Artificial intelligence analysis of the impact of fibrosis in arrhythmogenesis and drug response

Abstract

Background: Cardiac fibrosis has been identified as a major factor in conduction alterations leading to atrial arrhythmias and modification of drug treatment response. Objective: To perform an in silico proof-of-concept study of Artificial Intelligence (AI) ability to identify susceptibility for conduction blocks in simulations on a population of models with diffused fibrotic atrial tissue and anti-arrhythmic drugs. Methods: Activity in 2D cardiac tissue planes were simulated on a population of variable electrophysiological and anatomical profiles using the Koivumaki model for the atrial cardiomyocytes and the Maleckar model for the diffused fibroblasts (0%, 5% and 10% fibrosis area). Tissue sheets were of 2 cm side and the effect of amiodarone, dofetilide and sotalol was simulated to assess the conduction of the electrical impulse across the planes. Four different AI algorithms (Quadratic Support Vector Machine, QSVM, Cubic Support Vector Machine, CSVM, decision trees, DT, and K-Nearest Neighbors, KNN) were evaluated in predicting conduction of a stimulated electrical impulse. Results: Overall, fibrosis implementation lowered conduction velocity (CV) for the conducting profiles (0% fibrosis: 67.52 ± 7.3 cm/s; 5%: 58.81 ± 14.04 cm/s; 10%: 57.56 ± 14.78 cm/s; p < 0.001) in combination with a reduced 90% action potential duration (0% fibrosis: 187.77 ± 37.62 ms; 5%: 93.29 ± 82.69 ms; 10%: 106.37 ± 85.15 ms; p < 0.001) and peak membrane potential (0% fibrosis: 89.16 ± 16.01 mV; 5%: 70.06 ± 17.08 mV; 10%: 82.21 ± 19.90 mV; p < 0.001). When the antiarrhythmic drugs were present, a total block was observed in most of the profiles. In those profiles in which electrical conduction was preserved, a decrease in CV was observed when simulations were performed in the 0% fibrosis tissue patch (Amiodarone ΔCV: -3.59 ± 1.52 cm/s; Dofetilide ΔCV: -13.43 ± 4.07 cm/s; Sotalol ΔCV: -0.023 ± 0.24 cm/s). This effect was preserved for amiodarone in the 5% fibrosis patch (Amiodarone ΔCV: -4.96 ± 2.15 cm/s; Dofetilide ΔCV: 0.14 ± 1.87 cm/s; Sotalol ΔCV: 0.30 ± 4.69 cm/s). 10% fibrosis simulations showed that part of the profiles increased CV while others showed a decrease in this variable (Amiodarone ΔCV: 0.62 ± 9.56 cm/s; Dofetilide ΔCV: 0.05 ± 1.16 cm/s; Sotalol ΔCV: 0.22 ± 1.39 cm/s). Finally, when the AI algorithms were tested for predicting conduction on input of variables from the population of modelled, Cubic SVM showed the best performance with AUC = 0.95. Conclusion: In silico proof-of-concept study demonstrates that fibrosis can alter the expected behavior of antiarrhythmic drugs in a minority of atrial population models and AI can assist in revealing the profiles that will respond differently.

Keywords: atrial fibrillation; cardiac fibrosis; machine learning; population of models; support vector machines.