Machine learning of electrophysiological signals for the prediction of ventricular arrhythmias: systematic review and examination of heterogeneity between studies

BACKGROUND: Ventricular arrhythmia (VA) precipitating sudden cardiac arrest (SCD) is among the most frequent causes of death and pose a high burden on public health systems worldwide. The increasing availability of electrophysiological signals collected through conventional methods (e.g. electrocardiography (ECG)) and digital health technologies (e.g. wearable devices) in combination with novel predictive analytics using machine learning (ML) and deep learning (DL) hold potential for personalised predictions of arrhythmic events. METHODS: This systematic review and exploratory meta-analysis assesses the state-of-the-art of ML/DL models of electrophysiological signals for personalised prediction of malignant VA or SCD, and studies potential causes of bias (PROSPERO, reference: CRD42021283464). Five electronic databases were searched to identify eligible studies. Pooled estimates of the diagnostic odds ratio (DOR) and summary area under the curve (AUROC) were calculated. Meta-analyses were performed separately for studies using publicly available, ad-hoc datasets, versus targeted clinical data acquisition. Studies were scored on risk of bias by the PROBAST tool. FINDINGS: 2194 studies were identified of which 46 were included in the systematic review and 32 in the meta-analysis. Pooling of individual models demonstrated a summary AUROC of 0.856 (95% CI 0.755–0.909) for short-term (time-to-event up to 72 h) prediction and AUROC of 0.876 (95% CI 0.642–0.980) for long-term prediction (time-to-event up to years). While models developed on ad-hoc sets had higher pooled performance (AUROC 0.919, 95% CI 0.867–0.952), they had a high risk of bias related to the re-use and overlap of small ad-hoc datasets, choices of ML tool and a lack of external model validation. INTERPRETATION: ML and DL models appear to accurately predict malignant VA and SCD. However, wide heterogeneity between studies, in part due to small ad-hoc datasets and choice of ML model, may reduce the ability to generalise and should be addressed in future studies. FUNDING: This publication is part of the project DEEP RISK ICD (with project number 452019308) of the research programme Rubicon which is (partly) financed by the 10.13039/501100003246Dutch Research Council (10.13039/501100003246NWO). This research is partly funded by the 10.13039/100019741Amsterdam Cardiovascular Sciences (personal grant F.V.Y.T).