Explainable deep learning outperforms guideline criteria and QRSarea for prediction of outcome after cardiac resynchronization therapy. (19th May 2022)

• Record Type: Journal Article
• Authors: Van De Leur, RR; Wouters, PC; Vessies, MB; Van Stipdonk, AMW; Ghossein, MA; Maass, AH; Prinzen, FW; Vernooy, K; Meine, M; Van Es, R
• Is Part Of: Europace. Volume 24:Supplement 1(2022)
• Publication Date: 2022-05-19
• Subjects: Arrhythmia -- Treatment -- Periodicals; Cardiac pacing -- Periodicals; Catheter ablation -- Periodicals; Heart -- Physiology -- Periodicals; Electrophysiology -- Periodicals; 617.4120645
• DOI: 10.1093/europace/euac053.482
• Languages: English
• ISSNs: 1099-5129

Abstract

Abstract: Funding Acknowledgements: Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Netherlands Organisation for Health Research and Development (ZonMw) Background: Electrocardiogram-based prediction models for cardiac resynchronization therapy (CRT) response mainly focus on the QRS complex, but other information in the electrocardiogram (ECG) is neglected. Purpose: We sought to identify and visualize ECG features using an explainable deep learning-based algorithm (FactorECG) to predict CRT outcome and echocardiographic response, and compare this to state-of-the-art parameters, including QRSarea. Methods: Patients who underwent CRT implantation in three academic hospitals were analyzed for clinical outcome (death, left ventricular assist device implantation, or heart transplantation), and echocardiographic response (≥ 15% left ventricular end-systolic volume reduction at 6 months). Pre-implantation ECGs were converted into their respective FactorECG. By using a deep learning algorithm trained on 1.1 million ECGs, a compressed version of the median beat ECG was obtained, where all ECG features are summarized in only 21 explainable factors of variation (interactive tool: https://decoder.ecgx.ai ). The 21 FactorECG values per patient were used in a Cox and logistic regression model, for outcome and response, respectively. Models were trained on data from two hospitals (n = 936 for outcome and n = 591 for response) and externally validated in a third hospital (n = 339 for outcome and n = 230 for response). Furthermore, ESC CRT guideline indications and vectorcardiographic QRSarea were used as a comparison. Results: The deep learning-based approach was able to predict clinical outcome (AUC = 0.74 [95% confidence interval (CI) 0.69 - 0.80]) and echocardiographic response (AUC = 0.70 [95% CI 0.63 - 0.77]). Moreover, it significantly outperformed a model based on the ESC CRT guidelines for outcome (AUC = 0.57 [95% CI 0.50-0.63]) and response (AUC = 0.57 [95% CI 0.51 - 0.64]). In comparison with QRSarea, the deep learning-based approach performed significantly better for outcome (AUC = 0.61 [95% CI 0.53-0.69], but similar for response (AUC = 0.70 [95% CI 0.64-0.77]. Based on QRSarea and predicted probabilities of the deep learning approach, for both outcome at three years and response, four groups of similar size were identified and compared to the ESC CRT guidelines (Figure 1). Important ECG factors for poor response and clinical outcome were identified as anterior T-wave inversion (F9), increased heart rate (F10), non-LBBB morphology (F26) and increased PR-interval (F8, Figure 2). Conclusion: Without compromising interpretability, the deep learning-based algorithm objectively identified CRT recipients with good clinical outcome and echocardiographic response. This approach outperformed QRSarea for outcome, and QRS morphology and duration as used in the ESC CRT guidelines for both outcome and echocardiographic response. Hence, most predictive information for CRT response is found within the QRS complex, whereas the complete median beat ECG provides additional predictive information for clinical outcome.