IPM Take
Sudden cardiac death exposes one of cardiology’s most uncomfortable gaps: the people who die are often not the people the system predicted. Today’s dominant gatekeeper, left ventricular ejection fraction, misses many patients and flags others who may never benefit from a defibrillator. This new Nature study suggests AI may uncover risk directly from the ECG, one of the cheapest and most universal tests in medicine. But prediction is not policy. Before this becomes screening, health systems need trials, regulation, workflow, reimbursement, equity safeguards and clear rules for who gets a defibrillator, monitoring or further evaluation.
Executive Summary
A Nature study published has applied deep learning to population-based ECG data from Region Halland in Sweden, linked to death certificates and electronic health records. The model identified a high-risk group representing 2.2% of the sample, with a 7.0% annual rate of sudden cardiac death.
That risk was higher than in the group identified by reduced left ventricular ejection fraction, or LVEF, the current biomarker most widely used to guide defibrillator decisions. Reduced LVEF identified 1.9% of the sample, with a 4.6% annual sudden cardiac death rate. Importantly, 86.1% of the model’s high-risk ECG patients were not flagged by LVEF.
The model was externally validated in a US health system, where it predicted ventricular tachycardia and ventricular fibrillation, and in a Taiwanese hospital registry, where it specifically predicted future arrhythmic cardiac arrests. The researchers also used a generative model to visualise the ECG waveform morphology linked to risk, describing an apparently previously unrecognised biomarker visible in lead aVL.
The finding is potentially disruptive, but not yet practice-changing. Observational data suggested that high-risk patients with defibrillators had lower-than-expected mortality, but the authors state that a randomised trial will be crucial. The policy question is how to prepare for an AI-ECG tool that could expand risk detection far beyond today’s LVEF-based model.
Why it matters
- Policymakers and public authorities: AI-ECG screening could make sudden cardiac death prevention more scalable, but only if linked to evidence-based pathways, workforce capacity and registries.
- Regulators: An AI model that influences defibrillator decisions would require strong medical-device oversight, external validation, bias assessment, post-market monitoring and transparency about model updates.
- HTA bodies and payers: Cost-effectiveness will depend not only on prediction accuracy, but on whether high-risk ECG patients actually benefit from defibrillators, monitoring or other interventions.
- Clinicians and providers: ECG risk flags must not become unmanaged alerts. Cardiologists, electrophysiologists and primary care need referral criteria, shared decision-making tools and escalation pathways.
- Patients and advocates: A hidden risk signal can be empowering, but also frightening. Patients need clear communication about uncertainty, treatment options and the difference between risk prediction and diagnosis.
Sudden cardiac death is one of the most brutal failures of prediction in modern cardiology.
The technology to prevent some deaths already exists. Implantable cardioverter-defibrillators can detect and terminate lethal ventricular arrhythmias. The problem is knowing who needs one before the arrest happens.
For decades, cardiology has leaned heavily on left ventricular ejection fraction, or LVEF, to identify patients at high risk. It is practical, measurable and embedded in guidelines. It is also a blunt instrument. Many people who die suddenly never had low LVEF. Many people who receive defibrillators based on low LVEF never need a life-saving shock.
A new Nature study may shift that debate.
Researchers trained a deep-learning model on ECG data from Region Halland in Sweden, linking ECG waveforms to death certificates and electronic health records. The result was an AI-derived ECG biomarker for sudden cardiac death. In the Swedish lockbox dataset, the model identified a high-risk group comprising 2.2% of the sample with a 7.0% annual sudden cardiac death rate.
That is not a trivial signal. It was higher than the 4.6% annual sudden cardiac death rate in patients with reduced LVEF. More importantly, 86.1% of the AI-identified high-risk patients were not flagged by LVEF.
This is the political punchline: the current system may be missing many of the patients most likely to die.
The study did more than produce an algorithmic risk score. It tested the model across settings. In a US health system dataset from Sharp HealthCare, the model predicted ventricular tachycardia and ventricular fibrillation, the arrhythmias that often cause sudden death. In a Taiwanese hospital registry, it specifically predicted future arrhythmic cardiac arrests and performed much worse for non-arrhythmic arrests, supporting the idea that the model was detecting arrhythmic risk rather than general illness.
The model also produced a visible clue. By pairing the predictive model with a generative model of ECG waveforms, the researchers visualised a morphology associated with risk. They describe a slurred terminal portion of the R wave in lead aVL, replacing the sharper negative S wave seen in lower-risk morphs. The authors argue this feature is easily visible and robustly predictive, but not previously described in the literature.
That matters because explainability is not a luxury in medicine. A black-box risk score might be useful, but a visible biomarker gives clinicians and researchers something to interrogate, validate and challenge. It also creates a bridge between AI discovery and electrophysiology.
The study’s mechanistic hypothesis is early but interesting. The authors suggest that the waveform pattern could reflect increasingly disorganised depolarisation, potentially linked to diffuse myocardial fibrosis. Blinded review of cardiac MRI data in a small subset found more subtle diffuse late gadolinium enhancement among patients in the highest-risk predictions. That is hypothesis-generating, not definitive. But it points toward a bigger scientific opportunity: AI may not only predict who is at risk, but reveal new biology behind sudden cardiac death.
The policy implications are immediate.
First, this cannot become clinical practice by publication alone. The authors are explicit that a randomised trial in high-risk ECG patients will be crucial. That is the right standard. Many risk predictors look promising until they are asked the hard question: does acting on the prediction save lives without causing unacceptable harm?
For sudden cardiac death, that question is especially serious. A risk flag could lead to ambulatory monitoring, electrophysiology referral, cardiac MRI, genetic testing or defibrillator placement. Each has costs, risks and access implications. Defibrillators can save lives, but they are not harmless. They involve procedures, device complications, inappropriate shocks, psychological burden and long-term follow-up.
Second, the health system needs to decide what “high risk” means operationally. The Nature model identified a group whose annual risk exceeded trial-based benchmarks used to justify defibrillator placement. But guidelines are not currently built around AI-ECG risk. That creates a grey zone. Clinicians may face patients whose ECG says “high risk” but whose LVEF is normal or unknown. What happens next?
A serious pathway would need staged escalation: repeat ECG confirmation, review of clinical history, ambulatory rhythm monitoring, echocardiography, cardiac MRI where appropriate, medication optimisation, electrophysiology consultation and shared decision-making. Without that pathway, AI prediction becomes a dangerous orphan result.
Third, this raises reimbursement questions. ECGs are cheap. AI interpretation may be cheap at scale. But the downstream pathway is not. If a health system flags thousands of new high-risk patients, who pays for follow-up testing? Who pays for defibrillators? Who pays for monitoring? Who carries liability if a flagged patient is not referred, or if an unflagged patient dies?
Fourth, equity must be designed in from the beginning. ECGs are globally available in a way that cardiac MRI and advanced electrophysiology testing are not. That makes AI-ECG risk prediction potentially powerful for lower-resource settings. But the benefit is only real if high-risk patients can access follow-up care. A risk score without defibrillator availability is not precision prevention. It is a diagnosis of system inequality.
The model’s validation across Sweden, the United States and Taiwan is encouraging. But global deployment would still require validation across sex, age, ethnicity, comorbidity, device manufacturers, clinical settings and health systems. ECG acquisition quality varies. Population risk varies. Coding practices vary. The consequences of false positives and false negatives vary.
Fifth, regulation will matter. In Europe, AI-based software intended for medical purposes can fall under high-risk AI requirements and medical device rules. In the United States, the FDA has been developing pathways for AI-enabled software as a medical device, including predetermined change control plans for models that may evolve over time. A sudden-death prediction tool cannot be treated like a wellness algorithm. It would influence life-and-death clinical decisions.
This is where developers, regulators and health systems must resist the hype cycle. The question is not whether the AUC is impressive. It is whether the model is safe, generalisable, clinically actionable, equitable and beneficial when embedded in real care.
There is also a data governance issue. The Swedish model relied on large-scale linkage of ECGs, health records and death certificates. That kind of learning health system infrastructure is not available everywhere. Where it is available, patients and publics will need confidence that high-dimensional ECG and outcome data are used responsibly, securely and with appropriate oversight.
The most exciting part of this study is also the most politically uncomfortable. The ECG is old technology. It is cheap, routine and everywhere. If AI can extract hidden risk from something already sitting in health records, then many missed deaths may no longer be explained by lack of science. They may be explained by lack of implementation.
But cardiology should move carefully.
A new biomarker can change medicine only if it changes outcomes. The next step is not mass deployment. It is prospective testing, pathway design and policy preparation.
Sudden cardiac death prevention has been trapped for decades between too little prediction and too late action. This study suggests there may be a new signal hiding in plain sight.
Now the question is whether health systems can turn that signal into survival, without turning it into another inequitable AI experiment.
