How to address non-responders to cardiac resynchronization therapy issue using artificial intelligence: the potential role of an explainable electrocardiogram-based deep learning.

Cardiac Resynchronization Therapy (CRT) markedly improves outcomes in heart failure patients, including heart function, symptom alleviation, overall quality of life, and mortality reduction1. However, despite the evident benefits of CRT, not all patients respond favourably. This variable response can be attributed to multiple factors, such as the optimal placement of the Left Ventricular (LV) lead, the appropriate programming of the atrioventricular and interventricular intervals, and the maintenance of a biventricular pacing percentage over 99%. Nevertheless, the most influential factor in a patient's response to CRT is appropriate patient selection, which currently relies heavily on electrocardiographic criteria. Indeed, the concept of CRT is founded on the frequent observation of high-grade intraventricular conduction delays in patients with HF and LV systolic dysfunction2-3.

Current guidelines clearly outline ECG criteria for choosing the right patients for CRT, but what if there were other important ECG characteristics for selecting these patients, perhaps not clearly recognizable by the human eye? Could a deep learning approach help doctors predict patient outcomes to CRT using just a standard ECG?

In this study, an innovative deep learning-based algorithm, FactorECG, was presented and its performance compared with current guideline ECG criteria1 and QRSAREA4-5. Designed to utilize solely standard 12-lead ECG, the algorithm aimed to predict long-term clinical outcomes, HF hospitalization, and echocardiographic non-response in CRT-eligible patients. It was trained on 1.1 million ECGs from 251 473 patients, and adopted an explainable deep learning model, addressing the 'black box' concern often associated with AI.

The primary endpoint was a combined clinical endpoint consisting of left ventricular assist device (LVAD) implantation, heart transplantation (HTx), and all-cause mortality. The secondary endpoint was echocardiographic non-response, defined as a relative decrease in LVESV of < 15%. In addition, three tertiary endpoints were investigated: a composite of HF hospitalization and the primary endpoint, HF hospitalization alone, and ≥1 point of NYHA functional class improvement. Utilizing a large multicentre database for training, the algorithm underwent an internal validation process using bootstrapping to ensure unbiased results.
FactorECG predicted the combined clinical endpoint of death, left ventricular assist device, or heart transplantation [c-statistic 0.69, 95% confidence interval (CI) 0.66–0.72], significantly outperforming QRSAREA and guideline ECG criteria [c-statistic 0.61 (95% CI 0.58–0.64) and 0.57 (95% CI 0.54–0.60), P< 0.001 for both]. Moreover, FactorECG identified inferolateral T-wave inversion, smaller right precordial S- and T-wave amplitude, ventricular rate, and increased PR interval and P-wave duration to be important predictors for poor outcome. Finally FactorECG proved to be particularly effective in predicting outcomes in females and the non-ICM population.

In summary, from a clinical perspective, FactorECG could provide a continuous evaluation of the electrical substrate, thus enabling a departure from the traditional binary classification of LBBB morphology.
Despite these promising findings, the study has some limitations. The data used comes from a single vendor, and while internal validation was conducted, external validation in different types of patient populations is still needed for wider generalizability. The results need to be validated also for CRT-P patients and for those directed towards CSP (Conduction System Pacing) therapy as an alternative to CRT.
Lastly, prospective studies with FactorECG are warranted to acquire CE certification, allowing its use as a medical device.