ESC Congress 2026 Abstract · Oct 26, 2026
Transformer-based ascertainment of sudden cardiac death in electronic health records: development, validation and utility across national datasets
By Petrazzini B, Ahmed N, Lian J et al.
Background: Sudden cardiac death (SCD) accounts for 15-20% of global deaths1, yet it is poorly captured in death registries because ascertainment requires timeframe and origin of death, which are often missing. This limits statistical power for research. Electronic health records (EHR) could enable large-scale SCD research, however the latest phenotype of SCD in EHR shows suboptimal performance2. Transformer models, which can learn disease trajectories from EHR3, could help identify SCD in EHR. Purpose: To develop and validate a transformer model for ascertainment of SCD (aSCD) in EHR and evaluate its utility across national datasets. Methods: We trained a transformer model using diagnostic and procedural codes from >7 million individuals in the Clinical Practice Research Datalink (CPRD) and fine-tuned it using 1,084 definite cases of SCD and 6,142 curated controls. We evaluated the area under the curve (AUROC) and classification metrics using a held-out test set from CPRD and an external validation set from the UK Biobank (UKB). To define predicted SCD we ranked individuals based on aSCD and ascertained the top n matching the expected incidence of SCD in the UK (46 cases per 100,000 person-year)4. Then, we tested whether aSCD can capture suddenness and cardiac origin of death in CPRD, the defining attributes of SCD. Finally, we used ascertained cases of SCD to evaluate whether aSCD can inform 1) under-provision of automated external defibrillators (AED), 2) pharmacovigilance and 3) SCD risk marker discovery in CPRD, UKB and Mount Sinai Data Warehouse (MSDW). Results: aSCD showed AUROC of 0.95 and 0.90 in the held-out test and external validation sets, respectively. Furthermore, aSCD improved precision of SCD ascertainment by 49% at similar recall in both cohorts compared to the ESCAPE-NET Consortium's phenotype2. The tool tracked signatures of suddenness of death as shown by significantly higher scores in sudden versus expected types of death (P<0.0001) (Figure 1A). aSCD also captures cardiac origin of death as shown by an explainability analysis dominated by cardiovascular codes, strong discrimination between cardiac and non-cardiac causes of cardiac arrest (AUROC=0.95), and enrichment of SCD-related causes of death (β=0.49, P<0.0001) (Figure 1B-D). We then evaluated three use-cases for aSCD. First, predicting SCD incidence across regions of England revealed a disproportionate distribution of AED (concentration index=-0.14), with northern regions being the most under-provisioned (Figure 2A). Second, aSCD identified 7 out of 11 (63.64%) drug classes with known risk of QT-prolongation (Figure 2B). And third, aSCD tracked with known markers of SCD risk such as lower left ventricular ejection fraction (β=-4.2, P<0.0001) and longer corrected QT interval (β=14.5, P<0.0001), among others, in a meta-analysis of UKB and MSDW (Figure 2C). Conclusion: A transformer model improves SCD phenotyping in EHR enabling scalable SCD research.