Optimization of left ventricular pacing site in cardiac resynchronization therapy using combination of personalized computational modeling and machine-learning

Abstract

ABSTRACTBackgroundUp to 30-50% of selected patients with chronic heart failure do not respond to cardiac resynchronization therapy (CRT). Optimization of pacing lead placement in ventricles remains a challenge.ObjectiveWe utilize a machine learning (ML) classifier to predict the position of an optimal left ventricular (LV) pacing site maximizing the probability of CRT response for a certain patient.Materials and MethodsRetrospective data from 57 patients with implanted CRT devices were utilized. Positive response to CRT was defined by a 10% improvement in the LV ejection fraction in a year after implantation. For each patient, a personalized model of ventricular activation and ECG was developed based on MRI and CT images. The total ventricular activation time, QRS duration and electrical dyssynchrony indices during intrinsic rhythm and biventricular (BiV) pacing with clinical pacing lead position (ref-LP) were computed and used in combination with clinical data to train the ML algorithm. We built a logistic regression classifier predicting CRT response with a high ROC AUC=0.84 and an average accuracy of 0.77. It generates a ML-score estimating the probability of CRT response. ML-scores were computed from model-driven features for varying LV pacing sites. Then Bayesian optimization was used to interpolate the ML-score over the available LV surface and an optimal LV lead position for BiV pacing that maximizes ML-score (ML-LP) was defined.ResultsThe optimal LV pacing site position increased the average ML-score by 17% in the patient cohort. Moreover, 11 out of 34 (20%) non-responders classified as true negative at ref-LP were re-classified as positive at ML-LP. In a patient group (n=14, 25% of the cohort) with LV pacing lead deployed in close proximity to the optimal position, the ratio of responders to non-responders was three times higher than in the entire cohort.ConclusionWe have developed a new technique based on simulations and ML to define an optimal position for LV lead for BiV pacing maximizing the ML-score of CRT response on the available LV epicardial surface. This technique demonstrates a high potential for the improvement of CRT outcome with guided lead implantation.