Sudden cardiac death prevention in general population cohorts: prospects of machine learning

This editorial refers to ‘Machine learning-based prediction of sudden cardiac death in the general population using electronic health record data', Y . Youssfi et al., https://doi.org/10.1093/eurjpc/zwag087. Sudden cardiac death (SCD) accounts for 15–20% of deaths worldwide.1 In the decades following the landmark randomized trials establishing the survival advantage offered by implantable cardioverter defibrillator (ICD) therapy, SCD risk stratification has evolved minimally. Whilst it is accepted that SCD risk is elevated in defined populations such as those with ischaemic cardiomyopathy, the continued reliance on the left ventricular ejection fraction (LVEF) inherently prohibits the screening of those without known cardiac disease, paradoxically where the absolute rates of SCD are highest.1 The central dogma surrounding the use of a dichotomized LVEF for primary prevention ICD allocation is increasingly being acknowledged as imprecise, patients are being exposed to the procedural risk of ICD insertion despite the majority of ICDs never providing appropriate therapy during their lifespan.2 Novel approaches must be generated to provide an individualized SCD risk assessment. Machine learning (ML) offers a unique avenue of investigation, capable of integration into clinical systems to interrogate high-dimensional datasets. The utility of ML in SCD risk prediction was first summarized during the 2022 systematic review by Barker et al.3 illustrating the predictive performance of contemporaneous ML models relative to traditional regression analyses; despite remaining limited by small sample sizes and lack of robust external validation.3 The study by Youssif et al.4 represent a shift in the field of ML approaches for SCD. Authors generated a ML model aimed at identifying those at elevated risk of SCD within a general population based on electronic health record (EHR) data. During their case-control study, 12 338 SCD cases were identified from the Paris Sudden Death Expertise Centre (SDEC) registry and were matched by age, sex, and residential area to national medical records from the Système National des Données de Santé between 2011 and 2015. Model variables were selected by first incorporating all EHR medical codes up to 5 years prior to the SCD event without pre-selection. The model performed feature selection and retained the most important codes for SCD prediction. Overfitting was mitigated with hyperparameter optimization through cross-validation and feature selection. Of the six models used, the best performing (average AUC) during cross-validation was CatBoost; an ensemble method optimized for categorical data using iteratively layering of simple decision trees. New decision trees are added to fix inaccuracies of the preceding existing model; thus, refining the predictive ability in a step-wise manner known as gradient boosting. In the derivation cohort, the model achieved an AUC of 0.83 (95% CI: 0.82–0.84) with a sensitivity of 70% (95% CI: 68–71) and a specificity of 83% (95% CI: 81–83). Due to its iterative nature, overfitting of spurious data is a consideration for validity. External validation was performed within two additional cohorts. First, the Paris SDEC registry was analysed at a temporally distinct period between 2016 and 2020. Of the 11 620 SCD cases the model correctly predicted 69% of SCD events, AUC 0.81 (95% CI: 0.80–0.82), sensitivity 69% (95% CI: 67–70), and specificity of 79% (95% CI: 78–80). To examine the geographical generalizability, the Paris model was deployed on data from the Seattle-King County Cohort SCD cases using University of Washington Medicine EDW hospital-based population data between 2013 and 2021. Predictive performance metrics dropped considerably with an AUC of 0.66 (95% CI: 0.58–0.73), sensitivity of 56% (95% CI: 46–65), and a specificity of 74% (95% CI: 65–82). Recalibration of the Paris-trained model on the Seattle cohort did not improve performance, whereas full retraining on Seattle-King County data improved the AUC to 0.77 (0.73–0.81). Authors performed several sensitivity analyses to test the model’s robustness. Consistent performance was achieved across multiple demographic subgroups. Notably the model proved resilient to short term coding biases, including potential clinically irrelevant noise in the immediate period preceding SCD during a 30 and 90-day data washout period analysis. Since myocardial infarction (MI) represents the major general population-level risk factor for SCD, authors verified the model’s capacity to differentiate between these two closely linked endpoints. When the Parisian temporal validation cohort was divided into deciles of predicted SCD risk, rates of MI showed no associated increase, thus adding weight to the model’s specificity for SCD and not its co-founders. A drawback of Youssif’s 1:1 case-control design for model development and validation is the predicted probabilities cannot be directly interpreted as absolute risk estimates for population-level screening or clinical decision-making owing to the sampling fraction not reflecting the general population incidence of SCD. When the CatBoost model’s accuracy was assessed correcting for SCD prevalence seen in the general population, the limitations for screening were clearly apparent. When a sensitivity threshold of 95% was applied, 14% of the SCD cases were allocated as high risk with a false positive rate of 1.46%. This model is not positioned as a readily deployable screening tool for ICD-therapy; instead it offers a route of exploration to tackle the paradox where the highest absolute SCD prevalence occurs in those without known cardiac disease. A recent publication by Perry et al.5 also investigated a ML approach to tackle this paradox using a large quaternary-care dataset from the University of Washington Medicine cohort as seen during the geographical validation by Youssif. Similarly, they found a gradient boosting algorithm (XGBoost) produced the best AUC of 0.95 (95% CI: 0.94–0.95) with a sensitivity of 51% and a strong predictive ability maintained during external validation using the unseen cohort data, despite some performance degradation AUC 0.85 (95% CI: 0.84–0.87). Notably the Paris derived model exhibited lower performance when applied to this geographically distant population which was partially mitigated on recalibration, suggesting local health systems niches and practices have a significant impact on a model’s portability in the real world. Whilst the ML literature in SCD has expanded across different study populations, so too has the interest in increasing the dimensionality of the source data used for model derivation. In their 2023 study, Kolk et al.6 examined a ML based prediction of malignant arrhythmia using combinations of EHR data with non-discrete variables. This study encompassed more than 1000 primary prevention ICD patients with LVEF ≤35% from 12-lead-ECG recordings. Intriguingly the most predictive features of appropriate ICD-therapy were weighted towards ECG based metrics with a minority contribution from EHR data. This was shown as a modest accuracy difference with the inclusion of EHR data in the multi-modal XGBoost based model AUC 0.792 (95% CI: 0.746–0.841) compared with the model based solely on ECG data AUC 0.742 (95% CI: 0.689–0.797). In a later study Kolk et al.7 integrated short-axis cardiac MRI as an additional layer to 12 lead-ECG and EHR data7 during a smaller study of 289 patients with non-ischaemic cardiomyopathy and primary prevention ICD with LVEF ≤45% supporting the achievable benefits of adding advanced imaging data to ML models. However larger studies of this nature are needed to understand the performance of cardiac MRI in greater granularity for SCD prediction. Despite its potential offering, complex phenotyping with ML using advanced imaging and genetics is clearly limited in a large population setting. Youssif et al.4 have authored a timely study applying ML to predict SCD from large-scale electronic health records. Although their model showed guarded generalizability across populations, it illustrates the potential utility behind ML models, and the rapid progress made over a short period of time. The trajectory of ML for SCD risk prediction has evolved from small disease specific cohort’s analyses3 to large-scale population studies. Current sophisticated models better account for high-dimensional data with non-linear relationships. The interest in multi-modal approaches highlights the capability to develop an individual electrophysiological and structural risk profile from existing datasets. However, any future integration into population screening will depend on accuracy, scalability and digital equity (Figure 1). Caution remains where premature implementation into a field currently lacking prospective data would leave clinicians in a dilemma regarding an evidence-based approach to accurate SCD risk prediction and prevention. Considerations for the application of machine learning models for sudden cardiac death in a general population. Edward Y.M. Lau (Data curation, Resources, Writing—original draft [lead]), Xin Li (Data curation, Methodology, Resources, Writing—review & editing [supporting], Supervision [equal]), and G. Andre Ng (Conceptualization, Supervision, Writing—review & editing [lead], Data curation, Resources [supporting]) G.A.N. and X.L. are supported by British Heart Foundation Research Excellence Award (RE/24/130031), NIHR Leicester Biomedical Research Centre, Medical Research Council Biomedical Catalyst Developmental Pathway Funding Scheme (MR/S037306/1) and NIHR i4i grant (NIHR204553). E.Y.M.L. is supported by a Kusuma Trust Research Fellowship. G.A.N. is also supported by British Heart Foundation Programme Grant (RG/17/3/32774). The data underlying this article are available in the article.