Sepsis machine learning research surges, bibliometric map reveals hotspots

Researchers have produced the first bibliometric analysis dedicated specifically to machine learning prediction models for sepsis, mapping more than three decades of publications to reveal how the field has grown from a trickle of studies into a global research enterprise. The analysis, published in Critical Care: Sepsis and Severe Infection, examined 857 publications indexed in the Web of Science Core Collection between 1990 and September 18, 2024, and traced the trajectory of a field that now sits at the intersection of intensive care medicine and artificial intelligence.

The team, led by Xintong Wang and colleagues including Lei Qi, Qin Liu, Feng Ji and Xiaobo Hu, searched the Science Citation Index Expanded and the Social Sciences Citation Index using a comprehensive string of terms covering artificial intelligence, deep learning, neural networks, random forests, big data and related methods, combined with prediction terms and sepsis terminology. The initial search returned 1,060 records. After excluding review articles, editorials, letters, meeting abstracts and other non-article document types, and restricting the set to English-language original research articles, 857 publications remained for analysis.

The scale of the resulting landscape is striking. The 857 articles were written by 5,557 researchers from 3,772 institutions across 65 countries and appeared in 333 different journals, accumulating 24,156 citations. According to the authors, this breadth of participation reflects a highly globalized and interdisciplinary research area with active academic collaboration, and the frequent cross-citation among publications points to the interconnected nature of work combining critical care medicine with computational methods.

The annual publication trend tells a story of three distinct phases. From 1990 to 2005, the field remained in an early development stage, with most years producing between zero and three publications, and some years producing none. After 2005, output entered a slow growth phase, gradually increasing year by year. Then, from 2020 onward, the field entered what the authors describe as an explosive expansion. Publication output reached a historical peak of 200 articles in 2023, roughly 4.1 times the 49 articles published in 2020. In 2024, the count fell to 151, though the authors note this figure covers only part of the year, since data collection stopped on September 18, and the total still remained at a relatively high level. Linear fitting analysis confirmed an overall positive growth trend.

Journal analysis showed that the research is distributed across critical care medicine, medical artificial intelligence, bioinformatics and related fields. PLOS ONE and Scientific Reports published the largest number of articles in the field, while Critical Care, with an impact factor of 8.8, and Critical Care Medicine, with an impact factor of 7.7, ranked among the top in both impact factor and total citation frequency. Interdisciplinary journals such as Computers in Biology and Medicine and Artificial Intelligence in Medicine also showed strong publication performance. In a journal co-occurrence network built with VOSviewer, including journals with at least three occurrences, the highest total link strength belonged to Critical Care Medicine with 247, Scientific Reports with 187 and Computers in Biology and Medicine with 147. A separate bibliographic coupling analysis found Scientific Reports at the top with a total link strength of 18,019, followed by Frontiers in Medicine with 13,817 and BMC Medical Informatics and Decision Making with 11,898, suggesting these journals draw on a similar knowledge base.

Individual researchers also emerged as central figures. Shamim Nemati held the highest h-index in the field at 9, reflecting a combination of consistent productivity and citation impact. Ritankar Das had the highest total citation count at 709, while Rishikesan Kamaleswaran had the highest publication output with 20 papers. The authors also highlighted Zhongheng Zhang’s continuous research output as evidence of the growing contribution of Chinese researchers to intelligent sepsis prediction models.

The most cited publication in the dataset was not a sepsis model at all but a 2014 Health Affairs paper titled “Big data in health care: using analytics to identify and manage high-risk and high-cost patients,” which accumulated 621 citations. That study proposed using big data analytics to mine electronic health data and identify high-risk patients and disease deterioration, providing methodological foundations for early prediction models of critical illness, including sepsis. The second most cited work, with 407 citations, was the 2018 Critical Care Medicine paper “An interpretable machine learning model for accurate prediction of sepsis in the ICU,” which used real-time ICU data to build an interpretable early warning model capable of predicting sepsis onset 4 to 12 hours in advance. Third, with 326 citations, was the 2017 Lancet Respiratory Medicine study “Classification of patients with sepsis according to blood genomic endotype: a prospective cohort study,” which combined genomics and machine learning to sort sepsis patients into four molecular endotypes.

Keyword analysis offered a window into the field’s research hotspots. Using author keywords and Keywords Plus, the team identified 146 keywords with at least five occurrences, relying primarily on Keywords Plus results because they proved more stable. In the co-occurrence network, “sepsis” dominated with 251 occurrences and a total link strength of 861, followed by “mortality” with 175 occurrences and a link strength of 696, “septic shock” with 131 occurrences and 478, and “prediction” with 89 occurrences and 339. These high-frequency terms point to the field’s core concerns: sepsis prediction, mortality prediction, risk assessment, disease severity assessment and predictive model development.

Burst detection in CiteSpace revealed which topics have recently captured research attention. “Procalcitonin” showed the strongest citation burst, with a strength of 3.91 between 2020 and 2021. The authors urge caution in interpreting this, however, noting that procalcitonin is a widely investigated biomarker in sepsis research generally, so its prominence may reflect biomarker-oriented studies rather than machine learning prediction work specifically, and cannot simply position the biomarker as a core topic of the field. Other keywords showing notable bursts between 2020 and 2024 included “system,” “classification,” “mortality prediction,” “electronic health record,” “intensive care,” “risk factors” and “health care,” suggesting growing interest in risk stratification, predictive modeling, healthcare data utilization and critical care applications. The authors stress that citation bursts reflect changes in research attention over time and should not be read as direct evidence of clinical importance or future impact.

Beyond mapping trends, the analysis confronts an uncomfortable gap: clinical translation lags well behind algorithm development. Many machine learning models have shown strong predictive performance in retrospective studies and public databases, but implementation-focused research has highlighted persistent barriers to adoption, including limited external validation, poor generalizability across healthcare systems, difficulties integrating with electronic health records and the absence of prospective real-world evaluation. Even the highly cited Artificial Intelligence Sepsis Expert model, which achieved accurate prediction of sepsis 4 to 12 hours before clinical recognition, was accompanied by its authors’ emphasis that prospective studies are still needed to determine its clinical utility. Variations in patient populations, sepsis definitions and data collection practices across databases may also produce dataset shift, limiting how well models generalize when moved to different clinical environments. And even highly accurate models may have limited clinical impact if they are not integrated into existing workflows, since alert fatigue, workflow disruption and insufficient user acceptance can erode the practical value of predictive algorithms despite strong technical performance.

Based on these findings, the authors identify priorities for future work. They argue that research should emphasize clinical applicability rather than purely chasing predictive performance, developing transparent and clinically actionable prediction systems while addressing interpretability, external validation and prospective evaluation. The rising prominence of keywords related to electronic health records, validation and healthcare systems suggests a need to improve model generalizability through data harmonization, interoperability and multicenter validation. They also call for standardized methodological and reporting frameworks, noting that differences in data preprocessing, outcome definitions, model development strategies and validation procedures currently hinder meaningful comparisons across studies.

The study has limitations the authors acknowledge. It drew only on the Web of Science Core Collection, potentially missing relevant publications in other databases. Only English-language publications were included, which may limit regional representativeness. The choice of search terms and retrieval criteria may have shaped the final dataset and omitted some relevant studies. And because 2024 data were collected only through September 18, the analysis of annual publication trends for that year may be affected. The authors also note that a preliminary PubMed search using the terms sepsis, machine learning, prediction model and bibliometrics identified only three potentially relevant publications and no prior bibliometric study focused specifically on machine learning-based sepsis prediction models, supporting the novelty of their analysis.

For a field that barely existed before 2005 and has exploded since 2020, the bibliometric map offers both encouragement and a warning. The integration of artificial intelligence and critical care medicine is clearly accelerating, with a stable academic system forming around leading critical care and interdisciplinary AI journals. But the authors conclude that future progress will depend not only on improvements in predictive accuracy, but also on solving the harder problems of validation, interpretability, interoperability, workflow integration and prospective clinical evaluation that stand between promising algorithms and real benefit for sepsis patients.

Subject of Research: Bibliometric analysis of machine learning prediction models for sepsis

Article Title: Global trends and hotspots of the sepsis machine learning prediction model: a bibliometric analysis

Article References: Wang, X., Qi, L., Liu, Q., Ji, F., & Hu, X. (2026). Global trends and hotspots of the sepsis machine learning prediction model: a bibliometric analysis. Critical Care: Sepsis and Severe Infection, 1(1), Article 14. https://doi.org/10.1186/s44541-026-00017-3

Image Credits: AI Generated

DOI: 10.1186/s44541-026-00017-3

Keywords: sepsis, machine learning, prediction models, bibliometrics, critical care, Web of Science, research trends, procalcitonin, electronic health records, clinical translation, early warning, artificial intelligence