Cardiac arrhythmogenesis is an inherently multi-scale phenomenon. Dysfunction of specific ion channels or abnormalities in subcellular Ca2+ handling have been causally linked to arrhythmias in vivo and have been investigated in detailed computational cardiomyocyte models. However, such (sub)cellular models are computationally demanding, precluding their use in organ-level simulations required for investigating the proarrhythmic consequences of subcellular alterations or the complex pharmacodynamic effects of antiarrhythmic drugs. Recent work has shown how detailed cellular models can be used to parameterize simpler models that can be used in tissue-level simulations [6, 9], providing examples of how challenges related to multi-scale simulations can be partially overcome.

Another significant challenge for progress in the clinical translation of digital twins is the reproducibility and technical validation of the numerous different frameworks and implementations maintained by different researchers worldwide. While the sharing of models has improved through simulation frameworks such as Myokit [7] and OpenCARP [21], modeling standards such as CellML, and open source pipelines for mesh generation from imaging data [23], the fact that many models are still only available in custom-made, often poorly documented code, or that exact mesh details and initial conditions are often not reported, hinders the reproducibility and technical validation of model implementations.

Precision cardiology aims to deliver targeted, mechanism-based treatment for an individual patient. Until now, most digital twin studies have used non-invasive imaging data (CT, CMR) to obtain patient-specific anatomies and indicators of structural remodeling (generally fibrosis obtained through late gadolinium-enhanced CMR) [28]. However, existing modalities have important limitations, e.g., related to spatial resolution, imaging artefacts due to implanted devices, and standardization of cut-off values, leading to overestimation of fibrotic content and inability to identify all channels of surviving myocardium in a complex scar [12]. Recent advances, including dedicated high-resolution CMR protocols and photon-counting CT, may help to address some of these limitations.

Compared to structural remodeling, personalization of electrical parameters is less developed in most digital twin approaches. Standard 12-lead electrocardiograms (ECGs) are available for most patients and can be used to parameterize digital twins [13]. However, standard ECGs provide limited information on spatial repolarization patterns, resulting in a non-unique relationship between ECG and spatiotemporal electrophysiological properties, limiting the identifiability of model parameters. Invasive mapping data can be used to further personalize the electrophysiological properties of digital twins [23], but require an invasive procedure with data points collected over multiple heart beats and with limited spatial relation to the imaging-based mesh. Non-invasive electrocardiographic imaging may reveal local repolarization abnormalities that are not visible on a standard 12-lead ECG, e.g., in patients with idiopathic ventricular fibrillation [8], potentially providing a non-invasive data source to personalize model parameters. However, its spatial accuracy is limited and primarily reflects epicardial information. Importantly, none of these approaches provides clear information on the underlying biophysical mechanisms, with multiple combinations of ionic changes potentially giving rise to the same electrophysiological phenotype at baseline, but responding differently during tachyarrhythmias. Since molecular or functional data at the level of the cardiomyocyte are not available in the vast majority of patients, no truly personalized cellular electrophysiology models are currently available. Perhaps future advances, e.g., in the generation of patient-specific induced pluripotent stem cell-derived cardiomyocytes, may enable additional personalization of electrophysiological parameters, although numerous limitations remain, including functional immaturity and the inability to control the regional differences in electrophysiological properties (e.g., apicobasal or transmural).

Finally, even if high-quality patient-specific data can be obtained for model personalization, these data are typically only available at a single moment in time, ignoring the dynamic nature of cardiac electrophysiology and the sudden occurrence of arrhythmias. Recent work has shown how the performance of a static machine learning-based VT risk predictor derived from the baseline ECG dropped over time, whereas a dynamic model incorporating time-varying ECG data showed increased performance over time [19]. More research is needed to identify methodologies to incorporate dynamic changes in arrhythmogenic risk in digital twins, e.g., based on blood biomarkers or wearables.

Ultimately, the proverbial proof of the pudding that will determine the clinical success of digital twins for cardiac arrhythmia management will be high-quality large-scale RCTs demonstrating the benefit of digital twin-guided care over routine clinical care on clinically relevant outcomes [12]. However, such trials are complex and time consuming, requiring close collaboration between clinicians, computational modelers, and industry [12, 15]. Even if such trials were to show positive results, significant efforts would be required to enable routine clinical application of digital twins, e.g., in terms of work-flow automation, standardization, and processing times, in order not to disrupt normal clinical care. On the other hand, digital twins based on mechanistic models have the advantage of being easily interpretable, potentially facilitating acceptance of model-guided care over other black-box approaches. Ultimately, the feasibility and cost-effectiveness of the use of digital twins for the treatment of the large number of patients with cardiac arrhythmias will need to be demonstrated to ensure reimbursement and subsequent routine clinical application (Fig. 2).