The protocol adheres to the reporting guidance provided in the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) [16] (Table S1 in Multimedia Appendix 1). The design of this study is conceived according to the Joanna Briggs Institute Reviewers’ manual for evidence synthesis [17].

We selected studies based on the following criteria (Textbox 1).

First, the literature was searched for open access ICU databases. A recent review systematically identified publicly available adult clinical ICU databases [9]: Amsterdam University Medical Centers Database (AmsterdamUMCdb) [7], eICU Collaborative Research Database (eICU-CRD) [18], High time-resolution ICU dataset (HiRID) [19], MIMIC clinical databases version II, III, and IV [10,20,21].

We then used PubMed-Medline (US National Library of Medicine), Embase (Elsevier), and Web of science (Clarivate Analytics) to search articles published from the inception of each database to August 1, 2022. The search terms of the different databases are included in Table S2 in Multimedia Appendix 1. A senior librarian [GK] developed and validated search queries. Finally, the references of the included articles were examined for potentially relevant medical articles that may have escaped the literature search.

Two independent reviewers, a senior intensivist [BP] and a junior intensivist [JK], conducted the search strategy and retrieved the references. An initial selection was made after examining the titles and abstracts of the medical articles resulting from the search strategy. The reviewers [BP and JK] performed an initial blind and independent selection of the articles. Subsequently, the reviewers met to reach a consensus on any disagreements. If the 2 primary reviewers could not reach a consensus, the third reviewer, a health data engineer (AL), intervened to resolve the disagreements through arbitration. The disagreements typically revolved around the eligibility criteria for study inclusion. The selection process was facilitated using the Rayyan software [22].

Medical articles were referenced by their Digital Object Identifier, the name of the first author, and the date of publication. Two reviewers [BP and JK] independently extracted data using a custom electronic form designed specifically for this study. The form was pilot tested with a random sample of 20 articles to ensure its effectiveness. Upon achieving consistent data abstraction (κ≥0.8) [23], reviewers proceeded with data extraction for the entirety of included articles using the Goupile software [24].

Data for journal research fields and impact factors were obtained from Web of Science [25]. In addition, research topics were predefined and listed in the research protocol. Collected data details are outlined in Table S3 in Multimedia Appendix 1. Results were aggregated into a table, organized with 1 column per database and 1 row per variable.

A descriptive analysis was then carried out on the gathered data using R (version 4.1.2; R Foundation for Statistical Computing) [26]. Categorical variables were summarized as counts and frequencies per category, while quantitative variables were expressed as medians and interquartile ranges. Bar charts were used to visualize the distribution of categorical variables and histograms and density curves for the quantitative variables. In addition, the temporal evolution of the studied variables was analyzed.

From the inception of these databases up to August 2022, a total of 4466 publications were identified. After excluding duplicates, 2063 (46.2%) titles and abstracts were retained. Among these articles, 1362 (66%) met the inclusion criteria. Ultimately, 1307 (96%) articles were selected for the final analysis after thorough reading. The selection flowchart is shown in Figure 1, with the characteristics of the included studies are detailed in Table 1.

Since 2016, there was a consistent rise in the number of publications using these databases. Most studies were published after the year 2020, with 967 (74.0%) publications. The MIMIC databases were the most frequently used. No publications using the HiRID database were identified (Table 1, Figure 2). China and the United States were the leading countries in terms of number of publications, with 689 (52.7%) and 370 (28.3%) articles, respectively. European countries contributed to 109 (8.3%) publications, led by the United Kingdom leading with 39 (3.0%) publications (Table 1, Figure 3).

The median impact factor of publications was 3.8 (IQR 2.8-5.8) (Table 1). As for the journal field, the most commonly represented journals were those in medicine, computer science, and intensive care, with 712 (54.5%), 373 (28.5%), and 242 (18.5%) publications, respectively (Table 2).

Studies included a median number of 4282 (IQR 1468-12,740) patients with a maximum sample size reaching up to 219,306 by combining patients from several databases. The median inclusion period was 11 (IQR 11-11) years (Table 1).

Research predominantly explored the cardiovascular system, particularly hemodynamics with 333 (25.5%) publications and infectious diseases, specifically sepsis with 324 (24.8%) publications. Renal failure and metabolic disorders were also heavily studied, with 231 (17.7%) publications, along with respiratory failure and mechanical ventilation, which had 203 (15.5%) publications. Notably, many studies investigated overall mortality among critically ill patients admitted to the ICU, with 178 (13.6%) publications (Table 3).

Furthermore, 422 (32.6%) studies investigated the entire population admitted to the ICU. Most studies focused on patients experiencing specific organ failures, represented in 818 (62.6%) publications (Table 4). Various exposure or prediction factors were evaluated. Biological data and vital signs were the most analyzed, featuring in 673 (51.5%) and 428 (32.7%) studies, respectively. Demographic information, medication prescriptions, and comorbidities were also widely included, with 332 (25.4%), 278 (21.3%) and 271 (20.7%) publications, respectively (Table 5). The most frequently assessed primary outcome was mortality, observed either within the ICU, the hospital, or after a certain period of time, across 878 (67.2%) studies (Table 6).

Regarding the study objectives, 713 (54.6%) studies addressed an inference question, while 594 (45.4%) studies targeted a prediction question (Table 7, Table 8). Supervised learning and deep learning methods were the most commonly applied machine learning techniques, with 432 (33.1%) and 247 (18.9%) publications respectively (Table 7).

When examining specific models, logistic or linear regression was the predominant choice for inference questions with 396 (55.5%) publications (Table 9). Regarding prediction questions, logistic or linear regression was the most widely used classical machine learning method, surpassing survival models and generalized models with 281 (47.5%), 29 (4.9%), and 12 (2%) publications respectively. Neural networks were the most frequently applied advanced machine learning methods outpacing tree-based methods and boosting methods, with 237 (40.0%), 186 (31.4%), and 169 (28.5%) publications respectively (Table 10).

Inference studies reported statistically significant results in 84% of cases. Furthermore, the prevailing performance measure in prediction studies was the area under the curve (AUC) with 436 (73.4%) studies. The median AUC value was 0.840 (IQR 0.780-0.890; Table 8).

Preexisting clinical scores were applied for comparison with the performance of prediction models implemented in the studies. The most used scores were SAPS (Simplified Acute Physiology Score), SOFA (Sepsis-related Organ Failure Assessment), and APACHE (Acute Physiology and Chronic Health Evaluation) scores, with 55 (42.6%), 24 (18.6%), and 14 (10.9%) publications, respectively. The median AUC for these scores was 0.733 (IQR 0.677-0.781; Table 8).

This review examined clinical publications from open databases in the field of intensive care. We observed a consistent increase in the number of publications from these databases since 2016, with the majority of articles being published after 2020. MIMIC databases were the most frequently used, while the countries contributing the most were China and the United States. Most studies were published in journals with an impact factor ranging from 3 to 6. Cardiovascular and infectious topics, particularly those related to hemodynamics and sepsis, were the most represented. Other significant subjects included renal failure coupled with metabolic disorders and respiratory failure with mechanical ventilation. The most studied outcome measure was mortality in the ICU. Biological data, vital signs, demographic details, medication prescriptions, and comorbidities were extensively used as exposure or predictor factors.

Regarding statistical methods, logistic regression was the most used model for both inference and prediction questions. Neural networks were the most frequently used advanced machine learning methods surpassing supervised and reinforcement methods. A majority of the inference studies presented statistically significant results. In prediction studies, the most recurrent performance measure was the AUC, with a median value of 0.840.

The rising number of publications using open databases is steadily increasing, highlighting the potential impact of these resources on clinical research within intensive care. Collaborative initiatives have led to the emergence of large open-access databases, enhancing data availability in this field. Researchers have now access to data from numerous centers, enabling the production of robust results through larger sample sizes. Furthermore, this access has significantly boosted statistical power, enabling the study of specific topics and populations that were previously overlooked due to insufficient sample sizes. In our review, the median sample size in the studies was 4282 (IQR 1468-12,740), in contrast to typical sample sizes in prospective studies among critical care patients, that barely reach a few hundred patients. In a literature review by van de Sande et al [27] studying machine learning in intensive care, the median sample size was 179 (IQR 94-1411) and 142 (IQR 40-380) across all prospective observational and clinical studies, respectively.

The wealth and diversity of information contained in these databases have opened a broad spectrum of research possibilities, resulting in a wide array of topics covered in publications. Our review categorized these into over 20 different categories. Cardiovascular, infectious, respiratory, and metabolic issues emerged as the most commonly studied research domains, mirroring the reality of clinical practice in intensive care [28]. Mortality in ICU was the most frequently analyzed outcome measure. In a literature review by Syed et al [14], which examined the application of machine learning using the MIMIC dataset, mortality prediction was also the most studied outcome measure, followed by sepsis prediction, cardiac events, and acute kidney injury prediction.

Despite the significant number and diversity of publications, the clinical relevance of studies from these databases remains heterogeneous. The results were often statistically significant but had limited effect sizes. Nevertheless, there are few criteria to objectively evaluate the clinical relevance of a publication. In addition, the quantitative nature of our analyses did not allow us to fully explore this aspect. A qualitative methodology specifically assessing clinical relevance would provide more insights and discussion points.

Recently, a bibliometric analysis study using machine learning and natural language processing methods Bidirectional Encoder Representations from Transformers (BERTopic) compared research topics in traditional intensive care unit studies and those conducted with open access databases. A total of 1307studies from open access databases were identified versus 145,426 traditional studies. Among these 1307 studies, the most predominant topic was sepsis and kidney injury, and over 40% of the articles were based on predictive models [29].

In recent years, due to advancing technology, data availability, and the need to analyze increasingly larger databases, machine learning has emerged to develop increasingly accurate predictive models. Machine learning is a field of computer science and a part of artificial intelligence that defines both the science and engineering for which computer systems can analyze data and learn from the information contained within [30].

In our study, 594 out of 1307 (45.4%) publications used at least 1 machine learning model, with a significant increase in their implementation after 2015. The most commonly used machine learning method was logistic regression and neural networks. Literature review by Shillan et al [31], summarizing the characteristics and results of machine learning methods used in intensive care, found similar results. Nearly half of the studies identified in that review were published after 2015, and the most commonly used methods were neural networks, support vector machines, and decision trees.

Although the use of machine learning models in intensive care has increased in the scientific literature, their adoption in actual clinical practice remains limited or even absent at present. Machine learning models are often complex and require specialized skills to develop and implement. Their interpretation also poses a significant barrier to their use, with the phenomenon known as the “black box” issue [32]. Furthermore, these models need to be rigorously validated to ensure their reliability and accuracy. External validation of models on independent datasets is crucial to evaluate their performance under real-world conditions. In our review, the vast majority of publications analyzed came from databases implemented in the United States of America, limiting the generalizability of these models to the rest of the population. In addition, only 92 out of 2063 (4.5%) articles focused on external validation using an independent dataset, which is consistent with Shillan et al findings [31]. Furthermore, prospective clinical evaluation of these models is still rare. In a recent review, only 10 out of 494 (2%) articles clinically evaluated artificial intelligence in real clinical settings, with 5 studies being randomized clinical trials [27]. Finally, the use of machine learning models raises ethical and regulatory questions, particularly concerning data privacy, automated decision-making, and liability in case of errors. Clear guidelines and regulations must be established to govern the use of these models in a clinical context.

This study, to our knowledge, is the first to examine the contribution of open databases in clinical research in intensive care. This study follows a rigorous methodology, and its research protocol was made public before the study [33]. Furthermore, this review adopts a comprehensive approach by including all clinical research publications generated from open ICU databases. By examining a wide range of publications, it provides a detailed overview of research themes, methodologies used, and results obtained. This can assist clinicians and researchers in understanding how these databases can contribute to improving patient care in intensive care and offering a solid foundation for future research and discussions.

However, it is important to acknowledge the limitations of this literature review. First, only publications written in English were included, potentially introducing a linguistic bias. Second, there exists and inherent risk of publication bias in literature reviews. Published and accessible studies may not represent all research carried out on open databases in intensive care. Studies with negative or nonsignificant results are less likely to be published, potentially resulting in an overestimation of positive results. While several open ICU databases were included in the study, it is possible that there exist other databases that were not taken into account. This could limit the representativeness of the sampled studies. Finally, despite exhaustive research, no publications from the HiRID database were identified. With these limitations in mind, it is important to consider the results of this study with caution and interpret them in the appropriate context.

Open databases in intensive care have facilitated clinical research and provided new perspectives for enhancing care in intensive care. The abundance of scientific outputs resulting from these databases and the diversity of topics addressed highlight the importance of these databases as valuable resources for clinical research and suggest their potential impact on clinical practice in intensive care. However, the quality of studies and their clinical relevance remain highly heterogeneous among publications and are challenging elements to evaluate.