Kurzfassung

BACKGROUND: Open access intensive care unit (ICU) databases, including MIMIC-III, MIMIC-IV, the eICU Collaborative Research Database (eICU-CRD), AmsterdamUMCdb, HiRID-high time resolution ICU data set, and the Salzburg Intensive Care Database (SICdb), have fundamentally transformed critical care research over the past decade. Despite their widespread use, no comprehensive bibliometric analysis has systematically mapped the scientific output generated by studies relying on these resources.

METHODS: A structured search was conducted in PubMed and Web of Science covering 2018 to 2025, retrieving original articles citing at least one of six defined open access ICU databases in title or abstract. Database selection followed the methodology of two prior systematic reviews of publicly available adult ICU datasets. Structured fields covering database used, machine learning (ML) methods, study design, clinical outcomes, AUROC values and patient counts were extracted via a semi-automated AI pipeline (DeepSeek API, multi-label JSON schema). The extraction pipeline was validated against a manually coded reference sample of 110 publications. Bibliometric analyses used R (bibliometrix v4.0.1) and Python. Citation counts, open-access status and SCImago journal quartile were retrieved via OpenAlex and SCImago to characterise the visibility and venue quality of the corpus. Four network analyses were performed: database-to-ML-method co-occurrence, country-to-database, international collaboration, and clinical topic co-occurrence networks.

RESULTS: A total of 4,648 publications were included. Annual publication counts increased 45-fold from 42 (2018) to 1,907 (2025), accelerating markedly after 2021 following the release of MIMIC-IV. MIMIC-IV dominated database attributions (54.6%), followed by MIMIC-III (30.0%) and eICU-CRD (13.5%); European databases (AmsterdamUMCdb, HiRID, SICdb) collectively accounted for only 1.9%. China contributed 65.6% of publications (n = 3,048) and the USA 10.6%. Median citations per paper were 5.0 (1.50 per year). 76.4% of publications were open access and 49.3% appeared in a Q1 SCImago-ranked journal. AUROC showed no meaningful correlation with citation impact (p = 0.10). Logistic regression was the most frequent ML method (38.5%), followed by neural networks (15.6%) and gradient boosting (13.8%). Retrospective observational designs dominated (67.9%), with mortality prediction in sepsis as the primary research focus. Median AUROC among 1,263 publications with extractable values was 0.798.

CONCLUSION: Open access database-based critical care research is growing almost exponentially, driven primarily by Chinese groups using MIMIC-IV for retrospective mortality studies. European databases remain underutilized. The field exhibits methodological homogeneity and a heavy bias toward US-American clinical practice patterns, with cross-dataset external validation rarely foregrounded in study abstracts. Greater use of European databases, systematic external validation and topical diversification are key priorities.

Keywords

Humans, Intensive Care Units/organization & administration, Bibliometrics, Critical Care/methods, Databases, Factual/statistics & numerical data