Artificial Intelligence in Intensive Care Medicine: Bibliometric Analysis

Background: Interest in critical care–related artificial intelligence (AI) research is growing rapidly. However, the literature is still lacking in comprehensive bibliometric studies that measure and analyze scientific publications globally. Objective: The objective of this study was to assess the global research trends in AI in intensive care medicine based on publication outputs, citations, coauthorships between nations, and co-occurrences of author keywords. Methods: A total of 3619 documents published until March 2022 were retrieved from the Scopus database. After selecting the document type as articles, the titles and abstracts were checked for eligibility. In the final bibliometric study using VOSviewer, 1198 papers were included. The growth rate of publications, preferred journals, leading research countries, international collaborations, and top institutions were computed. Results: The number of publications increased steeply between 2018 and 2022, accounting for 72.53% (869/1198) of all the included papers. The United States and China contributed to approximately 55.17% (661/1198) of the total publications. Of the 15 most productive institutions, 9 were among the top 100 universities worldwide. Detecting clinical deterioration, monitoring, predicting disease progression, mortality, prognosis, and classifying disease phenotypes or subtypes were some of the research hot spots for AI in patients who are critically ill. Neural networks, decision support systems, machine learning, and deep learning were all commonly used AI technologies. Conclusions: This study highlights popular areas in AI research aimed at improving health care in intensive care units, offers a comprehensive look at the research trend in AI application in the intensive care unit, and provides an insight into potential collaboration and prospects for future research. The 30 articles that received the most citations were listed in detail. For AI-based clinical research to be sufficiently convincing for routine critical care practice, collaborative research efforts are needed to increase the maturity and robustness of AI-driven models.


Background
Artificial intelligence (AI) refers to a system that mimics human intelligence as characterized by the ability to perceive, reason, discover meaning, generalize, draw lessons from past experience and solve problems, or make decisions [1]. Machine learning (ML), natural language processing, and capability to visualize and recognize objects (computer vision) are all commonly used AI technologies [2].
ML is the dominant technique for implementing Al systems. ML refers to the science of programming computers that use statistical analysis techniques to create algorithms to learn from data [1]. Because it uses statistical models and algorithms to evaluate enormous training data sets, it is also referred to as "programming with data." Supervised and unsupervised frameworks are two different types of ML (Figure 1). AI has been used in the medical field in molecular biology, bioinformatics, and medical imaging and to support population health management, provide tailored diagnosis and treatment, monitor patients, guide surgical care, and predict health trajectories [1, 3,4].
The intensive care unit (ICU) is the most suitable ward among all the hospital wards to begin the transition to big data and the application of AI in research and even clinical practice in the near future. In ICUs, patients are closely monitored to detect physiological changes associated with deterioration that might require an appropriate reevaluation of the treatment plan. Nursing staff closely monitor patients in the ICU by charting neurological status, input, and output (including medication administration), etc. Bedside monitors facilitate this and continuously stream large amounts of data [5]. With advances in computer science, it has become possible to integrate and archive data in clinical documentation from various information systems and build a comprehensive system that is later transformed into a research database. Large public ICU databases include eICU and MIMIC databases. Open access to these databases encourages the use of AI technology in clinical research in intensive care medicine and the development of decision support tools.
The application of AI in critical care mainly involves disease diagnosis, prediction of disease progression (clinical deterioration), and characterization of specific disease phenotypes or endotypes in sepsis, septic shock, and acute respiratory distress syndrome (ARDS), etc.
Bibliometric study is a quantifiable informatics technique that analyzes the academic literature [6,7]. A general, quantitative, and qualitative overview of a certain topic can be provided via bibliometric analysis. It specifically identifies the most active authors, organizations, publications, influential studies, and international collaborations [7].

Goal of This Study
Although there has been a growing interest in critical care-related AI research, the literature is still lacking in comprehensive bibliometric studies that measure and analyze scientific publications globally. In this study, we aimed to (1) provide a holistic view of the research trends in AI application in ICUs; (2) highlight trending research topics in AI-related research focused on health care in ICUs; (3) highlight the contributions of prolific authors, leading countries, and the most productive academic institutions; and (4) provide an insight into potential collaboration and research directions in the future [8].

Methods
A bibliometric analysis study uses a mechanistic method to comprehend the global research trends in a certain field based on the outputs of the academic literature database. This approach distinguishes bibliometric analysis from reviews that are primarily designed to discuss the most recent advancements, challenges, and future directions of a particular topic [8].

Ethical Considerations
Ethics committee permission was not required, as this study was a retrospective bibliometric analysis of the existing published studies.

Data Source and Search Strategy
Data mining was conducted on March 18, 2022, using the Scopus database. Scopus is recognized as the largest abstract and citation database of peer-reviewed literature covering a wide range of subjects [8]. This study conducted a search for articles mentioning artificial intelligence or AI-related terms (neural network*, machine learning, deep learning, or natural language processing) and intensive care or ICU-related terms (critical care, critically ill, high dependency, or ICU) in the title, abstract, and keywords. The oldest publication dates to 1986, and the more recent ones are from 2022. The reproducible query string used for the search was: TITLE-ABS-KEY ("artificial intelligence" OR "neural network*" OR "machine learning" OR "deep learning" OR "natural language processing") AND ("intensive care" OR "critical care" OR "critically ill" OR "high dependency" OR "ICU") SEARCHED ON 18, March 2022.

Screening Strategy
The query string yielded 3619 documents. From those, only articles (2050/3619, 56.64%) were included (Table 1). A total of 4 duplicate articles were removed by using Stata (version 17; StataCorp) for data cleaning. The papers analyzed were restricted to those that (1) focused on intensive care medicine and (2) involved AI technologies. Two coauthors (SZ and RT) reviewed the titles of all studies as a pilot screening and removed irrelevant articles. Papers from the preliminary searches were categorized into include, exclude, or unsure. The abstracts and keywords of papers marked as unsure were further screened by 3 authors (SZ, RT, and MZ) and discussed until a consensus was reached in team meetings.
After screening the titles of all articles and abstracts, when necessary, 848 articles were excluded either because they did not focus on intensive care medicine or because they did not involve AI technologies. Finally, 1198 papers were included in the bibliometric analysis.

Overview
We used VOSviewer (version 1.6.18; Centre for Science and Technology Studies, Leiden University), a software tool for constructing and visualizing bibliometric maps, for bibliometric network visualization. The citation, bibliographical, and author keyword information of 1198 articles were exported to VOSviewer. The countries, authors, institutions, or keywords were included as objects of interest when creating maps using VOSviewer. We computed the growth rate of publications, research keywords, and publication patterns (countries, institutions, and journals). Bibliometric analyses were performed according to the instructions provided in the VOSviewer user manual [9].

Publication Output and Growth of Research Interest
The publication years were sorted, and the number of publications each year was counted using Stata 17.
The growth rate of publications over time was computed using the following compound annual growth rate formula: Growth rate = ([number of publications in the last year or number of publications in the first year]1/(last year − first year) − 1) × 100 [10][11][12].

Preferred Journals
We used the citation analysis function of VOSviewer and set the unit of analysis as "sources." Of the 443 sources (journals), 44 (9.9%) had >5 publications in total on AI in intensive care medicine. Journals were sorted according to the number of publications. We listed the number of citations, an important index of the degree of attention and influence of the published papers [13,14]. CiteScore 2020 was obtained from the Scopus Preview website [15].

Leading Countries, International Collaboration, and Top Institutions
The citation trends of the top 10 most productive countries, top 15 most productive journals, and top 15 most productive research institutions were analyzed. The frequency and percentage of publications or citations in each country, journal, and institution were computed. This information was provided by Scopus and analyzed using the citation and coauthorship functions in VOSviewer. In the coauthorship analysis, the country-to-country link strength showed the number of publications coauthored by 2 linked countries. We created a thesaurus file to merge same institutions with different name variants.
Google Mymaps [16] was used for world map drawing. Using information from the International Monetary Fund's World Economic Outlook, the gross domestic product of the countries was estimated to ascertain whether the economic power of the countries had an impact on the productivity of publications [17].

Author Keywords
A total of 2267 keywords from 892 (74.5%) articles were analyzed for author co-occurrence. Owing to the lack of author keyword information, the remaining 306 (25.5%) articles were excluded. A thesaurus file was created to merge synonymic single words and congeneric phrases. For example, coronavirus, coronavirus disease 2019, and sars-cov-2 were merged into 1 keyword and relabeled as covid-19. We identified high-frequency keywords and classified them into 3 categories: diseases, technology, and function.

Publication Output and Growth of Research Interest
In all, 1198 research articles were published in 36 years (1986-2022; Figure 2). The oldest publication dates to 1986 [18].
It is suggested that a strong interest in AI in intensive care medicine started from 2018 when the annual growth rate increased by 135.3%. Since then, there has been a steep increase in annual publications, which has caused the cumulative number of publications to increase rapidly. The growth rate was 3.

Preferred Journals
Our results revealed that the top 15 most productive journals were owned by 9 different publishers ( Table 2). The most productive journal was Artificial Intelligence in Medicine with 42 articles, covering 3.5% of the total publications, followed by Critical Care Medicine (40, 3.3%), PLoS ONE (35, 2.9%), and Scientific Reports (33, 2.8%). Critical Care Medicine, a Lippincott Williams and Wilkins journal, received the highest number of citations-1166. One of their articles, "An Interpretable Machine Learning Model for Accurate Prediction of Sepsis in the ICU," published in 2018, was the most cited article, with 247 citations.
A total of 9 journals had a CiteScore of ≥5 according to the CiteScore 2020 report. Critical Care Medicine (CiteScore 12.7) and JMIR Medical Informatics (CiteScore 1.59) had the highest and lowest CiteScores, respectively. Although ranked 7th with 26 articles in Scopus, the total number of citations and number of citations per document of Frontiers in Medicine were significantly lower than those of other journals.

Leading Countries, International Collaboration, and Top Institutions
The top 15 most productive countries contributing to the growth of AI in critical care research activities globally are listed in Table 3 and Figure 3. The United States was the leading country with 488 publications, accounting for 40.73% of all publications (1198) worldwide. With one-third of the total publications in the United States, China was the second most productive country (173/1198, 14.44%).
Machine learning was the most frequently encountered keyword, with 347 occurrences and 809 links to other keywords. In the same cluster of machine learning, general terms included big data (9 occurrences, 15 links), data science (6 occurrences, 13 links), prediction models (5 occurrences, 10 links), and clustering (5 occurrences, 14 links). Machine learning co-occurred with ICU-related professional keywords, including critical care medicine, acute respiratory distress syndrome, acute respiratory failure, endotracheal intubation, mechanical ventilation, and personalized medicine.   We categorized the keywords into 3 different categories based on the diseases in ICU, ML technologies used, and the function of ML in the research ( Table 6). The top 5 diagnoses were sepsis, COVID-19, acute kidney injury, ARDS, and cardiac arrest. The top 5 AI technologies were ML, AI, deep learning, decision support systems, and neural networks. The top 5 functions were prediction (of clinical deterioration or mortality), mortality, monitoring, disease prognosis, and (disease phenotype) classification.

Citation Analysis
The 30 most cited articles among the 1198 articles on AI in intensive care medicine between 1986 and 2022 are presented in Table 7. The total number of citations and average number of citations per year are given. Of the 30 most cited articles, 10 (33%) were related to sepsis, one of the hottest topics in intensive care, and 11 (37%) articles were related to the use of AI technologies for the prediction of acute kidney injury, sepsis, hypotension diagnosis, clinical outcomes (mortality, survival, and ICU length of stay), and complications.

Discussion
Our study used a bibliometric method to analyze AI in intensive care medicine research by examining publication output, the growth of research interest, preferred journals, leading countries, international collaboration, top institutions, author keywords, and citation analysis.

Publication Output and Growth of Research Interest
Since the oldest publication in 1986, publications on AI in intensive care medicine had a slow growth for 30 years. The turning point appeared in 2018 when there was a significant growth in the interest in AI in intensive care medicine. This lags 6 years behind the rapid growth in the interest in AI in general medicine that started in 2012 [20]. This is likely owing to concerns regarding the safety and accountability of the AI model in critically ill patients. The AI technologies that emerged from 2014 to 2018 such as autonomous robots, voice recognition, neural networks, and ML provided unprecedented opportunities to predict, diagnose, and manage diseases. Large public critical care databases such as MIMIC and eICU became readily available to researchers in 2016 and 2018 [5,21]. Advancements in AI technology and large databases have contributed to the steep increase in annual publications since 2018. Based on the publication trend, it is anticipated that the annual publications will continue to increase.

Preferred Journals
Of the top 15 productive journals, 9 (60%) had a CiteScore of ≥5, which suggested that critical care medicine-related AI research is favored by the top journals in critical care and medical informatics. These include Critical Care Medicine, Critical Care and IEEE Journal of Biomedical and Health Informatics. Authors who want to publish critical care medicine-related AI research could first consider the top productive journals listed. CiteScore, an Elsevier-Scopus alternative to the Clarivate Analytics Impact Factor, is a metric for assessing journal impact based on citation data from the Scopus database. However, CiteScore is not the only factor considered when deciding which journal to publish in. Authors should consider the ability of the journal to disseminate the research work to the right audience and contribute to the progression of the field [8].

Leading Countries, International Collaboration, and Top Institutions
The fact that 9 of the top 15 productive institutions are among the top 100 best universities demonstrated that AI in critical care medicine has received attention at the top universities worldwide. Authors could consider joint research with those institutions or apply for their visiting scholar or educational programs.
When the distribution of publications by countries was examined, high-income countries were the leading force in critical care medicine-related AI research. The top 10 most productive countries are among the top 25 in terms of world gross domestic product, which suggested that the economic power of the countries affects the productivity of their publications. This result is the same as that of the bibliometric research on many other medical subjects [13,22,23]. About 60% of the global publications were contributed by the United States, China, and the United Kingdom, indicating that these 3 countries contributed the most to AI in critical care research. These countries also had the highest citations, although China had relatively low citations per document compared with the United States and the United Kingdom.
The number of citations was lower than that of other research hot spots in critical care medicine [24]. This is likely because of the limitations of AI-related studies in critical care, which include low maturity of AI in real-world application [25] and a lack of external validation process, prospective evaluation, and clear protocols to examine the reproducibility of AI solutions [26].
Coauthorship analysis revealed that the United States, the United Kingdom, Italy, and China were the most affiliated countries with >90 coauthorships. The diversity of research partners, a high proportion of foreign postgraduates or visiting scholars, and adequate research funding were all factors that contributed to improved international collaboration. To ensure the sustainability of international collaboration, a flexible and stable research policy is also crucial [8].

Author Keywords and Citation Analysis
According to the author keywords in the identified categories, the top domains of disease covered in critical care medicine-related AI research were sepsis, COVID-19, acute kidney injury, ARDS, and cardiac arrest. These most prevalent ICU conditions have become a popular target for AI algorithms.
The top functions of AI include the prediction of clinical deterioration or disease evolution or mortality, monitoring, disease prognosis, and disease phenotype or subtype classification. The literature reported other important functions such as disease identification and guiding decision-making (reinforcement learning) [26]. The keywords of high occurrence in the titles of the 30 most cited articles include sepsis, prediction, early detection, and clinical decision support. Keyword analysis showed that machine learning was frequently related to respiratory diseases, especially COVID-19. The most widely used AI technologies in critical care include ML, deep learning, decision support systems, and neural networks.

Limitations
Our research provided a general review of the research trends and hot spots in critical care medicine-related AI research, highlighted the most productive countries and academic institutions to facilitate potential collaboration, and provided directions for future research. However, this bibliometric study has several limitations.
First, even though we included the most widely used AI technologies and made an effort to be specific about AI-related terms in the search keywords (eg, neural network, machine learning, deep learning, and natural language processing), they were still quite general and did not include all AI technologies. Because of restricting the search to only those keywords in the title, abstract, and keywords, the search result may not have covered all AI in critical care medicine-related studies available on Scopus. Furthermore, because of missing author keyword information, the co-occurrence analysis of author keywords included only 74.4% (891) of the 1198 articles.
Second, our study used only the Scopus database, which is the largest abstract and citation database that we think should be sufficient for our analysis. Bibliometric analysis using multiple data sources such as the Web of Sciences, PubMed, and Google Scholar will be more comprehensive. For instance, Web of Science has a feature called "hot paper" that is not available in Scopus, which automatically displays the most popular articles in the field [27]. The hot paper feature displays important papers that were identified immediately after publication, as indicated by a sharp rise in the number of citations [8].
Finally, we did not include papers published in the form of conference papers, reviews, editorials, notes, letters, book chapters, short surveys, or data papers owing to concerns about the low clinical readiness of those publications. As a result, we may have missed relevant studies published in forms other than articles. Because AI technology is a cutting-edge and rapidly evolving research area, papers published in conference proceedings and letters may have reviewed the latest updates in the field.
Future bibliometric analyses could use more specific AI technology terms in the research keywords; use other databases such as Web of Sciences, PubMed, and Google Scholar; and include conference papers in the type of articles to explore more potential papers.

Conclusions
Our study has provided an overview of the research trends of AI in critical care medicine based on 1198 publications retrieved from the Scopus database. Publication growth was rapid in the last 5 years and is expected to further increase. We have reviewed countries and academic institutions (eg, the United States, China, and the United Kingdom) that have a substantial number of publications and solid international collaborations. This provides potential collaboration opportunities to other countries, especially to low-and middle-income countries that lack AI technologies but have an increasing demand for health care resources.
We have discussed several conditions in critical care that are currently actively explored using AI technology, such as sepsis, COVID-19, acute kidney injury, ARDS, and cardiac arrest. AI research hot spots in critically ill patients involve detecting clinical deterioration, monitoring, predicting disease evolution, mortality, disease prognosis, and classifying disease phenotypes or subtypes. The most widely used AI technologies in critical care research are ML, deep learning, decision support systems, and neural networks. The 30 articles that received the most citations between 1986 and 2022 have been listed in detail.
AI research on critical care is a rising hot spot in both critical care and AI research, with potential applications being demonstrated across various domains of critical care medicine. However, the development and implementation of AI solutions still face many challenges. AI research application in clinical settings has been constrained by a lack of external validation processes, prospective evaluation, and clear protocols to examine the reproducibility of AI solutions. For AI-based clinical research to be sufficiently convincing for routine critical care practice, collaborative research efforts are needed to increase the maturity and robustness of AI-driven models [26].

Conflicts of Interest
None declared.