Accuracy of cardiac scar and fibrosis geometry in ventricular models affects electrical dyssynchrony biomarkers and ML prediction score of CRT response

Cardiac resynchronization therapy (CRT) has become the gold standard for the treatment of patients with chronic heart failure (CHF). However, up to 30% of patients has no improvements after treatment, namely, no reverse remodeling in the left ventricuar (LV) geometry and ejection fraction (EF) improvements. Thus, the selection of patients for CRT, the planning of device implantation, and the optimization of the pacing electrode positions are still important clinical tasks. Personalized 3D computer models, as well as predictive models based on machine learning (ML), can help to solve these problems. Such models should take into account the peculiarities of cardiac geometry, including the structure of myocardial fibrosis and scarring regions. An open question is the accuracy of such fibrosis models to make reliable predictions. In our study, we compared the electrophysiological biomarkers produced by cardiac models with different accuracy of the fibrosis/scar geometry. Additionally, we used an ML-score of CRT success developed recently by our group to evaluate effect of the fibrosis/scar geometry on CRT prognosis for such models. To construct tetrahedral mathematical models of the heart, torso, and lungs, we used computed tomography (CT) data from four patients. We used two types of fibrosis/scar geometry representation in the ventricular models based on the contrasted Magnetic Resonance Imaging (MRI) data: in the first case, fibrosis geometry was derived from a schematic report on the damaged scarring and fibrosis regions made by a radiologist; in the second case, the fibrosis was manually segmented by an expert. Using ventricular models with myocardial scarring regions, we calculated the electrical activation map, ECG, and several electrical dyssynchrony biomarkers such as the total activation time, ECG duration, intraventricular and interventricular dyssynchrony indices. Then we fed the simulated biomarkers to a ML classifier to compare ML-scores predicting EF improvement after CRT for the two types of fibrosis geometry models. Results of our study show that a detailed structure of myocardial damage is essential for improving the accuracy of predicting CRT response.