AI-Enabled ECG Models Show Consistent Prognostic Signal Across Mortality and MACE Studies
Artificial intelligence is beginning to extract prognostic information from the electrocardiogram that conventional interpretation often leaves untouched. In a systematic review and narrative synthesis of 13 studies, AI-enabled ECG models repeatedly showed the ability to predict all-cause mortality, cardiovascular mortality, and major adverse cardiovascular events, with reported c-statistics for mortality prediction ranging from 0.70 to 0.89.
The finding matters because the ECG is already embedded in routine care, inexpensive, and widely scalable. A prognostic model that can operate on a test already collected in emergency departments, wards, clinics, and wearable-adjacent settings has a plausible path to clinical deployment, provided the evidence becomes more standardized and externally validated.
A Prognostic Signal Hidden in Raw Waveforms
The review focused on adults and included studies that used 12-lead or single-lead ECGs as the primary model input. Twelve of the 13 studies used deep learning architectures, most often convolutional neural networks, with one study comparing a ResNet-based model against an XGBoost approach. The core technical advantage is straightforward: these models do not depend on human-selected morphology alone. They process raw voltage-time data across leads and can detect low-amplitude, temporally distributed patterns that may reflect occult ventricular stress, ischemic burden, electrical remodeling, or systemic disease burden long before a clinician would infer risk from the tracing by eye.
The included literature was broad in geography and setting, spanning the United States, Canada, Taiwan, Switzerland, South Korea, the United Kingdom, and Brazil. Sample sizes ranged from several hundred ECGs to more than two million, giving the field unusual scale for a prognostic imaging-adjacent technology. Even with that spread, the review did not try to force the studies into a single pooled estimate. The outcome definitions, populations, follow-up windows, and performance metrics were too heterogeneous for a mathematically defensible meta-analysis.
Mortality Prediction Was the Most Consistently Studied Endpoint
Eleven studies reported all-cause mortality. Eight of those evaluated short-term mortality, defined across the literature as 30-day to 1-year follow-up, while five examined longer-term mortality, with some overlap between the two categories. Reported performance for mortality prediction was generally strong, with c-statistics falling between 0.70 and 0.89.
A few points stand out clinically. First, the model signal was present across both near-term and extended horizons, suggesting that the ECG may encode immediate physiologic instability as well as chronic substrate. Second, the strongest use case may not be broad population screening but risk stratification in already-encountered patients, where a routine ECG could add information to bedside judgment, lab data, and comorbidity burden. Third, the variation in endpoints makes it difficult to know whether a c-statistic near the upper end of that range came from a narrowly selected cohort, a highly enriched disease group, or a more general hospital population.
The review also identified four studies that reported cardiovascular mortality. The specific details of each model’s calibration, external transportability, and clinical integration were not uniform across studies, which limits direct comparison. Still, the consistency of direction across the mortality literature suggests that AI-ECG is detecting something more than random pattern recognition.
MACE Prediction Extends the Signal Beyond Death Alone
Eight studies addressed major adverse cardiovascular events, including four that assessed composite MACE outcomes and four that reported individual components such as myocardial infarction, heart failure, stroke, or ventricular arrhythmia. This distinction matters clinically because death is only one part of the cardiovascular burden that matters at the bedside. A model that flags impending infarction or decompensated heart failure may be useful even when mortality is not the immediate endpoint.
The biologic plausibility is strong. A surface ECG reflects conduction timing, chamber strain, repolarization abnormalities, prior infarction, and indirect signatures of structural remodeling. Deep learning may be detecting relationships among subtle QRS and T-wave features, interlead variance, and waveform morphology that mirror cumulative cardiovascular injury. Those patterns can plausibly correlate with future ischemic events, pump failure, arrhythmic death, and other downstream complications long before overt clinical deterioration becomes obvious.
One important limitation runs through the MACE literature: endpoint definitions were not standardized. Some studies used composite outcomes, while others tracked single components, and the constituent events were not always aligned across datasets. In practical terms, a model trained to predict one version of MACE may not behave the same way in a different institution using another definition.
Why the Review Stopped Short of Meta-Analysis
The authors did not pool the results, and the reason is methodologically sound. The studies were too different in baseline population, follow-up duration, endpoint definition, and statistical framework. Some relied on binary classification measures such as AUC, while others used time-to-event measures such as the C-index. Several large studies also drew from overlapping regional datasets, which raises the risk of double-counting patients and inflating apparent certainty.
The review therefore used a narrative synthesis and assessed risk of bias with PROBAST. That approach is slower to read than a pooled forest plot, but it is more honest about what the evidence can and cannot support. A combined summary would have suggested a precision the field does not yet possess.
Clinical Use Remains Ahead of the Evidence Base
The review points to real promise, but it also shows how far the field still is from routine implementation. Prospective validation remains essential, especially in external cohorts and in real-world workflows where prevalence, referral patterns, and ECG acquisition conditions differ from development datasets. Standardization of outcome definitions will matter just as much as model architecture.
At the bedside, the most realistic near-term role may be as an adjunctive triage tool rather than a stand-alone decision-maker. An ECG-derived risk score could eventually help identify patients who need closer follow-up, more intensive evaluation, or earlier preventive therapy. Yet that use case depends on reproducibility across hospitals and health systems, and the current literature remains uneven on that point.
Short-term and long-term prediction were both represented, but not in a way that allows clean cross-study comparison. The same is true for model transparency, calibration reporting, and integration into actual clinical decision-making. Until those gaps close, AI-ECG should be viewed as a promising prognostic technology with substantial technical momentum, not a finished clinical product.