The use of an artificial intelligence (AI)-based tool results in a 31-percent enhanced precision in sudden cardiac death (SCD) phenotyping in electronic heart records (EHR), a recent study has shown.
This AI-based model “improves ascertainment of SCD in EHR, capturing signs of suddenness and cardiac origin of death,” said lead study author Dr Ben Omega Petrazzini, University of Oxford, Oxford, UK, adding that this tool “shows utility across research fields.”
The AI-based model for the ascertainment of SCD (aSCD) demonstrated an area under the curve (AUC) of 0.95 (95 percent confidence interval [CI], 0.90–0.97) and a positive predictive value (PPV) of 0.86 (95 percent CI, 0.83–0.89) in the test set. [Petrazzini B, et al, EHRA 2026]
Signatures of suddenness of death were also tracked by the tool, as seen in the significantly higher scores in sudden vs expected types of death (p<0.0001).
Likewise, the AI-based model captured the cardiac origin of death through the explainability analysis dominated by cardiovascular codes, strong discrimination of cardiac vs noncardiac causes of cardiac arrest (AUC, 0.95, 95 percent CI, 0.92–0.98), and enrichment of SCD-related causes of mortality (β, 0.49; p<0.0001), according to Petrazzini.
Ascertainment of SCD with the help of AI also discovered a total of 138,387 cases of SCD in the UK following guidance from the ESCAPE-NET Consortium*, “a project aimed at improving understanding, prevention, and treatment of sudden cardiac arrest.” [https://tinyurl.com/ypnj4a8j]
“We then evaluated three use cases for aSCD,” Petrazzini said. “Analyses of regions in England revealed a mismatch between incidence of ascertained SCD and automated external defibrillator (AED) density, highlighting opportunities for redistribution of AED.”
In a pharmacovigilance study, the AI-based tool also pinpointed 10 out of 13 (76.9 percent) QT-prolonging drugs with hazard ratios ranging from 1.6 to 10.0, according to Petrazzini.
Furthermore, aSCD tracked with known SCD risk factors, such as lower left ventricular ejection fraction (β, –4.2; p<0.0001) and longer corrected QT interval (β, 14.5; p<0.0001), in a meta-analysis using the UK Biobank and Mount Sinai Data Warehouse (MSDW).
“An AI-based tool improves precision of SCD phenotyping in EHR by 31 percent (PPV of 86 percent for aSCD vs 55 percent for ESCAPE-NET),” Petrazzini said. “This can enable scalable SCD research in EHR with much needed applications to SCD prevention and health policy.”
Method
Petrazzini and his team trained an AI model using diagnostic and procedural codes of >7 million individuals in the Clinical Practice Research Datalink (CPRD) and fine-tuned using 1,084 definite cases of SCD and 6,142 curated controls. They also assessed the AUC in a test set designed to minimize label leakage and determined whether aSCD can detect suddenness and cardiac death origin in CPRD.
Subsequently, Petrazzini and colleagues used aSCD to retrospectively ascertain SCD matching the expected incidence rate in the UK. The authors also determined whether the tool could inform nationwide redistribution of AED, pharmacovigilance, and biomarker discovery in CPRD, UK Biobank, and MSDW.
“SCD accounts for ~15 to 20 percent of global deaths, yet it is poorly captured in national death registries because ascertainment of SCD needs timeframe and origin of death which are often missing,” said Petrazzini, noting that this could limit statistical power for research and informed health policy for SCD.
"Transformer-based AI models can potentially capture the timeframe and origin of death to improve ascertainment of SCD in EHR,” he added.
*European Sudden Cardiac Arrest network: towards Prevention, Education, New Effective Treatment