Artificial intelligence for cardiac arrest in the digital health era: From algorithmic performance to workflow-integrated, outcome-oriented digital health systems

Abstract

Objectives

To characterize the global research landscape, collaboration patterns, and thematic evolution of artificial intelligence (AI) in cardiac arrest care.

Methods

We conducted a bibliometric analysis of AI and cardiac arrest research using the Web of Science Core Collection and Scopus. We retrieved records on October 29, 2025, limited to English-language articles and reviews. After deduplication, 1,228 publications were included. CiteSpace, VOSviewer, and bibliometrix (R) assessed publication growth, country/institution contributions, collaboration networks, co-citation structures, and keyword dynamics.

Results

We included 1,228 publications (2012-2025) with a compound annual growth rate (CAGR) of 22.05%. The corpus comprised 6,994 authors (mean of 7.46 per paper), and international co-authorship accounted for 13.36% of all documents. Forty-eight countries contributed, with the United States leading (134 publications; 10.9%), followed by South Korea (101) and China (77). Seoul National University was the most productive institution (92), followed by Harvard University (60). Aramendi E. ranked first among authors (21), followed by Park J. (17) and Ong M.E.H. (16). Resuscitation ranked first among journals, with 54 publications and 1,350 citations. Keywords shifted from method-focused topics to more clinical, system- and outcome-focused themes. Burst and clustering analyses emphasized early warning/prediction and neurological outcomes, with recent burst terms including “Cerebral Performance Category (CPC),” “brain injury,” and “emergency medical services (EMS).”

Conclusions

AI–cardiac arrest research is entering a maturing expansion phase characterized by interdisciplinary linkages and a multipolar collaboration structure. The field is shifting beyond algorithmic performance toward transportable, workflow-aware, and outcome-oriented digital health systems.

Highlights

Artificial Intelligence for Cardiac Arrest in the Digital Health Era: From Algorithmic Performance to Workflow-Integrated, Outcome-Oriented Digital Health Systems

• • This bibliometric study identifies accelerated growth in AI–cardiac arrest research after 2020.

• • Global collaboration shows a multipolar structure spanning the United States, Europe, and Asia.

• • The field is moving beyond algorithmic performance toward workflow-integrated, outcome-oriented digital health systems.

• • Future clinical translation requires standardized neurological outcome assessment, multicenter validation, prospective evaluation, safety, and governance.

1. Introduction

Cardiac arrest is a life-threatening medical emergency characterized by the sudden cessation of effective cardiac mechanical activity and the consequent loss of systemic perfusion. Epidemiologically, cardiac arrest imposes a substantial burden in the United States, affecting hundreds of thousands of individuals each year, and survival after out-of-hospital cardiac arrest remains approximately 10%–12%; delayed treatment and limited immediate bystander cardiopulmonary resuscitation (CPR) contribute substantially to this low survival rate. ¹ Management centers on rapid recognition, immediate high-quality CPR, and early defibrillation when a shockable rhythm is present, followed by advanced life support and post–cardiac arrest care. ²

Artificial intelligence (AI) has become an increasingly important component of contemporary cardiovascular medicine. Beyond general risk stratification, AI methods have been applied to cardiovascular imaging, electrocardiographic interpretation, disease diagnosis, personalized risk prediction, and clinical decision support. Recent evidence suggests that AI may improve cardiovascular risk prediction by integrating multimodal clinical and imaging data. However, methodological heterogeneity, poor calibration, algorithmic bias, limited external validation, and regulatory uncertainty remain major barriers to clinical translation. ³ In addition, machine learning frameworks for cardiovascular diagnosis increasingly incorporate preprocessing, imputation, scaling, feature optimization, and diverse classifiers, reflecting the growing methodological sophistication of AI-enabled diagnostic pipelines. ⁴ The American Heart Association has also emphasized the potential of AI to improve cardiovascular outcomes, while highlighting the need for safety, transparency, validation, and responsible implementation. ⁵ More broadly, adjacent studies in measurement-intensive engineering highlight the importance of reliable characterization and robust performance assessment in technology development.^6–8 Complementary AI governance literature emphasizes privacy, transparency, accountability, ethical oversight, and user trust as key considerations in responsible technology deployment. ⁹ Together, these developments provide the broader methodological and translational context needed to examine AI applications in cardiac arrest research.

Against this background, AI has increasingly been applied to cardiac arrest research, enabling the analysis of large-scale healthcare datasets to identify outcome-associated patterns and risk factors. AI applications, such as machine learning models, have been developed to predict neurological outcomes in out-of-hospital cardiac arrest (OHCA) survivors, achieving high accuracy in both internal and external validation, which underscores their potential to improve risk stratification and clinical decision-making. ¹⁰ Additionally, AI systems such as DeepCARS have shown significant associations with cardiac arrest risk, supporting earlier identification of high-risk individuals and informing preventive or acute interventions. ¹¹ Integrating heterogeneous healthcare data into AI-driven prediction frameworks may improve model performance, facilitate earlier recognition of high-risk patients, and optimize resource allocation across prehospital and in-hospital care pathways. ¹² Collectively, these applications suggest that AI–cardiac arrest research is expanding beyond isolated prediction tasks toward risk identification, early warning, prognostication, and care-pathway optimization.

Within resuscitation science, AI has been explored across multiple stages of the cardiac arrest continuum, including early detection, emergency call recognition, rhythm interpretation, automated external defibrillator support, CPR quality feedback, post-resuscitation prognostication, and neurological outcome prediction. A recent scoping review reported that AI may support the prediction of cardiac arrest, heart rhythm disorders, and post-cardiac arrest outcomes, as well as defibrillator drone delivery and dispatcher notification. ¹³ CPR-specific reviews further suggest that AI and machine learning may enhance rhythm interpretation, CPR quality monitoring, real-time feedback, ROSC-oriented resuscitation support, and post-resuscitation care; however, the evidence remains heterogeneous and requires rigorous prospective validation.^14,15 Importantly, AI does not consistently improve real-world workflow performance. A randomized clinical trial of machine learning support for dispatcher recognition of OHCA did not significantly improve dispatchers’ recognition performance, underscoring the need to distinguish algorithmic promise from clinically demonstrated workflow benefit. ¹⁶

Representative empirical studies further illustrate the breadth of AI applications across this continuum. Real-time prediction of in-hospital cardiac arrest has been explored using continuously collected physiological signals, including ECG-derived features, and ICU/EHR-linked data streams to support earlier escalation decisions. ¹⁷ Parallel efforts have focused on post-OHCA prognostication, with models developed and externally validated to predict neurological outcomes and inform post–cardiac arrest care pathways. ¹⁸ Prehospital applications have also emerged, including machine-learning–assisted recognition of OHCA during emergency calls. ¹⁶ In prevention-oriented research, AI-enhanced electrocardiography has been used to detect conditions associated with sudden cardiac death risk, such as hypertrophic cardiomyopathy. ¹⁹ Taken together, these studies demonstrate an expanding evidence base spanning prediction, emergency response, post-resuscitation care, and prevention-oriented cardiovascular risk detection.

Despite this rapid growth, several challenges continue to limit the interpretation and clinical translation of AI research in cardiac arrest. First, existing studies are dispersed across emergency medicine, critical care, cardiovascular medicine, biomedical engineering, computer science, and digital health, making it difficult to evaluate the field as an integrated research domain. Second, studies differ substantially in data sources, model architectures, validation strategies, clinical endpoints, and reporting standards. Third, many AI studies emphasize discrimination metrics, such as the area under the receiver operating characteristic curve (AUROC), whereas fewer studies evaluate calibration, workflow impact, false-alarm burden, prospective performance, safety, fairness, or patient-centered outcomes. Fourth, although neurological outcomes and meaningful recovery are increasingly discussed, outcome definitions remain heterogeneous across out-of-hospital cardiac arrest (OHCA) and in-hospital cardiac arrest (IHCA) settings. Together, these challenges underscore the need for a reproducible bibliometric and thematic mapping study that examines not only publication volume but also collaboration structures, intellectual foundations, research hotspots, and translational gaps. Such an approach can quantify growth trajectories, identify leading contributors and collaboration networks, characterize the co-citation knowledge base, and track thematic evolution, thereby clarifying research gaps and informing future translational priorities.

Therefore, this study was designed to clarify how AI–cardiac arrest research has developed from algorithm-focused studies toward a broader digital health domain spanning prediction, emergency response, post-resuscitation care, and meaningful neurological recovery. Accordingly, this study makes four main contributions. First, it quantifies global publication growth and identifies the leading countries, institutions, authors, and journals in AI–cardiac arrest research. Second, it maps collaboration networks and co-citation structures to characterize the field’s intellectual foundations. Third, it tracks the temporal evolution of keywords, citation bursts, and thematic clusters to identify emerging priorities, including early warning, neurological outcomes, emergency medical services (EMS), and workflow-aware decision support. Fourth, it interprets these bibliometric patterns from a translational digital health perspective, with emphasis on external validation, prospective workflow integration, standardized endpoints, safety, fairness, and governance. By integrating productivity, collaboration, co-citation, keyword, and translational perspectives, this study provides a structured overview of the field and identifies priorities for validation, implementation, and outcome-focused AI research in cardiac arrest.

2. Methods

2.1. Database and search strategy

This study utilized the Web of Science Core Collection Database (WoSCC) and Scopus databases, both of which are widely recognized bibliometric datasets. ²⁰

The search strategy was organized into two concept blocks: cardiac arrest/resuscitation and artificial intelligence, as summarized in Box 1. The complete database-specific Boolean search expressions are provided in Supplementary Table 1. All records were retrieved on October 29, 2025. We excluded non-research document types (e.g., proceedings papers, corrections, news items, book chapters, retracted publications, and editorials). We included English-language articles and reviews only. Figure 1 summarizes the study workflow.

Box 1. Database-specific search strategy summary

Item	Description
Databases	Web of Science Core Collection (WoSCC) and Scopus
Search date	October 29, 2025
Time span	January 2012 to October 29, 2025
WoSCC fields	TI/AB/AK: title, abstract, and author keywords
Scopus fields	TITLE-ABS-KEY: title, abstract, and keywords
Disease/resuscitation concept block	“cardiac arrest” OR “heart arrest” OR “cardiopulmonary arrest” OR “out-of-hospital cardiac arrest” OR “OHCA” OR “in-hospital cardiac arrest” OR “IHCA” OR “sudden cardiac arrest”
Artificial intelligence concept block	“artificial intelligence” OR “AI (artificial intelligence)” OR “machine learning” OR “deep learning” OR “natural language processing” OR “NLP” OR “neural network” OR “neural networks” OR “convolutional neural network” OR “convolutional neural networks” OR “CNN” OR “recurrent neural network” OR “recurrent neural networks” OR “RNN” OR “graph neural network” OR “graph neural networks” OR “GNN” OR “ensemble learning” OR “supervised machine learning” OR “unsupervised machine learning” OR “transfer machine learning” OR “federated learning” OR “representation machine learning” OR “machine intelligence” OR “computer vision systems” OR “computer vision system” OR “reinforcement machine learning” OR “feedforward neural networks”
Boolean logic	Disease/resuscitation concept block AND artificial intelligence concept block
Field logic	In WoSCC, each concept block was searched across TI OR AB OR AK fields; in Scopus, each concept block was searched using TITLE-ABS-KEY.
MeSH note	Because WoSCC and Scopus do not use PubMed MeSH indexing, MeSH terms were not directly applied. Expanded free-text synonyms and database-specific field tags were used to approximate the conceptual coverage of relevant MeSH terms.
Limits	English-language articles and reviews
Exclusions	Proceedings papers, editorials, corrections, news items, book chapters, retracted publications, and duplicate records
Full expressions	Complete database-specific Boolean expressions, including all operators, field tags, and parenthetical groupings, are provided in Supplementary Table 1.

Figure 1.

Flowchart of literature retrieval, screening, deduplication, and bibliometric analysis.

For data integration, Scopus records were converted into a Web of Science–compatible plain-text format using the data conversion function in CiteSpace 6.4.R1. To minimize field-mapping discrepancies between WoSCC and Scopus, the converted records were harmonized according to a predefined bibliometric field schema. Core fields were checked and standardized, including article title, author names, affiliations, country/region information, source title, publication year, document type, DOI, author keywords, indexed keywords, and cited references. Scopus author keywords and indexed keywords were mapped to the corresponding keyword fields in the WoS-compatible format. For cited references, complete citation strings were retained when exact subfields, including volume, issue, page number, or DOI, were not fully equivalent across databases. To validate the conversion process, 100 converted Scopus records were randomly selected and manually compared with the original Scopus exports. The checked fields included article title, authors, affiliations/countries, keywords, journal or source title, publication year, DOI, and cited references. No systematic data loss or field mislabeling that could affect the bibliometric analyses was identified. Minor formatting differences, including punctuation, capitalization, journal-title abbreviations, and affiliation-name variants, were corrected during data cleaning in R and Microsoft Excel. The WoSCC dataset and the converted Scopus dataset were then merged and deduplicated using DOI, article title, first author, and publication year as matching criteria. After deduplication, 1,228 records were retained for analysis. Bibliometric analyses were performed on the combined plain-text files using CiteSpace 6.4.R1, VOSviewer 1.6.20, and the bibliometrix package in R version 4.4.3.

2.2. Data processing

2.2.1. CiteSpace

CiteSpace (version 6.4.R1) was employed to analyze the data. ²¹ The analysis encompassed the period from January 2012 to October 29, 2025, utilizing a time slice of one year. Node types included authors, institutions, and keywords. For author and institution nodes, the threshold was established at the top 25 per slice without pruning. In contrast, for keywords, a threshold of 25 was applied with pruning through pathfinder and merged network techniques. These predefined node types, thresholds, and pruning strategies were used to standardize the selection of bibliometric units, reduce network complexity, and preserve the main structural information. Visual analyses produced knowledge maps illustrating researchers, institutions, and keywords. All records obtained from WoSCC and Scopus were saved as “full records and cited references” in plain text format.

2.2.2. VOSviewer

Data were analyzed with VOSviewer 1.6.20, a software developed by the Centre for Science and Technology Studies (CWTS) at Leiden University. ²² Full counting was applied. Thresholds were set by analysis type to generate visualizations of collaboration networks.

2.2.3. Bibliometrix

The bibliometrix R package, developed by Massimo Aria and Corrado Cuccurullo, was used for historiographic analysis, trend monitoring of journals and authors, and calculation of key indicators (e.g., g-index, h-index, number of citations [NC], and number of publications [NP]). ²³

2.2.4. The other tools

Microsoft Excel 2021 (version 16.48) was used for initial data cleaning and organization. In addition, an online bibliometric analysis platform was used to support exploratory analyses and visualization of citation data. GraphPad Prism 8.0 was used to generate graphs.

2.3. Data analysis

Descriptive bibliometric indicators were calculated to summarize the overall characteristics of the included publications, including total publication output, annual output, cumulative counts, document types, countries/regions, institutions, authors, journals, references, and keywords. Publication growth was assessed using annual output, cumulative publication counts, and the compound annual growth rate (CAGR). CAGR was calculated from the first and last annual publication counts using the following formula: $\text{CAGR}={\left(\frac{N_{\text{final}}}{N_{\text{initial}}}\right)}^{\frac{1}{n}}-1$ , where N_initial denotes the number of publications in 2012, N_final denotes the number of publications in 2025, and n denotes the number of annual intervals between 2012 and 2025. Exponential growth fitting was performed to evaluate whether publication output conformed to Price’s Law. Goodness of fit was assessed using the coefficient of determination (R²). Because the search was conducted on October 29, 2025, the 2025 publication count was treated as partial-year output and interpreted cautiously.

Country-level collaboration was assessed using single-country publications (SCP), multiple-country publications (MCP), MCP percentage, total citations, and mean citations per article. Institutional, author, and journal contributions were evaluated using publication counts and citation-related indicators, where applicable. For journal analysis, bibliometrix was used to calculate bibliometric indicators, including the h-index, g-index, total citations, and number of publications (NP).

Network analyses were performed to characterize collaboration patterns and knowledge structures. CiteSpace was used to construct author, institution, keyword, and cited-reference networks. The time span was set from January 2012 to October 29, 2025, with one year per time slice. For author and institution analyses, the top 25 items per time slice were selected. For keyword analysis, a threshold of 25 was applied with Pathfinder and merged-network pruning to reduce visual complexity while preserving the main network structure. CiteSpace was also used for co-citation clustering, timeline visualization, and burst detection. To complement the quantitative co-citation cluster outputs, we qualitatively interpreted the major influential co-cited clusters. Representative references within each cluster were reviewed and summarized according to their bibliometric signals, influential cited evidence, DOI information, citation influence, qualitative meaning, and remaining research gaps. Citation burst analysis was used to detect cited references or keywords with abrupt increases in citation or occurrence frequency within a defined time interval. In CiteSpace, burst strength represents the intensity of this temporal increase, whereas the beginning and ending years indicate the duration of the burst period.

VOSviewer was used to generate collaboration, co-occurrence, and co-citation maps using full counting. Depending on the analysis type, node size represented publication output, frequency, or citation-related weight; link thickness represented collaboration, co-occurrence, or co-citation strength; and color represented clusters or the average publication year in overlay visualizations. Total link strength (TLS) was used to quantify the overall strength of relationships between nodes in co-occurrence and collaboration networks.

Because this study was a bibliometric analysis based on publication-level metadata rather than a predictive modeling study using patient-level data, no model training, machine-learning feature selection, inferential hypothesis testing, or dataset balancing was performed. To mitigate potential bibliometric bias, we used two major databases, applied predefined inclusion and exclusion criteria, harmonized Scopus records into a WoS-compatible format, deduplicated records using DOI, article title, first author, and publication year, and applied predefined node types, thresholds, and pruning strategies for network construction. All analyses were based on publication-level metadata and citation records. For author-level analyses, author names were standardized according to the name strings exported from WoSCC and Scopus. To reduce potential ambiguity among highly ranked author-name entries, we conducted targeted verification of the top 20 most prolific entries using available author identifiers, affiliation information, publication topics, and co-author patterns, where possible. Author productivity indicators were therefore interpreted as author-name-level summaries rather than definitive person-level rankings. From a computational perspective, the analytical workflow mainly involved record preprocessing, duplicate detection, frequency counting, network construction, clustering, burst detection, and visualization. These procedures scaled primarily with the number of bibliographic records (N) and network links (E). Preprocessing, deduplication, and frequency counting were approximately linear in N, whereas collaboration, co-occurrence, and co-citation network construction depended on both N and E. Because the final dataset contained 1,228 records and network visualizations were generated after predefined thresholding and pruning, the computational burden was modest and did not affect reproducibility or interpretation. Data cleaning and organization were performed using Microsoft Excel 2021 and R. Bibliometric analyses and visualizations were performed using CiteSpace 6.4.R1, VOSviewer 1.6.20, and the bibliometrix package in R version 4.4.3.

3. Results

3.1. Annual publications trends

The corpus contained 1,228 documents published between 2012 and 2025 (Supplementary Table 2). The corpus provides an overview of the field over nearly 14 years and reflects increasing publication volume and topical diversity. Annual publication output increased from 18 records in 2012 to 240 records in 2025, corresponding to a compound annual growth rate (CAGR) of 22.05% (Figure 2(a)). The cumulative number of publications increased over time and reached 1,228 by the retrieval date (Figure 2(b)). The literature search was conducted on October 29, 2025; therefore, the 2025 publication count represented partial-year output. Price’s Law was used to compare fitted and observed publication trends. The annual publication output showed a good exponential fit, with the fitted model y = 6.9093e^0.2647x and R² = 0.9455 (Figure 2(c)). After 2020, the observed annual publication counts were higher than the values estimated by the Price’s Law fitted curve, and the largest observed–fitted differences occurred in the most recent years. During 2012–2015, annual publication counts were relatively low, ranging from 10 to 19 records per year. Figure 2(d) shows the document types: articles (1,171; 95%) and reviews (57; 5%).

Figure 2.

Publication trends in AI–cardiac arrest research (2012-2025). (a) Annual publication output; (b) Cumulative number of publications; (c) Price’s Law–based exponential growth fitting of annual publication output. The fitted model was y = 6.9093e^0.2647x, with R² = 0.9455. Because records were retrieved on October 29, 2025, the 2025 count represents partial-year output; (d) Distribution of document types.

3.2. Distributions of countries

Figure 3(a) summarizes the international collaboration network, with the most active links involving the United States, European countries, and several Asian countries. In total, 48 countries contributed to AI–cardiac arrest research; the United States contributed 134 publications (10.9%), the highest output among countries/regions. The United States showed extensive co-authorship links, with prominent connections to European countries. In Asia, South Korea contributed 101 publications (8.2%), followed by China (77, 6.3%), India (34, 2.8%), and Japan (27, 2.2%). In Europe, Spain and France contributed 22 and 16 publications, respectively (Supplementary Table 3).

Figure 3.

Country-level contributions and collaboration network in AI–cardiac arrest research. (a) Country collaboration map. Darker blue indicates higher collaboration intensity, and line thickness reflects the strength of bilateral collaboration. (b) Global country collaboration network in AI–cardiac arrest research. (c) Country collaboration network (overlay visualization) in AI–cardiac arrest research. (d) Country distribution over time based on corresponding-author affiliations. Countries are ranked by total publication output; bars distinguish single-country publications (SCP) from multiple-country publications (MCP), reflecting domestic versus international collaboration. (e) Top 10 countries by citation impact.

Figure 3(b) shows the global collaboration network among countries/regions involved in AI–cardiac arrest research. The VOSviewer map depicts international collaborations based on country-level co-authorship. Nodes represent countries/regions, and node size is proportional to publication output; link thickness indicates the strength of collaborative ties, and colors denote collaboration clusters. The United States was the largest node and had prominent co-authorship links with countries in Europe and Asia, including the United Kingdom, Germany, the Netherlands, China, Japan, and South Korea. European countries formed a dense intra-regional cluster, and direct collaborative links were observed between Asian and European countries. The map displayed a multipolar collaboration structure rather than a U.S.-centered hub-and-spoke structure. International collaboration accounted for 13.36% of publications (Supplementary Table 2). Overall, the network indicates a geographically diverse and increasingly connected research landscape.

Figure 3(c) shows the temporal distribution of publications by country, with node size proportional to cumulative output. Figure 3(d) and Supplementary Table 3 summarize corresponding-author contributions by country, distinguishing single-country publications (SCP) from multiple-country publications (MCP). The United States had the largest numbers of both SCPs (n = 110) and MCPs (n = 24). By proportion, Spain had the highest MCP share (18/22, 81.8%), followed by France (7/16, 43.8%), the United Kingdom (5/14, 35.7%), Japan (9/27, 33.3%), and Germany (4/12, 33.3%). MCP% was calculated among publications assigned to each corresponding-author country; for Spain, the MCP count was 18 and the denominator was 22 corresponding-author publications. Figure 3(e) lists the top 10 countries/regions by total citations; the United States ranked first (2,062 citations; 15.4 citations per article). South Korea, China, and India followed (1,436, 594, and 438 citations, respectively), with mean citations per article ranging from 7.7 to 14.2.

3.3. Distribution by institutions

Figure 4(a) shows that Seoul National University (SNU) had the highest publication output (n = 92), followed by Harvard University (n = 60), the University of Toronto (n = 54), Harvard University Medical Affiliates (n = 48), and the Pennsylvania Commonwealth System of Higher Education (PCSHE; n = 47) (Figure 4(b); Supplementary Table 4). Annual output remained low before 2016, increased after 2018, and rose further after 2020. Harvard University’s annual output peaked in 2025, while PCSHE and SNU increased their outputs from 2021 onward.

Figure 4.

Institutional contributions and collaboration network in AI–cardiac arrest research. (a) Annual publication output of the top contributing institutions. (b) Top 10 institutions by publication volume. (c) Institutional collaboration network: node size represents publication volume, edge thickness indicates collaboration strength, and node color denotes collaboration clusters. (d) Timeline view of institutional collaboration, illustrating the entry and evolution of major institutions over time.

Figure 4(c) presents the institutional collaboration network and cluster structure in AI–cardiac arrest (AI–CA) research. Nodes represent institutions, and node size is proportional to publication output. Links represent co-authorship collaborations, with thicker links indicating stronger ties; colors denote clusters. The network comprised multiple hubs and clusters, with cross-regional links. The red cluster represents Korea–Singapore collaboration and is anchored by Seoul National University, with Yonsei University, Korea University, Samsung Medical Center, and Singapore General Hospital as additional members. This cluster shows dense connectivity between Korean and Singaporean institutions. The green cluster comprises a Nordic–European group, including Karolinska Institute, the University of Copenhagen, Oslo University Hospital, and Sahlgrenska University Hospital. The blue cluster includes major North American institutions, including the University of Chicago, Johns Hopkins University, and the University of Pennsylvania. The yellow cluster comprises Massachusetts-centered collaborations with links to selected European partners; key institutions include MIT, Massachusetts General Hospital, and the University of California, San Francisco.

Figure 4(d) overlays institutions by average publication year to depict temporal changes in institutional activity. Nodes represent institutions; node size is proportional to publication output, and color (blue to yellow) encodes the average publication year (yellow = more recent). Earlier activity (blue–green) was concentrated in Nordic and European institutions, whereas more recent activity (yellow) was concentrated in North America and East Asia. Recent activity clustered in North America and East Asia, with several emerging hubs. Yellow-coded institutions were concentrated in East Asia and North America, including Seoul National University, Massachusetts General Hospital, the University of Chicago, Johns Hopkins University, and MIT. These institutions formed prominent nodes in the collaboration network. Inter-institutional links were dense, particularly in the most recent five-year period. Across 2012–2025, the number of contributing countries/regions increased, and annual publication output rose over time.

3.4. Distributions of authors and co-cited authors

From 2012 to 2025, 6,994 unique authors contributed to AI–cardiac arrest publications (Supplementary Table 2). The mean number of authors per document was 7.46, and 25 papers were single-authored (Supplementary Table 2). The top 10 authors formed a productivity core in this literature (Supplementary Table 5). Irusta and Aramendi were among the leading contributors in AI-based defibrillation signal analysis, with publications spanning ECG waveform characterization, VF/VT detection, and algorithm refinement. The co-authorship network comprised several high-output teams and separated into subcommunities centered on signal processing, deep learning, and systems-oriented research.

Figure 5(a) lists the most prolific authors in AI–cardiac arrest research. Aramendi E. ranked first (n = 21), followed by Park J. (n = 17) and Ong M.E.H., Kim J., and Irusta U. (each n = 16). Figure 5(b) shows the temporal evolution of author productivity; Ong M.E.H. published in this field as early as 2012. Since 2016, Aramendi E., Irusta U., and Park J. have maintained high publication output. The VOSviewer co-authorship map (Figure 5(c)) further delineates several research groups, including a large cluster centered on Aramendi E. and Irusta U., another led by Eftestøl T.C. and Kramer-Johansen J., and smaller groups that include Kudenchuk P.J. and Rea T.D. The author co-citation network (Figure 5(d)) maps shared intellectual foundations and the influence structure of core authors in the AI–cardiac arrest knowledge base.

Figure 5.

Author collaboration network and co-cited author analysis in AI–cardiac arrest research. (a) Top 10 authors by publication output. (b) Annual publication trends of the top 10 authors. (c) Co-authorship network. (d) Co-cited author clustering in AI–cardiac arrest research. In panel (d), node size is proportional to co-citation frequency.

3.5. Journals and co-journals

Supplementary Table 6 shows that AI–cardiac arrest publications (2012–2025) were concentrated in 15 core journals, most of which were classified as JCR Q1. These journals covered three domains: resuscitation/critical care, cardiovascular medicine, and biomedical engineering/AI. Resuscitation ranked first by influence (h-index = 16; total citations = 1,350; publications = 54). Engineering and computational outlets ranked among the leading journals, including IEEE Access, Computers in Biology and Medicine, Computer Methods and Programs in Biomedicine, and Sensors. The Journal of Medical Internet Research was also included among the core journals. Multidisciplinary open-access journals were represented, including PLOS ONE and Scientific Reports.

Journal co-occurrence analysis (Figure 6(a)) placed Resuscitation at the center of the network. Three journal clusters were identified: (i) a clinical cluster (Resuscitation, Critical Care Medicine), (ii) an engineering/computational cluster (IEEE Access, Sensors, Computers in Biology and Medicine), and (iii) a cardiovascular cluster (Circulation, JACC). The journal co-citation network (Figure 6(b)) showed dense cross-links among journals. Three major co-citation clusters were delineated, covering resuscitation/critical care, cardiovascular medicine, and engineering/computational sources; Resuscitation remained centrally located and connected across clusters.

Figure 6.

Journal association networks in AI–cardiac arrest research. (a) Journal co-occurrence network. (b) Journal co-citation and association network. Nodes represent journals; node size is proportional to linkage strength and citation-based metrics, and node color indicates cluster membership. (c) Dual-map overlay of journals in AI–cardiac arrest research. Citing journals are shown on the left and cited journals on the right. Colored citation trajectories illustrate knowledge flows, with thicker paths indicating stronger citation links.

Dual-map overlay analysis (Figure 6(c)) depicts an interdisciplinary citation structure. Most cited references were located in the medicine/clinical medicine domain, and the dominant citation paths linked medicine/clinical medicine to health/nursing/medicine and to molecular biology/genetics. Additional citation paths were observed from systems/computing to systems/computing and mathematics.

3.6. References and articles

The corpus comprised 22,294 cited references and averaged 15 citations per document (Supplementary Table 2). The 10 globally cited publications in the bibliometric dataset (2012–2025; Supplementary Table 7) covered international resuscitation guidelines, machine learning–based deterioration prediction studies, and broader AI-related publications. The European Resuscitation Council Guidelines 2021: Basic Life Support ranked first, with 449 citations. ²⁴ The list also included studies on clinical monitoring, cardiac arrest prediction, and patient safety (e.g., Churpek 2016; Kwon 2018), as well as emergency medical dispatch (Blomberg 2019) and general AI reviews (Qayyum 2021).^25–28

Reference co-citation analysis (Figure 7(a)) maps the intellectual structure of AI–cardiac arrest research. Several core publications showed high co-citation frequencies (e.g., Churpek 2016; Kwon 2018; Hannun 2019; Nolan 2021).^25,26,29,30 Earlier co-cited studies centered on AI-based detection of clinical deterioration and early warning systems. From 2019 onward, deep learning–based cardiac arrest prediction featured prominently in the co-citation network. After 2020, frequently co-cited topics included multicenter datasets, updated resuscitation guidelines, and AI-assisted resuscitation decision support.

Figure 7.

Reference co-citation networks in AI–cardiac arrest research. (a) Reference co-citation network. Nodes represent cited references and edges indicate co-citation links. Node size is proportional to citation frequency, and warmer colors indicate more recent publications. (b) Co-cited reference clustering map. Colored regions denote distinct thematic clusters, with cluster labels extracted from the titles of citing articles. (c) Timeline view of co-cited reference clusters. Horizontal lines represent clusters, and node positions along the timeline indicate the year of strongest influence for each reference. (d) Top 25 references with the strongest citation bursts. Blue segments denote the observation period, and red segments indicate burst intervals, marking the onset and end of each burst.

The co-citation clustering map (Figure 7(b)) depicts the knowledge structure of AI–cardiac arrest research. Eleven major clusters were identified. Silhouette scores ranged from 0.90 to 1.00, with high within-cluster coherence and clear between-cluster separation. Cluster #0 (“predicting neurological outcome”; mean year 2020) was the largest cluster and had the most recent mean year; it centered on post–cardiac arrest neurological prognostication. This cluster contained terms related to deep learning, electroencephalography (EEG)–based prediction, and post-resuscitation prognostic assessment, and it included neurological recovery as an endpoint. Cluster #1 (“lethal ventricular arrhythmia”) and Cluster #2 (“cardiac arrest response”) focused on pathophysiology and emergency response systems. Cluster #3 and Cluster #5 focused on out-of-hospital and in-hospital cardiac arrest, respectively. Additional clusters covered shock-refractory ventricular fibrillation, sudden cardiac death, and outcome prediction. Supplementary Table 8 provides qualitative context for these co-citation clusters by summarizing representative influential references within each major cluster, including bibliometric signals, citation influence, qualitative summaries, and remaining research gaps. Cross-cluster co-citation links were most frequent among the “predicting neurological outcome,” “lethal ventricular arrhythmia,” and “cardiac arrest response” clusters. Timeline analysis (Figure 7(c)) depicts the temporal evolution of research themes from 2012 to 2025. Cluster #0 (“predicting neurological outcome”) expanded after 2018 and became the largest cluster in the most recent period. Mechanistic and electrophysiology-related clusters (e.g., #1 “lethal ventricular arrhythmia” and #8 “shock-refractory ventricular fibrillation”) appeared earlier and persisted in later periods. Process-oriented clusters (e.g., #3 “out-of-hospital cardiac arrest” and #5 “in-hospital cardiac arrest”) were observed throughout the study period. Cluster #16 (“quantitative electroencephalogram trend”) emerged in the later period.

Citation burst analysis identified 25 references with significant bursts, marking discrete periods of intensified citation activity. The burst references and their strengths are shown in Figure 7(d). References with the highest burst strength ranked among the top burst drivers in the network. Churpek et al. (2016) had the strongest reference citation burst (strength = 9). ²⁵ In multicenter ward datasets, random forest models outperformed logistic regression and the Modified Early Warning Score (MEWS) and reduced false alarms without loss of sensitivity. The study is frequently cited in the early warning system literature. Kwon et al. (2018) showed a strong reference citation burst (strength = 6.47). ²⁶ The study developed an RNN–LSTM model for in-hospital cardiac arrest prediction and reported higher performance than MEWS, logistic regression, and random forest using time-series data from four vital signs in multicenter datasets. Alonso-Atienza et al. (2014) showed a significant citation burst in ECG-based lethal arrhythmia detection. ³¹ Several recent burst references covered prehospital prognostication, ward early warning, and multicenter validation. Hirano et al. (2021) was among the burst references and used nationwide registry data for prehospital information–based prognostication in OHCA with initially shockable rhythms. ³² Lee et al. (2021) showed a strong citation burst and reported multicenter validation of a deep learning–based ward early warning score for in-hospital cardiac arrest. ³³ Across the study period, burst references spanned early algorithm development and ML-based risk identification, followed by deep learning–based early warning systems and later guideline- and registry-based, system-level studies.

3.7. Keyword analysis

We identified 2,761 author keywords and 7,322 Keywords Plus terms (Supplementary Table 2). Author-keyword co-occurrence analysis was performed in VOSviewer (Figure 8(a)), with total link strength (TLS) used to quantify co-occurrence intensity. “Machine learning,” “cardiac arrest,” “artificial intelligence,” and “deep learning” showed high TLS values and were among the most central terms in the network. The co-occurrence map delineated three major themes. Theme 1 focused on cardiac arrest and outcomes, with frequent terms including “mortality,” “prognosis,” and “neurological outcomes.” Theme 2 comprised AI-related terms (e.g., “prediction,” “early warning system,” “machine learning,” and “deep learning”). Theme 3 comprised signal-processing and arrhythmia-related terms, including “electrocardiogram,” “ventricular fibrillation,” and “arrhythmias.”

Figure 8.

Keyword co-occurrence, temporal evolution, and burst analysis in AI–cardiac arrest research. (a) Author keyword co-occurrence network. Nodes represent author-supplied keywords; node size is proportional to keyword frequency and edge thickness indicates co-occurrence strength. Colors denote thematic clusters. (b) Temporal overlay of author keywords. Keywords are color-coded by average publication year, highlighting the evolution of research hotspots in AI–cardiac arrest studies. (c) Timeline view of keyword clusters, showing the chronological development of major themes. (d) Top 25 keywords with the strongest bursts. Burst detection identifies keywords with rapid increases in frequency, indicating transient research frontiers and shifts in scholarly attention from 2012 to 2025.

The overlay visualization of author keywords (Figure 8(b)) depicts the temporal distribution of keywords across the study period. Earlier keywords were dominated by clinical topics (e.g., “cardiac arrest,” “resuscitation,” and “emergency medicine”), shown as blue nodes. In later years, data-driven terms (e.g., “machine learning,” “artificial intelligence,” “deep learning,” and “predictive models”) became more prominent and clustered in green–yellow nodes. “Electrocardiogram” and “ventricular fibrillation” also appeared as warmer-colored nodes in the overlay map, denoting later average publication years.

The timeline view (Figure 8(c)) summarizes the evolution of research topics over time. CiteSpace identified 10 keyword clusters; the largest were #0 “sudden cardiac death,” #1 “cardiac arrest,” and #2 “machine learning,” all of which remained active in recent years. Earlier periods were characterized by traditional machine-learning terms (e.g., “support vector machine” and “random forest”). After 2018, keywords increasingly included AI/deep learning, clinical deterioration detection, and CPR quality assessment. In later years, workflow-relevant prediction and resuscitation application terms appeared more frequently in the timeline map.

Keyword burst analysis (Figure 8(d)) summarizes time-varying research attention. During 2012–2016, burst terms centered on feature engineering and false-alarm management (e.g., “feature selection,” “false alarm reduction,” and “scoring system”). During 2017–2020, attention shifted toward ECG-driven machine learning; “electrocardiogram (ECG)” showed the highest burst strength (5.04). Bursts of “support vector machine,” “decision support,” and “early warning system” further point to increasing interest in workflow integration. Since 2020, burst terms have become more clinical, system-oriented, and outcome-focused. Bursts were observed for “Cerebral Performance Category (CPC)” and “brain injury” (neurological outcome–related terms) and for “emergency medical services (EMS)” (prehospital and system-level terms). Table 1 summarizes five major AI–cardiac arrest research domains based on co-citation, keyword co-occurrence, overlay, timeline, and burst analyses. Each domain is described according to its temporal trend, representative active countries/regions, methodological focus, clinical application area, and key implementation barriers.

Table 1.

Comparative summary of major thematic domains.

Cluster label and defining theme	Mean publication year and temporal trend	Representative active countries	Core methodological focus	Clinical application domain	Key implementation barriers
Automated external defibrillator and rhythm/shock-decision algorithms	Mean year 2016-2020; evolved from ECG feature engineering and shock/no-shock classification toward CNN/LSTM-based rhythm analysis during CPR.	Spain; Norway; USA	ECG waveform analysis, VF/VT detection, artifact filtering, shock-decision algorithms, CNN/LSTM models.	AED decision support, rhythm interpretation during CPR, defibrillation guidance.	CPR artifact noise, real-time signal instability, device integration, liability, and limited prospective deployment testing.
Post-cardiac arrest syndrome and neurological outcome prediction	Mean year around 2020; recent expansion toward CPC, brain injury, qEEG, and multimodal post-resuscitation prognosis.	USA; South Korea; China	ML/DL prognostic models using registry data, EHR, vital signs, EEG/qEEG, and multimodal clinical features.	ICU/post-resuscitation care, neurological prognostication, outcome-oriented risk stratification.	Heterogeneous CPC/mRS definitions, withdrawal-of-life-sustaining-treatment bias, missing data, calibration, and limited external validation.
In-hospital cardiac arrest early warning and deterioration detection	Mean year around 2017-2019; shifted from vital-sign/MEWS comparisons to deep learning and EHR-linked early warning systems.	South Korea; USA; China	Time-series vital signs, HRV, laboratory/EHR features, random forest, gradient boosting, RNN/LSTM, early warning scores.	General ward, ICU, rapid-response activation, IHCA prevention.	False alarms, alert fatigue, calibration drift, inconsistent monitoring frequency, missingness, and uncertain clinician response pathways.
Out-of-hospital triage AI and EMS dispatch recognition	Mean year 2017-2020; progressed from emergency-call recognition to prehospital triage, dispatcher support, and digital EMS systems.	Denmark; Japan; USA	Audio/NLP analysis of emergency calls, registry-based OHCA prediction, geospatial triage, dispatcher decision support.	Emergency call centers, EMS dispatch, OHCA recognition, transport and prehospital triage.	Language and regional generalizability, EMS documentation heterogeneity, latency, AI-human handoff, equity, and workflow integration.
Sudden cardiac death and ventricular arrhythmia risk stratification	Mean year 2019-2021; moved from ECG-derived risk markers to wearable, ambulatory, and EHR-based predictive models.	USA; China; South Korea	ECG AI, ambulatory/wearable monitoring, arrhythmia risk modeling, ensemble learning, deep learning, cardiovascular risk prediction.	Pre-arrest screening, sudden cardiac death risk prediction, cardiomyopathy/arrhythmia surveillance.	Low event prevalence, class imbalance, long-term follow-up needs, over-screening risk, uncertain clinical action thresholds, and external validation.

Abbreviations: AED, automated external defibrillator; AI, artificial intelligence; CNN, convolutional neural network; CPC, Cerebral Performance Category; CPR, cardiopulmonary resuscitation; DL, deep learning; ECG, electrocardiogram; EEG, electroencephalography; EHR, electronic health record; EMS, emergency medical services; HRV, heart rate variability; ICU, intensive care unit; IHCA, in-hospital cardiac arrest; LSTM, long short-term memory; MEWS, Modified Early Warning Score; ML, machine learning; mRS, modified Rankin Scale; NLP, natural language processing; OHCA, out-of-hospital cardiac arrest; qEEG, quantitative electroencephalography; RNN, recurrent neural network; VF, ventricular fibrillation; VT, ventricular tachycardia.

4. Discussion

We provide a global bibliometric map of AI research in cardiac arrest from 2012 to 2025. The intellectual base was both guideline-anchored and data-driven: resuscitation guidelines (notably ERC 2021 BLS) served as shared reference points, while highly cited and burst-driving studies increasingly emphasized multicenter early warning systems, registry-scale modeling, and clinically embedded decision support. Overall, these patterns suggest a maturing ecosystem in which AI is moving beyond algorithm-centric performance toward workflow-aware prediction, system optimization, and outcome prioritization, particularly neurological recovery. ³⁴ Accordingly, we interpret publication, collaboration, and thematic signals through a digital health lens, with a focus on translation to real-world resuscitation systems. The Discussion is structured around four themes: (i) global growth and the collaboration landscape of AI–cardiac arrest research; (ii) methodological and data shifts toward multimodal, validated models; (iii) the shift from survival to meaningful recovery, with an emphasis on neurological outcomes; and (iv) translation from prediction to trustworthy implementation, including workflow integration, safety, and governance.

4.1. Global growth and collaboration landscape in AI–cardiac arrest research

This bibliometric analysis indicates sustained growth in AI-enabled cardiac arrest research from 2012 to 2025, with accelerated expansion after 2020. The divergence between the observed publication counts and the fitted Price’s Law curve after around 2020 is compatible with this interpretation. Rather than being viewed only as a limitation of model fitting, this divergence may represent a phase of intensified growth, during which publication output exceeded the level predicted by the fitted exponential trend. This apparent acceleration was probably associated with several converging factors. These factors include the maturation of deep-learning methods, the wider availability of EHR-, ICU-, ECG-, and EMS-derived datasets, increased interest in digital health after the COVID-19 pandemic, and growing clinical demand for earlier risk stratification and workflow-aware decision support in emergency and critical care.^3,5,14–16 Beyond increased scholarly attention, this inflection likely reflects improved digital health readiness: emergency and critical care settings have become more instrumented and digitized, supporting higher-volume data capture and more routine computational modeling. ³⁵ Growth should not be interpreted as a simple “technology push.” Instead, it likely reflects co-evolution of (i) data infrastructure (e.g., maturing EHRs, monitoring platforms, and prehospital digital records), (ii) methodological capacity (e.g., deep learning and ensemble approaches), and (iii) clinical demand for timely risk stratification and decision support in high-stakes care pathways. This stepwise trajectory aligns with typical digital health maturation: technical feasibility precedes implementation-ready models, which in turn require governance, generalizability, and pathway integration. ³⁶ To contextualize our findings within the existing review literature, Supplementary Table 9 compares previous systematic reviews, scoping reviews, and meta-analyses with the present bibliometric analysis. Previous reviews have primarily focused on specific AI applications, model performance, and risk of bias in selected cardiac arrest and resuscitation contexts, such as early cardiac arrest prediction, defibrillation success prediction, OHCA care, and post-arrest outcome prognostication. In contrast, the present bibliometric analysis complements this literature by mapping publication growth, collaboration networks, influential journals and references, co-citation clusters, keyword evolution, and field-level translational gaps. This comparison supports the view that AI–cardiac arrest research is moving beyond isolated assessments of algorithmic performance toward workflow-aware validation, outcome-oriented endpoints, and implementation-focused digital health systems. At the collaboration-network level, the country collaboration map suggests a multipolar international structure rather than a single U.S.-centered hub-and-spoke network. Although the United States remained the leading contributor in publication volume and total citations, European countries showed dense intra-regional collaboration, and several Asia–Europe links were observed that did not necessarily pass through the United States. Spain’s high MCP% is noteworthy in this context, but it should be interpreted cautiously because MCP% is calculated from corresponding-author country assignments and is affected by the total number of country-assigned publications. Therefore, Spain’s MCP% should be understood as showing a high proportion of internationally coauthored publications among Spain-assigned records, rather than as definitive evidence, by itself, of leadership in multinational consortia. In cardiac arrest—where case mix, care pathways, and emergency medical systems are highly heterogeneous—collaborative infrastructure is not only an enabler of productivity but also a prerequisite for credible external validation and model transportability. ³⁷

4.2. Methodological and data shifts in AI–cardiac arrest: From features to multimodal, validated models

The shift from classical machine learning to deep learning and ensemble approaches reflects increasing reliance on complex, time-dependent physiological signals and heterogeneous clinical data. In digital health settings, the incremental value of deep models extends beyond improved discrimination. It also lies in learning clinically usable representations of streaming data (waveforms, trends, and sequences) that are difficult to capture with handcrafted features. ³⁸ This staged evolution—from early feature engineering and rhythm/signal analytics to deep learning–enabled rapid response and deterioration monitoring, and more recently to multicenter validation and registry-driven modeling—supports this interpretation. ³⁹

However, methodological progress must be interpreted in light of emergency-care realities. Author- and center-specific pipelines (e.g., preprocessing conventions, measurement frequency, and outcome definitions) can inflate apparent performance, and models optimized on local distributions may not generalize without explicit strategies to address heterogeneity, fairness, and generalizability under internal–external and external validation paradigms. ⁴⁰ This is why institutional and country-level collaboration matters: cross-site evaluation is inseparable from credible performance claims in safety-critical digital health. ⁴⁰ Importantly, the thematic shift toward “workflow-integrated” AI should not be conflated with widespread real-time testing in clinical workflows. Most AI–cardiac arrest studies in the current literature remain retrospective, simulation-based, or limited to offline validation, whereas prospective evaluations embedded in real clinical workflows remain uncommon. The randomized dispatcher-support study by Blomberg et al. represents a notable exception; however, the broader gap between algorithm development and implementation-ready clinical integration remains substantial.^14–16 Future comparative studies should evaluate AI-based prediction models against non-AI clinical benchmarks, including traditional early warning scores such as MEWS, cardiac arrest–specific tools such as CART, and conventional regression models. Such comparisons are needed to determine whether AI approaches provide incremental value beyond established risk stratification tools in discrimination, calibration, prediction lead time, interpretability, workflow burden, clinical actionability, and implementation impact.^{14,15,25,41,42}

In parallel, the shift from ECG-centered modeling to multimodal integration reflects a more systems-oriented framing of cardiac arrest. The warmer overlay colors observed for “ECG” and “ventricular fibrillation” should be interpreted cautiously. These terms remain central to AI applications in cardiac arrest prediction, rhythm classification, and resuscitation decision support.^13,14 However, these terms may have become less distinctive as emerging topics because they are increasingly used as routine input features, clinical descriptors, or baseline comparators in broader AI studies. Keyword mapping suggests convergence around (i) clinical outcomes, (ii) AI prediction/early warning systems (EWS), and (iii) rhythm/signal analytics. The burst timeline progressed from feature engineering and alarm management to ECG-centered false-alarm reduction, and then to post-2020 system-level, outcome-oriented topics (e.g., CPC/brain injury, EMS, and refined arrhythmia phenotypes). This trajectory aligns with a core digital health principle: clinical risk emerges from interacting subsystems (patient physiology, care processes, and context). It also maps onto the chain of survival, where prehospital recognition, in-hospital response, and post-arrest care contribute distinct data elements and intervention opportunities. ⁴³ Accordingly, multimodal strategies should be discussed not only as a technical upgrade but also as a pathway toward context-aware decision support, provided that data fusion includes transparent feature provenance, robust temporal alignment, and clinically meaningful endpoint definitions. ³⁹

4.3. From survival to meaningful recovery: Neurological outcomes and clinical relevance

The shift from mortality/survival to neurological outcomes reflects a clinically meaningful reframing aligned with patient-centered digital health objectives. Contemporary reporting standards increasingly prioritize survival with favorable neurological outcome, commonly defined using CPC or mRS thresholds. This neurological outcome definition requires careful interpretation. The Cerebral Performance Category (CPC) is widely used for post-arrest neurological outcome assessment, but it is a relatively coarse ordinal scale. Inter-rater variability has also been reported when CPC scores are assigned retrospectively from clinical records.^44,45 In many cardiac arrest studies, CPC is not operationalized as exact prediction across the full 1–5 ordinal scale. Instead, it is commonly dichotomized to distinguish favorable neurological outcome (CPC 1–2) from unfavorable neurological outcome (CPC 3–5). This distinction has important implications for AI model design and clinical utility. Binary classification may be more feasible for risk stratification and workflow decision support, whereas exact CPC prediction would require more reliable outcome labeling, ordinal modeling strategies, and validation of clinically meaningful thresholds. ¹⁴ This reinforces the view that survival alone is insufficient and that “meaningful recovery” should guide model objectives and translation. ⁴⁶ This outcome-oriented shift has direct methodological and clinical implications. First, neurological prognostication is intrinsically more challenging because outcomes are shaped by multi-stage post–cardiac arrest brain injury cascades (global ischemia–reperfusion injury followed by secondary injury processes) as well as by variation in care timeliness and interventions across settings. ⁴⁷ Second, it increases the need for temporally resolved, multimodal inputs (e.g., continuous vital signs/ECG and, where available, EEG monitoring) and explicit alignment between prediction time points and actionable decisions. ³⁰ Third, it raises the bar for external validation and subgroup robustness, particularly because cardiac arrest research spans both in-hospital and out-of-hospital systems with distinct decision points, data constraints, and documentation practices. ⁴⁸

From a digital health perspective, neurological outcomes are not merely another endpoint; they constrain translation. If models are intended to support triage, monitoring, or post-arrest management, they should be evaluated against clinically meaningful outcomes with standardized definitions. Their utility should also be expressed in workflow terms—namely whether predictions change escalation, resource allocation, or post-resuscitation pathway selection in time to act. ⁴⁸

4.4. From prediction to trustworthy implementation: Workflow integration, safety, and governance

A defining feature of recent work is the shift from standalone prediction to workflow-integrated clinical decision support. The mapped evidence suggests that the competitive frontier is shifting from “better AUROC” to “better decisions at scale,” with greater emphasis on external and prospective validation, workflow trade-offs (false alarms, calibration drift, and latency), and implementation outcomes alongside discrimination. This framing is particularly appropriate for cardiac arrest, where alerts must be timely, interpretable, and calibrated to resource constraints and escalation pathways, and where errors may carry disproportionate harm. ⁴⁹

In this context, interpretability and clinical safety should be treated as implementation requirements rather than optional add-ons. ⁵⁰ The field’s intellectual foundation appears both guideline-anchored and data-driven: resuscitation guidelines serve as shared reference points for practice targets and endpoint conventions, while multicenter registries and real-world datasets provide the empirical substrate for transportability testing and implementation-focused evaluation. ²⁴ However, guideline anchoring alone is insufficient for scalable translation. Without endpoint harmonization, benchmark datasets, and interoperable data models, heterogeneity in EMS–hospital documentation, care pathways, and measurement processes can induce dataset shift, degrade calibration, and compromise safety when models are deployed across sites. ⁴⁹ Therefore, strategies such as standardized endpoint definitions, harmonized data models, privacy-preserving collaboration, and rigorous reporting standards should be emphasized. ⁴⁸ Collectively, these considerations suggest that the next phase will be defined less by incremental AUROC gains and more by trustworthy system integration, in which models are embedded in workflows with explicit safety guardrails, transparent accountability, and evidence of benefit in real-world settings. ⁵¹

Overall, our mapping suggests that AI–cardiac arrest is a prototypical digital health domain in which impact depends on end-to-end integration across the care continuum (prehospital–ED–ICU), continuous multimodal data streams (vital signs/ECG/EEG), and governance for safety-critical deployment. The convergence of early warning/decision-support themes with guideline-centric co-citation reinforces a shift from “better AUROC” to “better decisions at scale,” favoring models that (i) demonstrate external and prospective validation, ⁵² (ii) quantify workflow trade-offs (false alarms, calibration drift, and latency), ⁵³ and (iii) report implementation outcomes alongside discrimination. ⁵⁴

Looking forward, high-impact research will likely prioritize: (i) outcome definitions aligned with meaningful recovery, particularly neurological outcomes and brain injury–oriented prognostication, supported by standardized endpoints and benchmark datasets ⁵⁵ ; (ii) interpretability and clinical safety as design requirements to enable accountable, clinician-trustworthy decision support ⁵⁶ ; (iii) multicenter registries and privacy-preserving collaboration (e.g., federated or distributed learning) to improve transportability across heterogeneous EMS and hospital systems ⁵⁷ ; and (iv) implementation-focused evaluation with robust reporting standards and ethical–safety frameworks to improve reproducibility, regulatory readiness, and real-world effectiveness. ⁵⁸ Together, these directions provide a practical roadmap to align AI innovation with the resuscitation chain and translate bibliometric “hotspots” into deployable digital health tools that improve meaningful survival. ⁵⁵ To make these priorities more actionable, Box 2 synthesizes the key bibliometric signals and translational gaps identified in this study and translates them into practical recommendations for future AI–cardiac arrest research.

Box 2. Practical recommendations for future AI–cardiac arrest research

Future research priority	Bibliometric signal from this study	Practical recommendation	Translational value
Standardize clinically meaningful outcomes	Reference co-citation Cluster #0 focused on predicting neurological outcome; keyword bursts included Cerebral Performance Category and brain injury.	Future studies should prioritize outcome definitions aligned with meaningful neurological recovery.	Aligns model development with patient-centered recovery and improves comparability across studies.
Develop standardized endpoints and benchmark datasets	The discussion identified endpoint heterogeneity across EMS, hospital, ECG, EEG, EHR, and post-arrest care settings.	Establish shared endpoint definitions, common data dictionaries, and benchmark datasets for AI-cardiac arrest research.	Enables fair model comparison, reproducibility, and more reliable evidence synthesis.
Treat interpretability and clinical safety as core requirements	The discussion emphasized trustworthy implementation, clinical safety, calibration, and workflow integration.	Interpretability, calibration, failure-mode analysis, and safety monitoring should be designed into models from the beginning.	Supports clinician trust and reduces risk when models are used in high-stakes resuscitation workflows.
Validate models across systems using multicenter registries and privacy-preserving collaboration	Country and institutional networks showed increasing collaboration, while the discussion emphasized transportability and external validation.	Use multicenter registries, external validation cohorts, and privacy-preserving approaches such as federated or distributed learning to test model robustness across EMS and hospital systems.	Improves generalizability across heterogeneous populations, documentation systems, and care pathways.
Move from retrospective prediction to implementation-focused evaluation	Recent clusters and burst themes emphasized early warning, EMS, decision support, workflow integration, and system-level outcomes.	Future studies should include prospective validation, workflow integration, false alarms, latency, calibration drift, and implementation outcomes.	Shifts the field from better prediction toward measurable improvement in care delivery and meaningful recovery.

4.5. Limitations

This study has several limitations. First, the findings are dependent on database coverage. Although WoSCC and Scopus were combined, PubMed Central, preprint servers, non-indexed open-access sources, and conference proceedings or engineering-oriented machine-learning outputs were not included as independent data sources. Therefore, rapidly evolving open-access or engineering-focused machine-learning developments may be underrepresented, particularly if they have not yet accumulated indexed citation records in WoSCC or Scopus. Second, author-, institution-, and country-level analyses are vulnerable to disambiguation errors, including homonyms, name variants, institutional mergers or renaming, and cross-database differences in affiliation reporting. Although targeted verification was conducted for the most prolific author-name entries where possible, comprehensive person-level disambiguation could not be performed for all author-name strings in the dataset. Therefore, individual author rankings should be interpreted cautiously as author-name-level productivity summaries rather than definitive person-level rankings. Third, network construction and clustering may be sensitive to parameter choices, including thresholds such as top-N per slice, counting methods, pruning strategies such as Pathfinder or merged-network pruning, and clustering algorithms. Accordingly, alternative settings may produce differences in map structure and cluster boundaries. Fourth, restricting inclusion to English-language articles and reviews may have underrepresented non-English evidence and excluded practice-oriented outputs, such as technical reports, implementation documents, and local registry reports. Fifth, although the thematic results showed increasing use of workflow-related terms, this pattern should not be interpreted as evidence that most AI models have undergone prospective testing in real-time clinical workflows. The gap between algorithmic development and prospective workflow integration remains largely unbridged in AI–cardiac arrest research. Most studies still rely on retrospective datasets, offline validation, or simulation-based evaluation rather than implementation-stage clinical testing. Finally, because this study was designed as a bibliometric analysis of publication-level metadata, it could not directly evaluate patient-level model development procedures, model performance, or implementation effectiveness. Therefore, the findings should be interpreted as a structured overview of the indexed literature rather than as evidence of the comparative clinical performance of specific AI models.

5. Conclusion

AI–cardiac arrest research is entering a maturing expansion phase, characterized by interdisciplinary journal linkages and a multipolar collaboration structure. Within this structure, the United States leads in publication volume and citations, Europe shows dense intra-regional collaboration, and Asian countries are increasingly contributing to regional and intercontinental partnerships. The field is moving beyond algorithm-centric development toward a digital health paradigm centered on workflow-integrated prediction, decision support, and clinically meaningful recovery. Future progress will likely depend not only on incremental improvements in discrimination but also on multicenter transportability, standardized endpoints—particularly neurological outcomes—prospective validation, and transparent reporting of workflow impact, safety, and governance.

Appendix.

Abbreviations

AI

artificial intelligence

CA

cardiac arrest

CPR

cardiopulmonary resuscitation

OHCA

out-of-hospital cardiac arrest

IHCA

in-hospital cardiac arrest

ECG

electrocardiogram

EEG

electroencephalography

ICU

intensive care unit

EHR

electronic health record

EMS

emergency medical services

CPC

Cerebral Performance Category

mRS

modified Rankin Scale

EWS

early warning system

MEWS

Modified Early Warning Score

VF

ventricular fibrillation

VT

ventricular tachycardia

BLS

basic life support

ED

emergency department

WoSCC

Web of Science Core Collection

NC

number of citations

NP

number of publications

SCP

single-country publications

MCP

multiple-country publications

TC

total citations

TLS

total link strength

JCR

Journal Citation Reports

IF

impact factor.

ORCID iD

Xidong Zhu https://orcid.org/0009-0004-5568-2245

Data Availability Statement

The datasets analyzed in this study were derived from the Web of Science Core Collection and Scopus databases. Processed data supporting the findings of this study are included in the article and its Supplementary Material. Additional details are available from the corresponding author upon reasonable request.*