AI Reveals Hidden Heart Failure and Valve Disease from Routine ECG in Seconds

At the European Society of Cardiology (ESC) conference in Munich, Imperial College London presented an AI model that extracts hidden disease signals from a standard 10-second electrocardiogram (ECG) in under two seconds. Tested on approximately 67,000 U.S. patients, the model identified 81% of heart failure cases and 90% of valve disease cases.

Large-Scale Training and Hidden Pattern Discovery

The team trained on over 1.6 million ECGs from Brazil linked to clinical histories, supplemented by millions of U.S. ECGs. Beyond cardiovascular diseases, the models also detected signals associated with diabetes and kidney disease. For cardiac conditions, diagnostic accuracy reached 83–93%; for non-cardiovascular diseases, 70–80% — all validated on international datasets from a single 10-second ECG.

The core insight: cardiac electrical activity is influenced by myocardial structure, cardiac load, pump function, and systemic diseases through complex physiological pathways. Clinicians recognize established waveform features, but AI can compare massive ECG libraries against subsequent clinical outcomes to uncover subtle patterns that do not appear as obvious abnormal waveforms.

The development process followed a stepwise validation: first stratifying patients into high- and low-risk groups, then estimating individual mortality risk, and finally predicting specific cardiac and non-cardiac diseases.

Clinical Translation via Cardiovolt.ai

In June 2026, the Imperial spin-out Cardiovolt.ai secured £1.4 million to bring the AI-ECG analysis layer into routine clinical workflows. The initial target is heart failure and valve disease: if the AI flags a risk signal, the patient proceeds to echocardiogram for confirmation.

Reference: https://www.bhf.org.uk/what-we-do/news-from-the-bhf/news-archive/2026/june/superhuman-ai-powered-ecgs-move-a-step-closer-to-clinical-use

Expanding Scope: Diabetes Prediction and Foundation Models

A 2026 study in European Heart Journal – Digital Health introduced AIRE-DM , an AI-ECG model predicting future type 2 diabetes risk in non-diabetic individuals. Developed and validated across multiple cohorts including the UK Biobank, it showed particular value for people not yet in the pre-diabetes range by traditional risk scores. This illustrates a paradigm shift: treating the ECG as a reusable data source that can reveal systemic metabolic states, not just a one-time cardiac snapshot.

On the foundational research front, Ahmed El-Medany presented an ECG-language multimodal contrastive pretraining model at ESC 2026, aiming to detect structural heart disease from 12-lead ECGs. The foundation model approach seeks to learn general representations from large-scale ECG and clinical data, then transfer to diverse downstream tasks.

Reference: https://esc365.escardio.org/presentation/326003?resource=abstract

Future Directions and Portable Monitoring

A 2026 review in European Heart Journal – Digital Health outlined three priorities:

Stronger model interpretability

Multimodal foundation models

Rigorous clinical implementation studies

To extend reach beyond hospital equipment, the team is also exploring portable ECG combined with AI and, in BHF-supported work, sensor-embedded smart shirts for continuous long-duration ECG-like signal recording.

Reference: https://www.theguardian.com/technology/2026/aug/31/superhuman-ai-tool-spots-heart-disease

Conclusion

Imperial College's trajectory moves from large-scale ECG-clinical data training, through clinical validation via Cardiovolt.ai, toward foundation models and wearable deployment. For clinicians, a familiar ECG waveform may still contain patterns not fully captured by human-defined rules — from hidden heart failure and valve disease to future diabetes risk — and researchers continue testing the limits of what a 10-second signal can reveal.

Code example

来源：ScienceAI
本文
约2000字
，建议阅读
5
分钟
AI从常规心电信号中捕捉人类难以识别的信息。