底图)进行时域对比时,也能观察到电位随时间的变化差异。值得注意的是,所有用于优化的心电图电极位置均未显示明显差异,所有电位曲线的预测损失几乎为零(见图5左图)。

Fig. 7. Preview of the distribution of the 20 samples, reconstructed from 12-lead ECG and BSPM with 128 electrodes as observations, both with and without physiologicalconstraints. We visualize the sample most closely representing the mean 𝜏𝜇 (top 2 rows) along with the deviation between t
图7. 预览从12导联心电图和128电极体表面电位图(BSPM)作为观测值重建的20个样本分布,分为有生理约束和无生理约束两种情况。我们可视化了最接近平均局部激活时间𝜏𝜇的样本(前2行),以及…(注:原文此处未完整显示,推测后续内容为“激活时间偏差”相关描述)。

Fig. 8. Comparison between pseudo-bidomain solutions on the torso in terms of errors in potential reconstruction dist𝜙 (18) (left panel) and in the ventricular activation sequencedist𝜏 (19) (right panel). Each pair of boxplots shows the errors for unrestricted (blue) and restricted (brown) cases, respectively. Higher density observations led to a significantimprovement in the reconstruction of the BSPM, but impact on the reconstruction of the ventricular activation map 𝜏 was margina
图8. 躯干伪双域解在电位重建误差dist𝜙(18)(左图)和心室激动序列误差dist𝜏(19)(右图)方面的对比。每对箱线图分别显示无约束(蓝色)和有约束(棕色)情况的误差。高密度观测显著改善了体表面电位图(BSPM)的重建,但对心室激动图𝜏重建的影响微乎其微。

Fig. 9. Comparison between LATs as reconstructed with an increasing number of torso electrodes, both for ignoring (left panel) and enforcing (right panel) physiological restrictions.For each boxplot, the average dist𝜏 with respect to the GT of the 20 samples is shown as a number at the bottom line
图9. 对比在忽略(左图)和施加(右图)生理约束条件下,随着躯干电极数量增加所重建的局部激活时间(LAT)。每个箱线图下方的数字表示20个样本相对于真实数据(GT)的平均距离dist𝜏。

Fig. 10. Hyperparameter study of the number of PMJs. The 12-lead ECG was fitted using a variable number of initial PMJs and the error in reconstructing the ECG and theactivation map 𝜏 was measured against the GT model in terms of dist and dist𝜏 . A zoom-in on the results for 𝑁 ≥ 500 is shown in the inset. In the top panel optimization resultsare shown for a number of PMJs of 𝑁 ∈ {10, 50, 300, 5000}, comprising the reconstructed LAT maps and the reconstructed Einthoven II ECG (solid) next to the GT ECG (dashed)
图10. 浦肯野-心肌连接处(PMJ)数量的超参数研究。使用可变数量的初始PMJ拟合12导联心电图,并根据真实数据(GT)模型,通过距离distΦ和distτ衡量心电图重建误差与激动图τ重建误差。插图放大显示N≥500时的结果。上图展示了N∈{10,50,300,5000}时的优化结果,包括重建的局部激活时间(LAT)图和重建的肢体Ⅱ导联心电图(实线),旁边为真实心电图(虚线)。

Fig. 11. Convergence history of Geodesic-BP: We show the evolution of the solution in terms of the ECG on the left 𝑦-axis (logarithmic), over the iterations (first row) andapproximate computational time (second row) on the 𝑥-axis. The top 4 plots show the computed lead II ECG in the current iteration (solid) next to the GT ECG (dashed).Additionally, we show the averaged movement of the initial sites 𝒙𝑖w.r.t. the last iteration in red (right 𝑦-axis, also in logarithmic scale).
图11.Geodesic-BP的收敛过程:我们通过迭代次数(第一行)和近似计算时间(第二行)为横轴,展示了基于心电图(左纵轴,对数刻度)的解的演化过程。前4幅图显示了当前迭代计算的Ⅱ导联心电图(实线)与真实心电图(虚线)的对比。此外,我们用红色标注了初始位点𝒙𝑖 相对于最后一次迭代的平均移动距离(右纵轴,同样为对数刻度)。