Can an ECG AI explain its risk predictions?
A language-model framework reached AUROCs of 0.98 for a composite cardiovascular outcome and 0.86 for mortality in an internal 10,000-patient holdout. The evidence comes from one Ohio health system, and its explanations were judged for plausibility rather than proved correct.
Editorial responsibility: The Impact of AI Editorial Desk · Report a factual concern
At a glance
- 1The final modelling cohort contained 60,000 adults from one Northeast Ohio health system: 40,000 for training, 10,000 for internal validation and 10,000 for a holdout test.
- 2The all-feature model reported AUROCs of 0.98 for major adverse cardiovascular events and 0.86 for all-cause mortality, outperforming XGBoost, a conventional transformer and a multilayer perceptron on the same partitions.
- 3Four physicians rated explanations for 20 positive predictions highly for plausibility and coherence, but that small exercise measured expert perception, not whether the proposed causal pathways were true.
Research topic
Whether a language model can predict later cardiovascular outcomes from structured ECG measurements while producing explanations that clinicians regard as physiologically plausible
The study separates prediction from explanation—but does not prove either will improve care
An electrocardiogram is inexpensive, fast and already embedded in routine care. That makes it an attractive source for models trying to identify risks that may not be obvious in a clinician's visual reading. The new study presents CLAIRE, a two-stage framework built around a distilled DeepSeek-R1/Qwen3 language model. One stage predicts all-cause mortality and a composite major adverse cardiovascular event, or MACE; the other ranks influential ECG measurements and turns them into clinician-readable pathways.
The paper's strongest result is technical, not clinical. On an internal holdout, the full model discriminated people who later met an outcome from those who did not. It also produced explanations that a small group of physicians regarded as plausible. The study did not test a hospital deployment, compare decisions made with and without the system, or show that an explanation caused a safer diagnosis or better outcome. Those are later questions, and the difference matters for patients who may read a high AUROC as proof of individual benefit.[1]
The Impact Brief · Free
Follow the evidence in health & life sciences.
Get a five-minute weekday briefing on what changed, why it matters and where the evidence comes from. Choose the topics you care about.
The source population was large, but the model used one system and one selected ECG per person
Researchers began with 795,778 ECGs from 599,606 adults treated in University Hospitals Health System in Northeast Ohio between January 2015 and March 2025. They linked the records to the health system's clinical data and the Ohio death registry. After excluding incomplete records and 5,796 patients whose selected ECG fell on the same day as an event, 170,104 people were eligible. The modelling work used the first 60,000 eligible patients rather than the full cohort.
Each person contributed one index ECG. For someone who experienced MACE or died, the researchers selected the closest ECG before the event; for someone without an event, they selected the most recent follow-up ECG. MACE combined non-fatal myocardial infarction, non-fatal stroke and all-cause death. In the 60,000-person set, 22.8% met the MACE definition and 18.4% died. Median follow-up was 4.0 years; median time to MACE was 2.3 years and to mortality 3.1 years. The different rules for choosing case and non-case ECGs are another reason to test a locked model prospectively.[1]
The language model read structured measurements, not the raw waveform
The input was 636 numerical ECG features—53 measurements from each of 12 leads—plus age and gender. Values came from the GE MUSE system and included familiar electrophysiological quantities such as amplitudes, durations and intervals. The team converted the matrix into structured text, then fine-tuned the eight-billion-parameter DeepSeek-R1-0528-Qwen3 model to predict both outcomes. The full version required about 13 days on one NVIDIA A100 GPU and roughly 60 to 90 seconds of inference per patient, according to the paper.
CLAIRE's second stage used Gradient × Input attribution over token embeddings to rank features. The researchers retrained smaller variants with roughly 73 or 50 endpoint-specific ECG variables, still including age and gender. That design makes the system different from waveform foundation models that learn directly from the electrical trace. Structured features may be easier to inspect and store, but they also inherit the extraction system's assumptions and can miss subtle morphology that a waveform model might detect.[1]
The holdout was independent of training, but it was not an external hospital
Of the 60,000 patients, 40,000 were used for training, 10,000 for internal validation and 10,000 for final holdout testing. Because each patient supplied only one ECG, the unit of analysis was a person rather than repeated recordings. The authors estimated 95% confidence intervals with 1,000 bootstrap resamples of the holdout and reported calibration plots, intercepts, slopes and Brier scores as well as discrimination. This is a stronger evaluation than quoting training accuracy.
It remains internal validation: all partitions came from the same health system, period and ECG infrastructure. The all-feature model reported AUROC 0.98 for MACE, with a 95% confidence interval of 0.97 to 0.99, and 0.86 for mortality, with a confidence interval of 0.83 to 0.90. Its MACE and mortality Brier scores were 0.048 and 0.132. Those results show strong separation inside this dataset; they do not establish calibration or clinical thresholds in another hospital, country or device environment.[1]
Conventional comparators strengthen the technical result, while the composite endpoint needs care
The team trained XGBoost, a multilayer perceptron and a conventional transformer using the same 636 ECG inputs, age, gender and data partitions. For MACE, their AUROCs were 0.84, 0.65 and 0.83, compared with CLAIRE's 0.98. For mortality, the corresponding figures were 0.74, 0.51 and 0.82, compared with 0.86. The roughly 73-feature CLAIRE variant retained AUROC 0.91 for MACE and 0.86 for mortality; the 50-feature version reached 0.90 and 0.76.
That is a meaningful internal benchmark, but it does not reveal how the system compares with established clinical risk scores or a cardiologist using the full chart. MACE also includes death, so its apparently exceptional discrimination should not be interpreted as an equally strong ability to predict each non-fatal event. The study reports discrimination for the composite and for mortality, not separate external estimates for myocardial infarction and stroke. Decisions would need endpoint-specific calibration, a prespecified operating threshold and evidence about false alarms and missed events.[1]
Twenty explanations were scored for plausibility, not objective truth
For the explanation exercise, the researchers randomly selected 50 people from the holdout, then randomly selected 20 cases with a positive event prediction. Four physicians—two cardiologists and two internists, one with cardiology training—rated each pathway across five domains. Across 80 ratings per domain, mean scores were 4.69 out of five for physiological plausibility, 4.47 for support in prior evidence, 4.74 for logical coherence and 4.62 for clinical utility. Novelty was lower and more variable at 3.84.
Those ratings show that the prose often sounded clinically coherent to experienced readers. They do not verify that a named ECG feature caused an intermediate condition or that the pathway faithfully represents how the model arrived at its probability. The authors call the outputs hypothesis-generating and acknowledge that no ground truth exists for many explanations. Reviewers saw only 20 positive cases, and the study did not test negative predictions, misleading rationales at scale or whether explanations make clinicians appropriately sceptical when the model is wrong.[1]
Practical value depends on transportability, workflow and access
A reliable system could turn measurements already produced during routine ECG acquisition into an additional risk signal, helping clinicians identify people who may need closer review. Compact feature sets might reduce compute and make local, firewall-restricted deployment more feasible. Explanations could also help a clinician decide whether a warning aligns with known physiology instead of presenting only an opaque score. Those are plausible benefits, not measured outcomes from this study.
The constraints are immediate. The final 60,000 patients were the first eligible records rather than a random sample and included a larger share of Black patients than the 170,104-person source cohort, 42.9% versus 26.5%. The data came from a single region and GE MUSE-derived features. The patient dataset and fine-tuned models are not public because of institutional review-board restrictions; the authors say qualified researchers may request access, and code and prompts are available on request. The paper declares no competing interests, and no separate funding statement appears in the article text reviewed for this analysis.[1]
What would change the assessment
The decisive next test is external and prospective. A locked model should be run across hospitals with different populations, ECG machines and clinical practices, with outcomes defined before analysis and no retuning on the test sites. Researchers should report calibration, sensitivity, specificity, positive and negative predictive values at clinically chosen thresholds, performance for myocardial infarction and stroke separately, and subgroup results with enough observations to be informative. Comparison with established risk scores and clinicians using the same available information would show whether the language-model architecture adds value.
A deployment study should then test what happens to people: whether alerts change appropriate follow-up, whether explanation text improves or distorts clinician judgment, whether false positives create unnecessary investigation, and whether any gains persist over time. Explanation evaluation needs many more positive and negative cases, blinded reviewers, explicit error categories and a method for testing faithfulness rather than fluency alone. Until those results exist, CLAIRE is a notable internally validated research system—not a clinical instruction, a causal account of disease or a ready-made population screening tool.[1]
What this means for people
- Patients should not interpret a high internal AUROC as a personalised forecast or reason to change treatment without clinical assessment.
- Clinicians may gain a reviewable risk signal from measurements already collected during an ECG, but external validation and workflow testing are prerequisites for routine use.
- Health systems would need to monitor calibration, false alerts, subgroup performance and the effect of explanation text on decisions after deployment.
Global context
The study is based entirely in Northeast Ohio, even though ECGs are globally available and cardiovascular disease is a leading burden across very different health systems. Structured features may make replication easier than raw-waveform pipelines, but device software, disease prevalence, follow-up, access to care and population characteristics can all shift performance. International relevance therefore depends on locked external tests in regions with different vendors and patient mixes, followed by local clinical-utility studies rather than simple reuse of the reported internal thresholds.
What the evidence does not yet show
- All development, validation and testing data came from one Northeast Ohio health system; no external institution or prospective cohort was used.
- The final 60,000-patient dataset was the first eligible subset rather than a random draw and differed from the source cohort in racial composition.
- Case and non-case index ECGs were chosen under different timing rules, and the study did not test performance at a fixed prospective clinical timepoint.
- The MACE endpoint combined non-fatal myocardial infarction, non-fatal stroke and all-cause death, so composite discrimination does not establish equal performance for each event.
- Four physicians evaluated explanations for only 20 positive predictions, measuring perceived plausibility rather than causal correctness or model faithfulness.
- Patient data, fine-tuned models and pipeline code were not publicly released; access is available only on request and subject to institutional review.
What to watch next
- External validation across health systems, countries and ECG vendors without site-specific retuning.
- Prospective comparison with established risk scores and clinicians using the same information.
- Endpoint-specific calibration and clinically chosen thresholds, including false-positive and false-negative consequences.
- Larger blinded tests of whether explanations are faithful, not merely plausible, across both correct and incorrect predictions.
- Evidence that use changes appropriate care and patient outcomes without producing avoidable investigation or inequity.
Living evidence record
Impact record IAI-09952XH
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
5 October 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Scientific Reports published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 5 October 2026
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Continue the story
Related reporting
Health & Life Sciences
Does 99% MRI accuracy mean an Alzheimer’s diagnostic is ready?
No. A peer-reviewed model classified images from one 6,400-image Kaggle collection with reported 99% test accuracy, but it has no external clinical validation and the published test counts do not reconcile cleanly.
8 min · 2 sources
Health & Life Sciences
Can routine tissue slides reveal a tumour’s molecular subtype?
A graph-constrained model separated two pancreatic-cancer subtypes reasonably well when transcriptomic signals were clear, but performance weakened sharply in 361 ambiguous cases and did not improve overall ranking over simpler image baselines.
9 min · 2 sources
Health & Life Sciences
Can an AI risk message raise flu vaccination?
Three preregistered US field trials found that telling algorithm-identified patients they were at high risk produced a small absolute rise in flu vaccination. Naming the algorithm or explaining its reasons did not add a measurable benefit.
9 min · 5 sources
Reader commentary
Add evidence, experience or a question
No account is required. Reader notes are published after a brief civility, relevance and safety check; disagreement is welcome.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.