Using Computational Modeling and Artificial Intelligence to Uncover Arrhythmogenic Mechanisms and Advance Arrhythmia Management

Computational modeling of the heart has become an important research tool in cardiac electrophysiology and arrhythmias with recent advances in real-world applications. Artificial intelligence (AI) is revolutionizing cardiac arrhythmia management by aiding detection, predicting risk, and contributing to clinical decision-making.

Characterizing the electrophysiological effects of stretch-activated ion channels

The interaction between cardiac electrophysiology and mechanics is modulated through the engagement of stretch-activated ion channels (SACs), however, their role in arrhythmogenesis is not well understood. Using a computational model of human ventricular myocyte including SACs, a recent study¹ systematically characterized the effects of stretch on cell dynamics. The model incorporated 3 types of SACs, potassium-selective, calcium-selective, and non-selective, all calibrated to experimental data. The study demonstrated that stretch may provoke afterdepolarizations, give rise to new triggered beats, change the morphology and duration of a subsequent action potential, or prevent triggered activity. Some of these effects could be proarrhythmic, depending on stretch characteristics (amplitude, time of application, duration) and disease-induced SAC remodeling. The study serves as template for deciphering the mechanisms of stretch-induced arrhythmias.

A simulator of neurocardiac modulation

As cardiac function is regulated by the autonomic nervous system (ANS), any imbalances between sympathetic and parasympathetic inputs could lead to rhythm disorders. Dissecting the effects of autonomic stimulation remains a major challenge as synapses in different heart regions result in multiple changes in function. Computational modeling can aid understanding as it integrates inputs at disparate spatial scales. The Clancy lab developed² a multiscale neurocardiac simulator that recapitulates features of autonomic control and predicts the effects of efferent stimulation of the two ANS branches on the cardiac sinoatrial node and ventricular myocardium. The model was able to make predictions across all system scales, from subcellular signaling to pacemaker frequency to tissue level responses, in close agreement with experiment. The study ascertained the conditions under which autonomic imbalance may increase propensity to arrhythmias or be used to terminate arrhythmia.

Assessing atrial arrhythmogenicity due to fibrosis to inform atrial fibrillation ablation

For patients with persistent atrial fibrillation (AF) and fibrosis proliferation, the standard AF treatment, pulmonary vein isolation (PVI), has poor outcomes. Atrial fibrosis creates substrates capable of sustaining reentrant activations. In a recent study³, persistent AF patients scheduled for ablation were enrolled prospectively to generate their bi-atrial digital twins (personalized multi-scale models) from the pre-procedure contrast-enhanced MRIs. The digital twins were used to uncover substrate locations possessing reentry-attracting capabilities that represent potential extra-PVI ablation targets. The study found that not all substrate locations capable of giving rise to sustained reentries are made equal. Some locations lose their reentry-attracting capabilities upon ablation at another location, and thus do not need to be targeted. Others persist in forming reentries and need to be ablated to eliminate propensity to arrhythmia. This ensures elimination of arrhythmia propensity with minimum lesions while also minimizing risk of AF recurrence.

Identifying low-burden AF by deep leaning on echocardiograms

The advent of medical artificial intelligence (AI) that can predict adverse events and outcomes has led to focused efforts on early detection of subclinical AF, thereby decreasing stroke risk. While most attempts have focused on deep learning on the ECG, a recent study⁴ sought to determine whether deep learning on transthoracic echocardiogram (TTE) videos could identify patients with low-burden AF. The two-stage algorithm distinguished whether TTEs were in sinus rhythm or AF with high accuracy, outperforming current clinical risk factors, and then predicted which of the sinus rhythm TTEs were from patients who had experienced AF within 90 days. An ensemble model in a cohort subset combining the TTE model with an ECG deep learning performed better than analysis on ECG alone. Applying deep learning to routinely acquired TTEs and bedside TTEs at the point-of-care presents a new opportunity for improving AF screening.

AI-enabled ECG analysis for risk stratification of occlusion myocardial infarction

Patients with occlusion myocardial infarction (OMI) with no ST-elevation on ECG typically have poor prognosis. Currently, there are no accurate tools to identify such patients during initial triage so that emergency reperfusion therapy could be administered. A multi-center prospective observational cohort study⁵ evaluated the diagnostic accuracy of ECG machine learning analysis (a random forest model) for risk stratification of OMI in the absence of a STEMI pattern in patients with chest pain at first medical contact. The results demonstrated the superiority of AI-ECG in detecting subtle ischemic ECG changes, outperforming practicing clinicians and widely used commercial ECG interpretation software. The study outlined the most important ECG features responsible for the model’s prediction and identified ECG patterns indicative of acute coronary occlusion beyond the clinical guidelines’ recommendations. When combined with the judgment of trained personnel, the algorithm helped correctly reclassify one in three patients with chest pain.