Researchers from Imperial College London announced at the European Society of Cardiology (ESC) 2026 congress that their artificial‑intelligence system, dubbed AIRE, identified 81 % of heart‑failure cases and 90 % of significant valvular disease from a standard 12‑lead electrocardiogram (ECG) in a validation cohort of about 67 000 patients.
Validation performance across a multinational cohort
The core claim comes from a Les Numériques article that cites the ESC presentation and the underlying European Heart Journal paper. In the validation cohort – defined as the group of patients on which the model’s performance was measured – the AI achieved an 81 % detection rate for heart‑failure and a 90 % detection rate for clinically relevant valvular disease. The valve‑leak risk model, a separate component of AIRE, produced a concordance (C‑index) ranging from 69 % to 79 % when evaluated on 34 214 patients from Beth Israel Deaconess Medical Center in Boston.
| Metric | Value | Cohort |
|---|---|---|
| Heart‑failure detection | 81 % | ≈67 000 patients |
| Valvular disease detection | 90 % | ≈67 000 patients |
| Valve‑leak risk C‑index | 69‑79 % | 34 214 Boston patients |
| Source: Les Numériques – AI detects cardiac disease from ECG in <2 seconds | ||
All three figures are explicitly tied to the validation period described in the source: the 81 % and 90 % rates refer to the ~67 000‑patient cohort, while the C‑index range refers to the Boston validation set of 34 214 patients.
Training data that underpins the model
AIRE’s training regime combined massive ECG repositories from multiple continents. The model was trained on 1.6 million Brazilian ECG recordings and “several million” additional US recordings, according to the Les Numériques excerpt. For the valve‑leak risk component, the team used 988 618 paired ECG‑echocardiogram records sourced from Zhongshan Hospital in Shanghai, China. These training sets were assembled before 2025, as indicated in the timeline.
By leveraging such heterogeneous data, the researchers aimed to capture subtle electrical signatures that are invisible to the human eye but predictive of structural heart disease. The French‑language source notes that the AI extracts “signaux invisibles à l’œil” from a routine ECG, a capability that could bypass the months‑long wait for an echocardiogram.
Timeline from data collection to clinical presentation
- Pre‑2025: Model training on Brazilian, US and Shanghai ECG datasets.
- 2025‑2026: Validation on two independent cohorts – 34 214 patients in Boston and a larger ~67 000‑patient cohort spanning multiple sites.
- Nov 2025: Valve‑leak risk results (C‑index 69‑79 %) presented in the European Heart Journal.
- 2026: ESC congress presentation of heart‑failure detection (81 %) and valvular disease detection (90 %).
Both the heart‑failure and valve‑disease results remain at the conference‑presentation stage; they have not yet been subjected to formal peer review.
Potential clinical impact and remaining questions
If the performance holds in prospective studies, AIRE could become a rapid triage tool for general practitioners and emergency‑department physicians. A standard 12‑lead ECG can be recorded in seconds, and the AI analysis reportedly takes less than two seconds, according to the source. This speed could allow clinicians to flag patients who need urgent echocardiography, shortening diagnostic pathways that currently involve waiting weeks for imaging.
However, the packet flags several uncertainties. The results are based on retrospective validation cohorts; prospective performance in a real‑world clinical workflow has not yet been demonstrated. The technology has not undergone peer review, meaning the methodology and statistical robustness have not been independently vetted. Moreover, the source does not provide information on false‑positive rates, which are critical for assessing the risk of over‑referral.
Regulatory clearance is also pending. In the United States, the Food and Drug Administration would likely require a de‑novo classification or a 510(k) pathway, depending on the intended use. In Europe, a CE mark under the Medical Device Regulation would be necessary before the model could be marketed. No such approvals are mentioned in the packet.
Who stands to benefit and next steps
The immediate beneficiaries would be health systems that struggle with echocardiography capacity, such as the National Health Service (NHS) in the United Kingdom, where a prospective NHS study has just begun, according to the commission brief. If the AI can reliably pre‑select high‑risk patients, it could reduce the backlog for echocardiograms and accelerate treatment decisions for heart‑failure and valve disease, conditions that together affect millions worldwide.
Future research will need to address three key gaps:
- Prospective validation in diverse clinical settings, including primary‑care clinics and emergency departments.
- Full reporting of specificity and false‑positive rates to gauge the net clinical benefit.
- Regulatory pathways and reimbursement models that would allow the technology to be deployed at scale.
Until those questions are answered, the AI’s impressive detection rates remain a promising but unproven step toward faster cardiac diagnosis.