IPM Take
Cardiology has largely moved beyond asking whether AI will enter clinical practice. At ESC Congress 2026, the harder question was whether its arrival will make cardiovascular care more equitable or simply make existing inequalities faster and harder to see. ESC itself framed AI as a major congress theme, spanning ECG interpretation, imaging, risk prediction and clinical decision support.
A 2026 JACC: Advances review shows why the distinction matters. Bias can emerge because certain populations are poorly represented in training data, because healthcare utilisation or cost is used as a misleading proxy for clinical need, or because a model developed in a well-resourced academic centre is deployed in a rural or otherwise different population without adequate validation. Even a dataset that appears balanced does not guarantee equitable performance if the model itself is optimised primarily for aggregate accuracy.
This is the political problem hiding inside the technical one: an algorithm can be highly accurate overall while being systematically worse for the patients the health system already serves least well.
Precision medicine cannot call that precision.
Executive Summary
At ESC Congress 2026 in Munich, a session chaired by Northwestern cardiologist Sadiya Khan focused on equity, ethics and privacy as AI becomes embedded in cardiovascular care. The message was not that AI should be rejected, but that clinical value cannot be assumed simply because a model performs well technically.
The concern is supported by a state-of-the-art review published in JACC: Advances in August 2026. Authors Aparna Kulkarni and Shyam Visweswaran identified several pathways through which generative AI can worsen cardiovascular disparities, including representation bias, measurement and labelling bias, deployment bias and model architecture bias. They propose stronger representation, transparency, accountability, community engagement, local validation and continuous recalibration across the AI lifecycle.
The implementation gap is already visible. The review cites U.S. data indicating that approximately 65% of hospitals use predictive models, while only 44% have assessed their models for bias. That matters because a system can perform strongly at population level while underperforming in racial and ethnic minorities, women, rural populations or socioeconomically disadvantaged groups.
Europe is simultaneously building a regulatory environment around healthcare AI and health data. The European Health Data Space entered into force in 2025 and is intended to enable more secure and structured use of diverse health data, although its major operational requirements will be introduced gradually over several years.
Why it matters
- HTA bodies: AI assessment cannot stop at average accuracy. Evidence will increasingly need to demonstrate external validity, subgroup performance, real-world clinical utility and whether performance deteriorates when an algorithm moves between hospitals, populations or health systems.
- Payers: Algorithms that improve efficiency for some patients while systematically missing others can shift rather than reduce costs. Coverage decisions may eventually need evidence on equity, workflow impact and downstream resource use alongside conventional performance metrics.
- Industry / innovation partners: Training on larger datasets is not enough. Developers increasingly face expectations for representative data, local validation, subgroup benchmarking, transparency and continued monitoring after deployment. Equity is becoming part of product quality, not an optional social-impact claim.
Artificial intelligence has become one of cardiology’s dominant technology stories. Algorithms can interpret ECGs, analyse imaging, predict deterioration and extract patterns from clinical records at a scale no individual clinician could reproduce. ESC Congress 2026 placed that transformation at the centre of its programme, but the discussion in Munich increasingly moved beyond technological capability toward accountability.
That change is overdue, because AI learns healthcare partly from the healthcare system that already exists.
If some populations are underrepresented in historical datasets, the model inherits that absence. If hospitalisation or healthcare spending is used as a proxy for disease burden, an algorithm may learn differences in access rather than differences in illness. If a model trained in a large academic centre is deployed in a rural hospital with different patients, disease prevalence and workflows, its original performance may not travel with it. The JACC: Advances review describes all three as routes by which apparently sophisticated AI can reproduce inequity rather than correct it.
The problem becomes particularly uncomfortable when aggregate performance looks good.
A model can achieve an impressive overall metric while making more mistakes in smaller patient groups. Optimising an algorithm for the average patient therefore risks creating a digital version of a familiar healthcare problem: the patients best represented in the system receive the most reliable care, while everyone else carries more uncertainty.
More data will not automatically solve the problem
The obvious response is to demand more diverse datasets, and that is necessary, but it is not sufficient.
The review argues that bias can also originate in model architecture, outcome definitions and deployment choices. That means equitable AI requires subgroup testing, local validation and continued recalibration after implementation, rather than a one-time accuracy score before launch.
Europe’s developing policy environment makes this increasingly relevant. The European Health Data Space is intended to create more structured, secure access to health data for care, research and innovation, while EU AI regulation places growing emphasis on data governance and risk management. Yet regulation cannot decide whether a particular algorithm genuinely improves outcomes for women, minority populations or patients in a smaller regional hospital. That still requires evidence.
Privacy adds another layer. Cardiovascular AI may use ECGs, imaging, electronic records, genetic information and longitudinal patient histories. As Khan highlighted at ESC 2026, extracting more value from those datasets creates an obvious governance question: who controls the data, who can access them and what safeguards remain once they enter complex AI systems?
None of this argues against AI in cardiology. It argues against confusing deployment with progress.
AI could help extend specialist knowledge to places where cardiologists are scarce, detect disease earlier and personalise cardiovascular care more effectively. But those benefits are not built into the technology by default. They depend on whose data shaped the model, where it was validated, how performance is measured and who remains accountable when it fails.
The next test for cardiovascular AI is therefore not whether it can outperform a clinician on another benchmark. It is whether it can improve care without making the patients already at the margins even easier to miss.
