AI Reads the Heart’s Electrical Whispers to Predict Death and Cardiac Disasters

Every heartbeat writes an electrical signature, and for more than a century physicians have read that signature through the electrocardiogram, or ECG, to diagnose heart attacks, rhythm disorders, and structural heart disease. What the ECG has rarely been asked to do is predict the future. Now a systematic review published in PLOS ONE suggests that artificial intelligence can transform this routine, inexpensive recording into a prognostic instrument, one that flags patients at risk of dying or suffering major cardiovascular events with accuracy that clinical risk scores have struggled to match. The review, led by Elhaitham Yasir Awad Musa and Amar Yasir Awad Musa, pooled the evidence from thirteen studies encompassing roughly 5.9 million patients, offering one of the broadest syntheses yet of what AI-enhanced ECG can and cannot deliver.

The stakes could hardly be higher. Cardiovascular disease remains the leading cause of death worldwide, and while clinicians rely on established risk scores to estimate a patient’s outlook, those tools provide only modest predictive accuracy. The ECG, meanwhile, is recorded millions of times a day in hospitals and clinics, yet its prognostic potential has gone largely untapped. The reason is that the human eye can extract only a fraction of the information embedded in the waveform. Deep learning models, by contrast, can mine high-dimensional features from the raw electrical signal, detecting subtle patterns that correlate with outcomes as distant as death from any cause months or years down the line.

To map this emerging field, the researchers searched PubMed, the Cochrane Central Register of Controlled Trials, IEEE Xplore, and Google Scholar through November 2025. They included studies that used artificial intelligence or machine learning models fed by twelve-lead or single-lead ECGs to predict all-cause mortality, cardiovascular mortality, or major adverse cardiovascular events, the composite outcome researchers call MACE, in adults aged eighteen and older. The ECG had to serve as the primary input for the model. Two reviewers independently screened the records and extracted the data, and because the studies differed so widely in design, populations, and outcome definitions, the team conducted a narrative synthesis rather than a formal meta-analysis.

The thirteen studies that made the cut reveal a field converging on a common technical approach. The models predominantly used convolutional neural networks, or CNNs, applied directly to raw ECG waveforms. CNNs are deep learning architectures originally developed for image recognition, and they excel at detecting local patterns in sequential data. When applied to an ECG, the network learns to recognize morphological features across the P wave, QRS complex, and T wave, along with intervals and amplitudes that no human reader could systematically quantify. This end-to-end approach, in which the algorithm learns directly from the signal rather than from hand-engineered features, appears to be the dominant paradigm for AI-ECG prognostication.

The headline numbers are striking. Eleven of the thirteen studies reported all-cause mortality as an outcome, with eight of those focusing on short-term prediction windows ranging from thirty days to one year. Four studies addressed cardiovascular mortality specifically, and eight assessed MACE, split evenly between composite endpoints and individual components such as heart attack or stroke. Across the mortality studies, the discrimination of the models, expressed as the C-statistic, ranged from 0.70 to 0.89. A C-statistic of 0.5 represents performance no better than chance, while 1.0 represents perfect discrimination, so values in this range indicate models that meaningfully separate patients who will die from those who will survive, approaching the accuracy of some validated clinical tools and in some settings exceeding them.

That consistency across studies, populations, and prediction windows is what gives the review its force. The authors conclude that AI-ECG models demonstrate strong potential for predicting mortality and cardiovascular events, with performance that held up across different cohorts and clinical settings. Part of the appeal is scalability. An ECG is cheap, noninvasive, and already collected in enormous volumes, so a model that adds prognostic value to an existing test could be deployed widely without new equipment, new blood draws, or new patient burden. In principle, a routine ECG performed for any reason could double as a screening tool for hidden cardiovascular risk.

Yet the review is equally candid about the obstacles standing between promising algorithms and bedside practice. Definitions of the outcomes varied considerably from study to study, as did the durations of follow-up, which limits direct comparison across models and makes it difficult to know whether a C-statistic reported in one cohort would hold in another. Heterogeneity in ECG acquisition, preprocessing, model architecture, and validation design further complicates the picture. A model trained on one hospital’s ECG machines and patient demographics may falter when transferred to a different population, a problem known as dataset shift that has plagued machine learning in medicine more broadly.

The authors are explicit that prospective validation and standardization are essential before clinical implementation. Retrospective studies, in which models are trained and tested on data already collected, can overstate real-world performance. Prospective validation means deploying the model forward in time on new patients, ideally across multiple centers, to confirm that its predictions hold when the algorithm encounters data it has never seen. Standardization, meanwhile, would require the field to agree on common outcome definitions, uniform follow-up intervals, and consistent reporting metrics, so that clinicians and regulators can compare models on equal footing rather than sifting through incomparable results.

If those hurdles are cleared, the implications for cardiovascular medicine could be profound. A prognostic layer on the standard ECG could help clinicians identify patients who look stable on routine assessment but carry elevated risk of death or major cardiac events, prompting intensified therapy, closer follow-up, or advanced testing. It could also reshape triage in emergency departments, where short-term mortality prediction from a single-lead tracing could inform admission decisions within minutes. The review’s synthesis of nearly six million patients suggests the signal is real and reproducible; what remains is the disciplined work of proving that these algorithms can change outcomes, not just predict them.

Subject of Research: Artificial intelligence in cardiovascular prognostication

Article Title: Artificial intelligence-enhanced electrocardiography for predicting mortality and major adverse cardiovascular events: A systematic review and narrative synthesis

Article References: Musa, E. Y. A., & Musa, A. Y. A. (2026). Artificial intelligence-enhanced electrocardiography for predicting mortality and major adverse cardiovascular events: A systematic review and narrative synthesis. PLOS One, 21(10), e0360756. https://doi.org/10.1371/journal.pone.0360756

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360756

Keywords: artificial intelligence, electrocardiography, cardiovascular disease, deep learning, mortality prediction, major adverse cardiovascular events, systematic review, convolutional neural networks, PLOS ONE, prognostic modeling, Artificial, intelligence-enhanced