Comprehensive Summary
In this systematic review, researchers assessed 46 studies that utilized predictive models and deep learning or machine learning approaches to predict malignant ventricular arrhythmias (VAs) or other adverse cardiac events from non-invasive electrophysiological signals. Bias of these studies was evaluated as low, high, or unclear risk of bias for participants, predictors, outcomes, and analysis. For studies with sufficient data provided, ML and DL models were assessed through meta-analysis, and researchers calculated diagnostic odds ratios (DOR) and area under the summary receiver operator curves (AUROC). Separate meta-analyses for studies that did or did not use ad-hoc datasets were performed. The former displayed a pooled DOR of 282.04 and AUROC of 0.919 for the seven best performing models including the studies developed for short-term prediction. Five short-term prediction studies from the “non” ad-hoc group displayed a pooled DOR of 21.45 and AUROC of 0.856. However, heterogeneity among studies varied from moderate to high, with a leave-one-out sensitivity analysis of each study showing significant changes in DOR.
Outcomes and Implications
Adverse cardiac outcomes such as sudden cardiac arrest are often preceded by VAs. However, the current diagnostic method used to identify VAs is based on left-ventricular dysfunction, which is absent in the majority of out-of-hospital sudden cardiac arrest. Utilizing AI to predict VA based on data from non-invasive testing like ECG and wearable sensors has the potential to become a vital tool for clinicians to provide early preventative personalized care for patients indicated with malignant VA. While meta-analysis displayed a high prediction performance, the majority of analyzed studies used small ad-hoc datasets leading to wide heterogeneity of the studies. Future research will need to focus on expanding these data sets, developing ways to display model explainability, and assessing the clinical integration and benefit of AI use.