---
title: "AI-enhanced ECG Predicts Mortality and MACE: Systematic Review Findings"
id: "plos-one-17-artificial-intelligence-enhanced-electrocardiography-for-predicting-mortality"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-17-artificial-intelligence-enhanced-electrocardiography-for-predicting-mortality"
content_type: "clinical_feed_article"
specialty: "Cardiology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0360756"
published_at: "2026-10-16T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# AI-enhanced ECG Predicts Mortality and MACE: Systematic Review Findings
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-17-artificial-intelligence-enhanced-electrocardiography-for-predicting-mortality
- **Specialty:** [Cardiology](https://medichelpline.com/clinical-feed/cardiology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0360756)
- **Published At:** 2026-10-16T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- Investigators conducted a systematic review of studies using surface **ECG** (12-lead or single-lead) as the primary input for AI/ML models to predict all-cause mortality, cardiovascular mortality, or **MACE** in adults. - The review followed PRISMA 2020 and was registered with PROSPERO (CRD420251244907); searches covered PubMed, Cochrane CENTRAL, IEEE Xplore, and Google Scholar through November 25, 2025. - Thirteen studies met inclusion criteria, representing roughly 5.9 million patients across healthcare settings in multiple countries and published from 2016–2025. - Eleven studies reported all-cause mortality outcomes (eight focused on short-term horizons up to 1 year; five reported long-term outcomes), four reported cardiovascular mortality, and eight assessed MACE (four composite, four individual components). - Models predominantly used convolutional neural networks and ResNet variants trained on raw waveform inputs; one study compared ResNet with an XGBoost approach. - Reported discrimination for mortality endpoints ranged roughly from AUC/C-index 0.70 to 0.89 across studies, with several models exceeding conventional clinical scores in head-to-head analyses. - Representative high-performance results: Sun et al. (AUC 0.843 for 30-day mortality, 0.812 for 1-year); Raghunath et al. (AUC 0.88 for 1-year); Lin et al. (C-index 0.860 for 1-year; other Lin models AUC 0.89). - Cardiovascular mortality models reported C-indices/AUCs in the 0.78–0.89 range, with Tsai, Lin (ECG-surv), Sau, and Hughes studies indicating strong discrimination in internal and some external validations. - Heterogeneity in outcome definitions, follow-up durations, model architectures, and overlapping datasets precluded statistical pooling; authors performed a narrative synthesis and assessed risk of bias with PROBAST. - The review authors concluded that **AI-ECG** shows consistent prognostic potential but emphasized the need for prospective validation, standardized endpoints, and broader external evaluation before clinical deployment.
## Clinical Analysis & Structured Key Points
Artificial intelligence-enhanced electrocardiography for predicting mortality and major adverse cardiovascular events: A systematic review and narrative synthesis | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Reader Comments Figures Figures Abstract Cardiovascular disease (CVD) remains the leading cause of death worldwide. While clinical scores provide modest predictive accuracy, the electrocardiogram (ECG) is rarely leveraged for prognosis. Artificial intelligence (AI) and deep learning now enable extraction of high-dimensional features from ECG waveforms for predicting cardiovascular outcomes. Our objective was to systematically evaluate the prognostic performance of AI-enabled ECG models for predicting all-cause mortality, cardiovascular mortality, and major adverse cardiovascular events (MACE). We searched PubMed, Cochrane CENTRAL, IEEE Xplore, and Google Scholar through November 2025 for studies using AI/ML models with 12-lead or single-lead ECGs to predict all-cause mortality, cardiovascular mortality, or MACE in adults (≥18 years old). Eligible studies included adults (≥18 years) using ECG as primary input. Two reviewers independently screened and extracted data. Narrative synthesis was conducted due to heterogeneity. Thirteen studies were included (~5.9 million patients). Eleven reported all-cause mortality (8 short-term: 30-day to 1-year), four reported cardiovascular mortality, and eight assessed MACE (4 composite MACE, 4 individual components). Models predominantly used convolutional neural networks [CNNs] on raw ECG waveforms. C-statistics ranged 0.70–0.89 for mortality prediction. Definitions and follow-up durations varied, limiting direct comparison. AI-ECG models demonstrate strong potential for predicting mortality and cardiovascular events with consistent performance and scalability. However, prospective validation and standardization are essential before clinical implementation. Citation: Musa EYA, Musa AYA (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 Editor: Chiara Lazzeri, Azienda Ospedaliero Universitaria Careggi, ITALY Received: May 14, 2026; Accepted: September 30, 2026; Published: October 16, 2026 Copyright: © 2026 Musa, Musa. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: All relevant data are within the paper and its Supporting information files. Funding: The author(s) received no specific funding for this work. Competing interests: The authors have declared that no competing interests exist. 1. Introduction Cardiovascular disease (CVD) remains the leading global cause of death [ 1 ]. Early risk identification is critical for preventive strategies [ 2 ]. Clinical prediction tools provide modest performance (AUCs ~ 0.65–0.75) [ 2 , 3 ] and often miss subclinical patterns indicating elevated risk. The 12-lead electrocardiogram (ECG) is a non-invasive, widely accessible tool in conventional medical practice. While consumer single-lead wearables are gaining popularity, the vast majority of validated prognostic data relies on standard clinical 12-lead recordings [ 4 ]. Artificial intelligence (AI), particularly deep learning, has emerged as a transformative approach for ECG analysis. AI models can extract latent features from raw waveforms and predict clinical outcomes without explicit feature engineering [ 5 – 8 ]. Unlike conventional ECG interpretation, which relies on pre-specified morphological criteria, deep learning models process raw voltage-time data across all 12 leads simultaneously, enabling detection of subtle prognostic patterns that are invisible to human readers. Several studies suggest AI-enhanced ECG models can predict cardiovascular mortality and major adverse cardiovascular events (MACE) [ 9 – 12 ]. However, no systematic synthesis has yet evaluated AI-ECG models specifically for mortality or MACE prediction. The evidence base has grown substantially over the past decade, spanning diverse populations, clinical settings, model architectures, and follow-up horizons — yet the heterogeneity in outcome definitions, validation approaches, and performance metrics makes cross-study interpretation difficult without structured synthesis. As AI-ECG tools approach clinical readiness, a rigorous evaluation of their prognostic performance and methodological quality is urgently needed to inform both clinical practice and future research priorities. In this review, we systematically evaluate the prognostic performance of AI-ECG models for predicting all-cause mortality, cardiovascular mortality, and MACE in adults, assess methodological quality, and identify priorities for future clinical validation and implementation. 2. Methods 2.1 Protocol and registration This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines [ 13 ] and was prospectively registered with the PROSPERO database (Registration ID: CRD420251244907). Full Details in S1 Appendix . 2.2 Eligibility criteria We included studies using AI/ML models with surface ECGs (12-lead or single-lead) as the primary input to predict all-cause mortality, cardiovascular mortality, or MACE in adults (≥18 years) in any clinical or community setting. Both internally and externally validated models were eligible, across retrospective and prospective cohort designs, model development and validation studies, and randomized controlled trials (RCTs). Studies employing a two-step design — wherein AI models first identify intermediate ECG phenotypes subsequently used to stratify mortality or MACE risk — were also eligible, provided the ECG remained the primary model input. We excluded pediatric-only studies, ECG-age or heart-age models, diagnostic-only models without prognostic intent, models not using ECG as a primary input, and conference abstracts without extractable outcome data. Following data extraction, individual MACE components (myocardial infarction [MI], stroke, heart failure [HF], ventricular arrhythmia) were additionally captured where reported by included studies, as these outcomes were consistently present and provided important prognostic context. Outcome definitions were extracted exactly as defined and reported by each study’s authors. This included extracting specific composite sub-endpoints, such as atherosclerotic cardiovascular disease (ASCVD), because its standard clinical components (e.g., non-fatal myocardial infarction, stroke, and cardiovascular death) directly overlap with classic MACE definitions, making it highly relevant to our prognostic scope. 2.3 Information sources We systematically searched four electronic databases: PubMed/MEDLINE, Cochrane CENTRAL, IEEE Xplore, and Google Scholar, from inception through November 25, 2025, with no date restrictions applied. All searches were limited to English-language publications. Google Scholar results were manually screened due to the platform’s limited advanced search capabilities, and the search was capped at the first 200 results. This aligns with established grey literature methodologies, as screening beyond this threshold introduces a high noise-to-signal ratio with exponentially diminishing scientific returns [ 14 ]. Conference abstracts were additionally screened where identified through the searches. 2.4 Search strategy Search strings were constructed using three concept blocks combined with Boolean operators: ECG-related terms (“electrocardiogram,” “ECG,” “EKG”), AI and machine learning terms (“artificial intelligence,” “machine learning,” “deep learning,” “neural network”), and prognostic outcome terms (“mortality,” “cardiovascular death,” “MACE,” “major adverse cardiovascular events”). Exclusion terms (“heart age,” “ECG age”) were applied across all databases to remove ECG-age estimation studies. Field-specific syntax was applied where supported — PubMed searches used [Title/Abstract] field tags, IEEE Xplore used metadata-only searching across title, abstract, and index terms, and Cochrane CENTRAL searches defaulted to title, abstract, and keyword fields. Full database-specific search strings are provided in S1 Appendix . 2.5 Study selection Two independent reviewers screened all records by title and abstract against the eligibility criteria, followed by full-text review of potentially eligible studies. Disagreements at either stage were resolved by discussion and consensus. A PRISMA 2020 flow diagram detailing the screening process is provided in Fig 1 . Download: PNG larger image TIFF original image Fig 1. PRISMA 2020 flow diagram. https://doi.org/10.1371/journal.pone.0360756.g001 2.6 Data extraction Data were extracted independently by two reviewers using a standardised extraction form. Extracted fields included: study design, country and clinical setting, sample size, ECG input type, AI model architecture, training and validation approach, outcomes reported, follow-up duration, and performance metrics (AUC, C-index, HR with 95% CI). Discrepancies were resolved by consensus. 2.7 Synthesis of results Statistical pooling and subgroup meta-analyses (e.g., for 1-year all-cause mortality) were deemed inappropriate a priori due to intractable clinical, methodological, and statistical heterogeneity. Even within specific temporal endpoints, baseline populations varied drastically, ranging from general unselected hospital cohorts to highly specific disease populations. Furthermore, the reported statistical metrics were mutually incompatible for quantitative synthesis, mixing standard binary classification metrics (AUC) with time-to-event survival metrics (C-index). Critically, several large-scale studies utilized overlapping regional datasets; pooling these datasets would introduce severe bias through the double-counting of patients. According to the Cochrane Handbook for Systematic Reviews of Interventions [ 15 ], attempting statistical pooling or calculating statistical heterogeneity metrics (such as I-squared) under these conditions violates core assumptions of baseline homogeneity and yields mathematically invalid summary estimates. Therefore, a rigorous narrative synthesis was performed. Risk of bias was assessed for each included study using PROBAST [ 16 ], with detailed rationale provided in S1 Appendix . 3. Results 3.1 Study selection Out of 512 records screened, 27 full-text articles were reviewed in detail. A total of 13 studies met inclusion criteria and were included in the final synthesis. Fourteen studies were excluded at full-text review: eleven due to incompatible focus or format (ECG-age estimation, diagnostic-only outcomes, pediatric population, or conference abstracts lacking full text), and three due to the use of traditional statistical regression models without neural network architecture (non-AI/ML model). A PRISMA flow diagram is provided in Fig 1 . 3.2 Overview of included studies The included studies were published between 2016 and 2025 and originated from diverse settings including the United States, Canada, Taiwan, Switzerland, South Korea, the United Kingdom, and Brazil. All studies used either 12-lead or single-lead surface ECGs as model inputs, with sample sizes ranging from several hundred to over two million ECGs. Twelve studies employed deep learning architectures (typically convolutional neural networks); one study systematically compared a ResNet-based deep learning model against an XGBoost approach [ 10 ]. Across the included literature: Eleven studies reported all-cause mortality, with eight evaluating short-term mortality (≤1 year), and five assessing long-term mortality (>1 year or extended follow-up), with some overlap between these categories. Four studies reported cardiovascular mortality. Four studies evaluated composite MACE outcomes, and an additional four reported on individual MACE components (e.g., MI, HF hospitalization, stroke) without composite MACE endpoints. Summary characteristics of the included studies are presented in Table 1 Download: PNG larger image TIFF original image Table 1. Characteristics of included studies. https://doi.org/10.1371/journal.pone.0360756.t001 3.3 All-cause mortality prediction Eleven studies reported all-cause mortality outcomes, spanning both short-term (30-day to 1-year) and long-term (>1-year) horizons. Short-Term All-Cause Mortality was reported by eight studies, Long-Term All cause mortality by five. Short-term all-cause mortality (n = 8). Eight studies evaluated AI-ECG models for predicting short-term all-cause mortality (30-day to 1-year): Sun et al. developed a ResNet-based deep learning model using over 1.6 million ECGs to predict short-term mortality at 30 days and 1 year, evaluated on an internal holdout dataset. In the validation cohort, the ECG-only model achieved an AUC of 0.843 (95% CI 0.838–0.848) for 30-day mortality and 0.812 (95% CI 0.808–0.816) for 1-year mortality [ 10 ]. Tsai et al. employed a CNN trained on a Taiwanese hospital cohort (Tri-Service General Hospital) ECG database to predict 1-year all-cause mortality. The model demonstrated strong performance, with an AUROC of 0.894 (95% CI 0.879–0.909) [ 17 ]. Raghunath et al. developed a deep learning model using a total cohort of 2,338,833 ECGs from Geisinger. In the total holdout test set, the model predicted 1-year all-cause mortality with an overall AUC of 0.88 and a hazard ratio of 8.1 (95% CI 7.9–8.2). Notably, the model maintained high performance even in a subgroup of patients with ‘Normal’ ECG interpretations (AUC 0.85) and a hazard ratio of 9.5 (95% CI 9.0–10.0) , highlighting its ability to detect subclinical prognostic signals [ 9 ]. Ouyang et al. developed PreOpNet, a deep learning model for postprocedural risk stratification in patients undergoing inpatient procedures. In a held-out test set of 4,581 patients, the model predicted 30-day all-cause mortality with an AUC of 0.83 (95% CI 0.79–0.87), and was subsequently validated across two external institutions [ 18 ]. Champetier et al. externally validated PreOpNet, an ECG-based model for 30-day mortality in high-risk surgical patients, with an AUC of 0.707 (95% CI 0.672–0.742) [ 19 ]. Lin et al. (ECG-MACE) reported 1-year all-cause mortality predicted by their multitask deep learning model ECG-MACE, developed using 2.8 million ECGs from Taiwan’s CGMH hospital network. The model achieved high performance for this endpoint, with an AUC of 0.89 for 1-year all-cause mortality prediction, the ECG-MACE model also achieved an external AUROC of 0.83 for 1-year all-cause mortality at an independent institution, and demonstrated long-term prognostic value in internal analyses up to 10 years [ 12 ]. Lin et al. (ECG-surv) developed a ResNet-18 model trained on 4.9 million ECGs, achieving a c-index of 0.860 for 1-year mortality [ 20 ]. Lee et al. developed a deep-learning model to predict 1-year all-cause mortality in heart failure with reduced ejection fraction (HFrEF) patients using ECGs. In the test set, the model achieved an AUC of 0.826 (95% CI 0.794–0.859), and the deep learning algorithm was an independent predictor with an adjusted HR of 4.12 (95% CI 2.32–7.33) [ 21 ]. Long-term all-cause mortality (n = 5). Five studies assessed long-term outcomes (>1 year): Sun et al. was also evaluated for 5-year all-cause mortality. The model achieved an AUC of 0.798 (95% CI 0.792–0.803) in the validation cohort, maintaining consistent discrimination across time [ 10 ]. Alrumayh et al. applied a single-lead AI model to detect low left ventricular ejection fraction (LVEF); patients flagged as high-risk had a significantly increased 2-year all-cause mortality (HR 1.56, 95% CI 1.06–2.29) [ 22 ]. Sau et al. developed a deep learning model using BIDMC, Boston, USA ECGs to predict long-term cardiovascular outcomes, including all-cause mortality. In a cohort of 189,539 patients with a mean follow-up of 5.46 years, the model achieved a concordance index (c-index) of 0.775 for predicting all-cause death. The AI-ECG model outperformed traditional clinical risk scores and demonstrated significant prognostic value even in patients with normal conventional ECG interpretations [ 23 ]. Tsai et al. employed a CNN trained on a Taiwanese hospital cohort to predict 10-year all-cause mortality. The model demonstrated strong performance, with a C-index of 0.858 in internal validation and 0.836 in external validation and an adjusted HR of 14.16 (adjusted for age and sex) (95% CI 11.33–17.70) for high-risk versus low-risk individuals [ 17 ]. Lu et al. developed an interpretable deep learning model that identified cardiac abnormalities from 2.3 million ECGs from a Brazilian primary care network, which were then used to stratify mortality risk over up to 7 years. Using ECG-predicted results, subjects with sinus tachycardia and atrial fibrillation demonstrated the highest adjusted mortality risk, with HRs of 2.24 (95% CI 1.96–2.57) and 2.22 (95% CI 1.99–2.48) respectively [ 24 ]. These findings support the use of AI-ECG as a robust prognostic tool for mortality risk stratification, with performance often exceeding traditional clinical scores across various timeframes and populations. 3.4 Cardiovascular mortality prediction Four studies predicted Cardiovascular Mortality outcomes: Hughes et al. developed the SEER model to predict 5-year cardiovascular mortality, with an internal AUC of 0.83 and external validation AUCs of 0.78 and 0.83 [ 11 ]. Lin et al. (ECG-surv) developed a deep learning model based on a ResNet-18 architecture to predict 1-year cardiovascular mortality using 4.9 million 12-lead ECGs from Chang Gung Memorial Hospital (CGMH), Taiwan. The model was trained with a time-to-event survival framework and achieved a c-index of 0.891 for cardiovascular death, This study represents one of the largest datasets used in ECG-AI modeling and demonstrated excellent prognostic discrimination for CV mortality at 1 year [ 20 ]. Sau et al. developed an ECG-based AI model using 1,163,401 ECGs from Beth Israel Deaconess Medical Center (BIDMC), Boston, USA to predict major adverse cardiovascular outcomes, including cardiovascular mortality. In this study, the model achieved a C-index of 0.832 (95% CI 0.831–0.834) for cardiovascular death over a mean follow-up of 5.46 years [ 23 ]. Tsai et al. developed a survival deep learning model using 512,805 ECGs from 271,407 patients across two Taiwanese hospital cohorts to predict long-term cardiovascular mortality. The high-risk group identified by the model had a significantly elevated cardiovascular mortality risk compared to the low-risk group, with an HR of 18.50 (95% CI 9.82–34.84) in internal validation and 13.74 (95% CI 8.57–22.04) in external validation. This study highlights the long-term predictive capacity of AI-ECG models and demonstrates applicability across extended follow-up horizons [ 17 ]. These studies support the prognostic use of ECGs beyond diagnostic boundaries and demonstrate that AI can detect latent features related to future risk, even in asymptomatic individuals. 3.5 Major adverse cardiovascular events (MACE) composite Four studies evaluated AI-ECG models for predicting composite MACE. Ouyang et al. developed PreOpNet, a deep learning model using preoperative 12-lead ECGs to predict adverse outcomes across inpatient and outpatient procedural settings. The composite MACE outcome included mortality, cardiac arrest, myocardial infarction, heart block, and pulmonary oedema. In the inte
## Related Clinical Research

- [Cardiovascular risk prediction models in women: protocol for an updated systematic review](https://medichelpline.com/clinical-feed/plos-one-7-cardiovascular-risk-prediction-models-for-women-a-systematic-review-protocol.md)
- [Comparing BMI, Waist Circumference, and BRI for Predicting ASCVD, AF, and Mortality in 1.2M Koreans](https://medichelpline.com/clinical-feed/plos-one-14-prognostic-associations-of-anthropometric-indices-with-major-health-outcomes-a.md)
- [Accuracy of EMR Retrieval Methods and a Large Language Model for Identifying Cardiovascular Events](https://medichelpline.com/clinical-feed/bmj-open-6-diagnostic-accuracy-of-electronic-medical-record-retrieval-methods-and-a-large.md)
- [Long-term cardiovascular risk after assisted reproductive technology and infertility: nationwide c](https://medichelpline.com/clinical-feed/plos-medicine-0-long-term-risk-of-cardiovascular-disease-after-assisted-reproductive-technology.md)
- [Intracoronary prourokinase for slow‑flow/no‑reflow during PCI in ACS: randomized double‑blind tria](https://medichelpline.com/clinical-feed/bmj-open-5-efficacy-and-safety-of-intracoronary-injection-of-recombinant-human.md)

## Navigation
- [← Back to Cardiology Feed](https://medichelpline.com/clinical-feed/cardiology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.