Dynamic machine learning model with latent ECG representations analysis for malignant ventricular arrhythmias prediction: ready to open the black box?

Ventricular arrhythmias (VAs) are among the leading causes of sudden cardiac death (SCD) in patients affected by structural heart disease, both ischemic and non-ischemic, and the use of implantable cardioverter defibrillators (ICDs) is the only effective therapy to prevent SCD in high-risk patients. Current non-invasive SCD risk stratification strategies are only based on the left ventricular ejection fraction (LVEF) and the presence and severity of heart failure symptoms (1). However, stratifying SCD risk based solely on these two risk factors has major well-recognized limitations. First, the majority of patients who experience SCD may not have abnormal LVEF or heart failure (2). And second, both risk factors are strongly associated with other modes of cardiovascular death. This has important clinical implications, as patients who receive primary prevention ICDs on the basis of LVEF and heart failure can be more likely to die from any other reason than from VA (main targets of ICD shocks). In recent years, many potential candidates have been explored, compared and evaluated, in terms of performance for VAs recurrence prediction, thus potentially being able to replace, or at least integrate, LVEF as decision factors. 

Artificial intelligence (AI) models that can process large amount of static and dynamic data may represent a viable future alternative to known current prognostic patterns. In a recent original research article published on eBioMedicine (part of Lancet Discovery Science), Kolk et al. (3), report a multicentre study aimed at evaluating the performance of a dynamic AI model trained to predict VAs based on multiple variables, including baseline clinical information and latent ECG representations obtained through variational autoencoders (VAEs). A total of 32,129 ECG recordings obtained from 2,942 ICD carriers (61.7 ± 13.9 years, 25.5% female) were analysed, with a mean follow up of 43.9 ± 35.9 months. Both a static and a dynamic random survival forest model were trained. The static model incorporated clinical variables and ECG latent variables obtained at device implantation to estimate a personalized survival function, while the dynamic model also included time-varying ECG latent variables collected during follow-up. The mean time-varying area under the ROC curve for the dynamic model was 0.738 ± 0.07, compared to 0.639 ± 0.03 for the static model. Feature analyses indicated dynamic changes in latent ECG representations, particularly in terms of T-wave morphology, were of highest importance for model performance.

First, it is worth mentioning that the authors propose an AI model trained on a peculiar form of data obtained through VAEs: latent ECG representations. Invisible to the human eye, these subtle static and dynamic ECG features can be seen as the minimum set of features representing the electrical and anatomical variations of the arrhythmogenic substrate observed on the ECG morphology and could also mirror unknown clinical events that may trigger arrhythmic episodes.

Overall, the electrocardiogram may be one the most cost-effective biosignal sources in the era of large volume AI models, given its easy and cheap availability and its intrinsic “digital” nature. Moreover, in recent years, this tool has witnessed a progressive miniaturization and “democratization” related to wearable personal devices and dedicated gadgets.

These high-performance predictive models may have potential large-volume implications when applied to standard office 12-lead tracings or even to single or multiple lead portable ECGs obtained with smartwatches or smartbands in large-scale screening, but we must not forget that a selected population of higher-risk patients who already carry an ICD may provide continuous ECG-like tracings from their farfield signal, and their remote post-processing may be possible by remote monitoring platforms, which could enable early event prediction and prompt preventive clinical action.

On the other hand, despite progress in interpretability and efforts to open the “black box”, we must ask whether the clinical cardiologist is ready to welcome in his daily practice new data sources, whose intrinsic nature is hidden, not completely understandable. Are we ready to take a leap of faith?