The start of critical care as a medical subspecialty can be traced back to a polio epidemic during which a substantial number of patients needed prolonged mechanical ventilation [1]. Over time, the field of critical care has experienced significant growth and continual evolution. Research in this field has played a pivotal role in unraveling the complexities of numerous diseases and treatment modalities, driving substantial advancements in clinical practice over the past decades [2]. Groundbreaking studies have investigated critical areas such as sepsis, mechanical ventilation, acute lung and kidney injuries, intensive care unit (ICU) delirium, and sedation in critically ill patients [3].

These research studies have often been conducted in traditional ways such as prospective and randomized controlled trials [4], cohort and observational studies, clinical trials [5], and clinical and translational research [6]. These traditional methods have revolutionized patient care and improved outcomes significantly. For instance, the implementation of protocol-driven, goal-directed management of sepsis and appropriate fluid therapy has led to remarkable reductions in mortality rates [7,8], and these findings have been integral in developing evidence-based practice guidelines that are now the gold standard [9,10].

Despite their undeniable merits, traditional research methods in intensive care also come with several limitations [11]. Clinical trials are known for their high costs [12], stringent standardization requirements, and ethical oversight [13]. Data collection can be laborious, prone to human errors, and constrained in terms of quantity and granularity [14]. Moreover, obtaining patient consent for most randomized controlled trials in the ICU poses challenges [15], necessitating alternative consent models. These limitations have become increasingly apparent as medical complexity continues to grow exponentially [16].

The advent of electronic health records (EHRs) has heralded a new era in clinical research by facilitating the digitization of health care systems [17]. In this era of data science, a more integrated approach can be adopted, using machine learning (ML) algorithms to tackle the complexity of critical illness [18]. Open-access databases (OADs), such as the Medical Information Mart for Intensive Care (MIMIC) database [19] and the Philips eICU Collaborative Research Database (eICU-CRD) [20], have played a transformative role by enabling free data sharing.

The concept of free and open databases plays a pivotal role in promoting data sharing and advancing medical knowledge in accordance with the findable, accessible, interoperable, and reusable (FAIR) guiding principle. The FAIR principles, which emphasize that data should be findable, accessible, interoperable, and reusable, are essential for fostering a collaborative and transparent scientific research environment [21,22]. By removing barriers to access, free, and open databases allow researchers, regardless of their affiliations or resources, to contribute to and benefit from the collective pool of information. Accessibility fosters inclusivity and diversity in research, promoting a broader range of perspectives and approaches to medical challenges. This democratization of knowledge leads to a more equitable distribution of information. Researchers can now leverage these vast repositories of information for ML and artificial intelligence studies, marking a departure from traditional intensive care (TIC) research approaches.

Conducting a literature review [23] to investigate the disparities between traditional ICU research and studies based on open-access data sets holds significant importance as it provides a comprehensive understanding of the strengths and limitations of the latter. However, conventional methods of literature reviews and bibliometric analysis have their limitations, especially when dealing with large-scale literature due to computational complexity and the labor-intensive nature of manual interpretations [24-26]. To address these challenges, natural language processing (NLP) offers a promising avenue, while topic modeling techniques can be used to extract various topic themes from extensive data sets [27,28].

Built on the foundations of bidirectional encoder representations from transformers (BERT), BERTopic introduces a novel approach to topic modeling [29,30]. Unlike traditional unsupervised models like latent Dirichlet allocation, which rely on “bag-of-words” model [31], BERTopic overcomes the problem of semantic information loss, significantly enhancing the accuracy of generated topics, and providing more interpretable compositions for each topic, which greatly facilitates the classification of topics.

With the aid of BERTopic, this study aims to shed light on the disparities and commonalities between studies conducted through OADs and TIC research. By analyzing the overall trends and patterns in these 2 groups, we seek to identify knowledge gaps and explore avenues for complementary contributions between these research approaches.

A total of 227,893 records were identified from the WoS database on January 18, 2023, of which 195,463 full records were subsequently processed. Records were excluded if they are not “article” or “review” or if they do not contain keywords related to intensive care. After exclusions, 145,426 articles were identified as TIC studies and 1301 articles were categorized as OAD (Figure 1).

We examined the number of articles published per year to analyze the trends in TIC and OAD studies (Figure 2). Over the past 2 decades, TIC studies have experienced exponential growth, culminating in a peak of 16,378 articles in 2021. A subsequent decline in the number of publications occurred in 2022, likely attributable to delayed indexing within the WoS database and a reduction in COVID-19–related studies as the pandemic stabilized [38]. In contrast, the first OAD study emerged in 2003, with its popularity experiencing a consistent upsurge since 2018. Nonetheless, the number of OAD publications remains markedly lower in comparison to TIC publications.

The OAD studies were published most frequently in new open-access journals such as Frontiers in Medicine, Frontiers in Cardiovascular Medicine, and Scientific Reports while the TIC studies were published most frequently in established journals like Critical Care Medicine, Intensive Care Medicine, and Critical Care (Multimedia Appendix 2). Further analysis of keywords from the abstracts showed 2.4% (3492/145,426) TIC studies were meta-analyses or systematic reviews, while only 0.08% (1/1301) OAD study was in this category. There were 5.61% (73/1301) OAD studies, and 7.43% (10,799/145,426) TIC studies that had the keyword of “cost.” Examples of the data fields that are available within OADs such as MIMIC and eICU-CRD are listed in Textbox 1. Some information fields such as end-of-life goals and values and health care provider psychology are not available within the current EHRs extracted for OADs.

The UMAP algorithm was used to project the high-dimension corpus to 4 dimensions and allowed exploration of the vocabulary patterns between the OAD and TIC studies (Figure 3). The projection values are represented by the x-axis, while the densities are represented by the y-axis. The presence of considerable overlap between TIC studies and OAD studies suggests that they share a substantial number of common terminologies, which may correspond to similar research topics. Nonetheless, TIC studies exhibit a more extensive coverage than OAD studies, which may stem from broader research scope and extended research duration.

Subsequently, the BERTopic model was then used to generate 30 topic IDs (Figure 4). The internal commonalities of each topic ID were reviewed by critical care physicians and assigned a specific subtopic in intensive care research. The model was able to automatically classify the topics with high interpretability and the topic components were interpreted with ease. For instance, components in topic ID 5 consist of, in decreasing order of weightage: “learning,” “model,” “machine,” “machine learning,” “models,” “data,” “prediction,” and “performance.” This topic was consequently labeled “predictive model” (topic ID 5 in Multimedia Appendix 3).

The overall topic distribution in TIC studies was more uniform, while the OAD studies tended to be concentrated on several topics including topic ID 2 (kidney injury), 5 (predictive model), and 13 (sepsis). Some topics that were missing in OAD studies included 6 (pediatrics care), 21 (viral infections), 23 (health personnel and psychology), and 28 (nutrition and rehabilitation).

The similarity matrix shows that there was little overlap between the topics (Multimedia Appendix 4). To facilitate the interpretability of the categories, the overlapping topic IDs were merged to form the final 22 topic families (Multimedia Appendix 3).

Topics such as “healthcare associated infection,” “thoracic surgeries,” and “pregnancy related” research were among the more frequently discussed 15 topics in TIC studies but have limited publications in OAD studies. The topics of “predictive model,” “obesity,” and “fungal infections” were popular in OAD studies but not the TIC studies. Overall, the topic distributions of the TIC studies were distributed more evenly with the topic family of sepsis accounting for a quarter of the studies, while publications in the OAD studies were heavily skewed toward the predictive model (>40%) and sepsis (>30%; Figure 5).