Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Med Syst . 2026 Apr 13;50(1):49. doi: 10.1007/s10916-026-02376-3 Search in PMC Search in PubMed View in NLM Catalog Add to search Evaluating Artificial Intelligence for Sepsis Prediction in Emergency Departments: A Systematic Review and Meta Analysis Yinan Zhang Yinan Zhang 1 DHI Lab, Biomedical Informatics and Digital Health, Sydney School of Public Health, The University of Sydney, Westmead, NSW 2145 Australia Find articles by Yinan Zhang 1 , Tim Kirchler Tim Kirchler 1 DHI Lab, Biomedical Informatics and Digital Health, Sydney School of Public Health, The University of Sydney, Westmead, NSW 2145 Australia Find articles by Tim Kirchler 1 , Audrey P Wang Audrey P Wang 1 DHI Lab, Biomedical Informatics and Digital Health, Sydney School of Public Health, The University of Sydney, Westmead, NSW 2145 Australia Find articles by Audrey P Wang 1, ✉ Author information Article notes Copyright and License information 1 DHI Lab, Biomedical Informatics and Digital Health, Sydney School of Public Health, The University of Sydney, Westmead, NSW 2145 Australia ✉ Corresponding author. Received 2025 Sep 11; Accepted 2026 Mar 30; Issue date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . PMC Copyright notice PMCID: PMC13076368 PMID: 41973329 Abstract This study aims to synthesize current evidence on artificial intelligence-based sepsis prediction models for emergency department patients and propose practical benchmarks that emphasize standardized data preparation and reproducible model characterization. Literature searches were conducted across Scopus, Web of Science, PubMed, MEDLINE, and Embase. Eligible studies were selected through a two-tiered screening process, followed by data extraction and assessment according to predefined criteria. Random-effects meta-analysis was used to quantify model performance, and heterogeneity was explored by subgroup, regression, and sensitivity analyses. A total of 36 studies comprising 98 predictive models were included, with a pooled area under the receiver operating characteristic curve of 0.87 (95% CI: 0.86–0.88). Differences in performance were associated with study-level methodologies, including target definition, data provenance, cohort scale, data preprocessing, feature representation, and model development. The integrated meta-regression further identified independent methodologies influencing model performance. Artificial intelligence-based models showed higher pooled predictive performance than widely used traditional scoring systems for sepsis in emergency departments. However, translation into practice remains limited by inconsistent evaluation and reporting, and by inadequate external validation. Standardized methodological benchmarks have the potential to improve reproducibility, comparability, and clinical applicability. Supplementary Information The online version contains supplementary material available at 10.1007/s10916-026-02376-3. Keywords: Sepsis, Artificial intelligence, Prediction models, Emergency medical services, Clinical decision support system Introduction Sepsis is a life-threatening condition indiscriminately affecting individuals of all ages, defined by the Sepsis-3 criteria as organ dysfunction caused by a dysregulated host response to infection [ 1 ]. Data released in 2020 showed that approximately 48.9 million sepsis cases occurred worldwide, leading to 11 million deaths, estimated to represent around 20% of global mortality [ 2 ]. Clinical management of sepsis is time-sensitive, with mortality risk escalating by 4–7% for each delayed hour in antibiotic administration [ 3 , 4 ]. In emergency departments (EDs), early sepsis recognition is limited due to rapid turnover, resource constraints, high cognitive loads [ 5 , 6 ], and different patient presentations [ 7 – 9 ], risking delayed interventions and adverse outcomes. Recent studies suggest that artificial intelligence (AI) increasingly outperforms rule-based screening tools such as Systemic Inflammatory Response Syndrome (SIRS), quick Sequential Organ Failure Assessment (qSOFA), and Modified Early Warning Score (MEWS) [ 10 ]. Traditional approaches have several limitations in spite of widespread clinical adoption, including insufficient sensitivity and specificity due to oversimplified thresholds [ 11 ], delayed responsiveness stemming from intermittently manually entered data [ 12 ], and limited adaptability across diverse patient groups [ 13 ]. By comparison, AI technologies have the potential to overcome these difficulties through real-time analysis of clinical datasets to detect physiological trajectories indicative of early-stage sepsis [ 12 ]. For example, a multicentre study successfully predicted 80% of septic cases an average of 3.7 h before onset using international databases [ 14 ]; and another triage-level analysis with over one million participants leveraged natural language processing to achieve an area under the receiver operating characteristic curve (AUROC) of 0.94 [ 15 ]. Integration of such models facilitates earlier alerts and intensive patient monitoring, helping EDs to mitigate diagnostic errors and mortality risks [ 16 , 17 ]. However, methodological variability and inconsistent results across studies limit comparability and generalizability. It remains unclear under which conditions AI-based models offer robust advantages over conventional warning systems across different healthcare contexts and patient demographics. Some evidence indicates superior AI performance primarily under optimal conditions, including extensive, high-quality datasets and higher prevalence [ 18 , 19 ]. Conversely, others report limitations arising from algorithm biases, incomplete clinical annotations, and inadequate external validations [ 20 , 21 ]. Thus, systematically evaluating AI effectiveness in EDs and establishing methodological guidelines are necessary steps toward bridging these gaps. Existing reviews on sepsis prediction focus primarily on isolated technologies and rarely conduct comprehensive analyses of model development. Previous studies discussed diagnostic ‘gold standard’ definitions [ 14 , 22 – 25 ], data processing strategies [ 23 , 26 ], predictive features [ 22 , 24 – 31 ], algorithm selection [ 22 , 23 , 26 , 29 , 32 , 33 ], and evaluation indicators [ 29 , 31 , 34 ]. Nevertheless, the absence of standardized evaluation criteria and cohesive modeling benchmarks complicates comparative analysis, undermines clinical interpretability, and limits actual adoption. Similarly, overlooking clarity in data preparation pipelines and consistency in model development may lead to performance biases, yielding high accuracy in controlled experiments but reduced performance in unfamiliar scenarios. Therefore, this review aims to evaluate recent AI-based sepsis prediction models in EDs with a focus on data preparation and model characteristics (Fig. 1 ). Theoretically, it provides a methodological benchmark to promote standardized and reproducible modeling approaches. Practically, it offers evidence-based recommendations to develop AI-enhanced clinical decision support systems, ensuring feasibility and effective integration into emergency medical workflows. The review is guided by three specific research questions: Fig. 1. Open in a new tab Overall workflow for sepsis prediction models How effectively do AI-based methods predict sepsis in EDs compared with traditional assessment tools? How can credible, generalizable performance be maintained or enhanced through data preparation and model characterization? How can standardized evaluation and reporting support reproducibility, cross-study comparability, and readiness for clinical implementation? Methods This systematic review and meta-analysis adheres to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [ 35 ]. The protocol has been pre-registered with the Open Science Framework (OSF) (osf.io/jkbr5). Search Strategy On November 8, 2024, an exhaustive search was conducted across five databases: Scopus, Web of Science, PubMed, MEDLINE, and Embase. The search strategy encompassed four primary concepts: ‘Sepsis’, ‘Artificial Intelligence’, ‘Emergency Department’, and ‘Prediction’ (see Table S1 ). Search terms included subject headings (MeSH in PubMed and MEDLINE, Emtree in Embase), along with synonyms and alternatives. These terms were orchestrated using Boolean operators, and their order of operation was grouped with parentheses (e.g., ‘Sepsis AND (Diagnosis OR Prognosis)’). Field-specific tags guided the searches within designated fields, while exact phrases were constructed using double quotes (curly brackets { } instead in Scopus). An asterisk ‘*’ functioned as a wildcard for truncation searches, accommodating variations of the root word (e.g., ‘Predict*’ could cover ‘Predicts’, ‘Predicting’ or ‘Prediction’). Additionally, searches were constrained to journals published from 2019 onwards to obtain recent developments in AI and excluded review-type articles. Detailed query strings are provided in Supplementary File S1 . Study Screening and Eligibility Criteria Retrieved records were imported into EndNote 21.2 for duplication and managed using Covidence. Two researchers (YZ and TK) independently performed a two-tiered screening to minimize biases. Inclusion criteria comprised studies that: (1) were peer-reviewed journal articles to mitigate the unreviewed bias; (2) were published in English for accessibility; (3) were published from 2019 onwards to ensure the latest development of AI technology; (4) explicitly mentioned sepsis prediction in titles or abstracts; (5) applied AI algorithms (e.g., Machine Learning or Deep Learning), and highlighted in titles or abstracts; (6) were conducted in emergency department facilities instead of non-ED or outpatient settings [ 36 ]; (7) reported predictive model effectiveness measures. We excluded case reports and review articles, as well as studies focused on non-human subjects or experimental research unrelated to human healthcare. Following a pilot screening of 100 publications, titles and abstracts were reviewed against predefined eligibility criteria (Table S2 and S3 ) to exclude obviously irrelevant studies. Full-text evaluations were subsequently conducted on candidate articles to determine eligibility for inclusion. Any discrepancies between two reviewers were resolved by consensus, with arbitration by the third reviewer (AW) when necessary. Screening processes were well-documented and reported using a PRISMA flow diagram, clearly outlining rationales for exclusion during the full-text phase. Data Extraction A standardized data extraction form was designed and applied to all included studies, structured around three domains: study details (Table S4 ), data preparation (Table S5 ), and model characteristics (Table S6 ). The extraction process was fully documented to support transparency and reproducibility. Extracted information included: (1) study details: authors, year of publication, country, clinical setting, demographics, sepsis definition, clinical comparator, research objective, study period; (2) data preparation: data source, collection method, sample size, training/validation/testing datasets, positive patients, sepsis prevalence, data preprocessing methods, feature engineering, feature numbers (initial and final), feature types, feature importance; (3) model characteristics: algorithms, optimization methods, AUROCs, other output metrics, accuracy, sensitivity, specificity, prediction window, effectiveness summary. Quality Assessment The quality assessment was conducted based on a predefined TRIPOD + AI framework to evaluate the reporting completeness and transparency of included studies [ 37 ], which was adapted specifically for this review. Each study was rated ‘YES’, ‘NO’, or ‘Not Applicable (N/A)’ across 52 items organized into five sections: introduction, methods, open science, results, and discussion (details are shown in Table S7 ). Statistical Analysis Descriptive statistics summarized key characteristics of included studies in terms of research objectives, publication years, countries, patient demographics, modeling features and algorithm types. Between-study heterogeneity was evaluated using the statistic. Pooled performance metrics (AUROC, accuracy, sensitivity, and specificity) and their 95% confidence intervals were estimated using random-effects meta-analysis. Subgroup analysis was performed stratified by prediction outcomes, sepsis definitions, data sources, collection methods, sample sizes, sepsis prevalence, feature numbers, dataset divisions, and algorithm types. Univariate and multivariate meta-regressions were fitted to examine variables influencing sepsis model performance and sources of heterogeneity. An integrated model was further used to estimate associations between candidate variables and model performance, and to quantify heterogeneity explained. Its inference was based on study-clustered robust standard errors to account for dependence among multiple effect sizes reported in the same study, thus yielding more conservative uncertainty than conventional model-based tests. All regression models accounted for within-study clustering of multiple reported models by including study-level random effects. Besides, the leave-one-out sensitivity analysis systematically excluded individual studies to test result robustness, and Egger’s funnel plot was used for assessing publication biases. Considering that only AUROC was universally reported across studies, it was selected as the primary measure. Since AUROC values were bounded between 0 and 1, a logit transformation was applied prior to regression analyses to stabilize variance and satisfy statistical assumptions. For clarity, factors used for model development were defined as ‘predictors’ and those used for study-level regression analyses as ‘variables’. All analytical procedures were conducted using R software version 4.4.2. Results Study Selection The data screening process is summarized in a PRISMA flowchart (Fig. 2 ). Initially, 5,527 records were retrieved from five databases and imported into EndNote, with 3,259 duplicates subsequently removed. Covidence facilitated the screening of remaining records, excluding 2,087 irrelevant titles and abstracts, leaving 181 articles for full-text review. 30 reports were not retrieved in full text because they were conference abstracts/records only, even after attempts to locate related full publications by the same authors. Ultimately, 36 studies satisfied the inclusion criteria. The main reasons for exclusion were the absence of AI tools or non-ED settings. Fig. 2. Open in a new tab PRISMA flow diagram of the study screening Study Characteristics The included studies ( n = 36) were observational cohort studies primarily aimed at prediction of early-onset sepsis ( n = 19, 53%) [ 15 , 38 – 55 ], sepsis-related mortality ( n = 11, 31%) [ 43 , 56 – 66 ], septic shock ( n = 5, 14%) [ 67 – 71 ], and sepsis prognosis ( n = 1, 3%) [ 72 ]. The majority originated from the U.S. healthcare system ( n = 12, 33%) and focused on adult patient visits ( n = 29, 81%). Frequently employed AI algorithms included boosting models ( n = 19, 53%), logistic regression (LR) ( n = 17, 47%), random forest (RF) ( n = 16, 44%), and neural network (NN) ( n = 15, 42%). AUROC values were model-level results reported in included studies and were presented descriptively by prediction outcomes, shown with sample sizes, feature and model numbers, algorithm types, and prediction windows (Fig. 3 , adapted from Fleuren et al. [ 25 ], extracted from Table S4 – 6 ). AUROCs ranged from 0.65 to 0.98 for early-onset prediction, 0.83 to 0.92 for septic shock and prognosis, and 0.63 to 0.98 for mortality prediction. Prediction windows were generally within 24 hours for sepsis onset and shock, whereas they spanned up to 30 days for mortality. Sample sizes varied widely (< 500 to > 2 million), with models incorporating up to 91 features. Fig. 3. Open in a new tab Overview of included studies. AB= Adaptive Boosting , ANN= Artificial Neural Network , CR= Cox Regression , CART= Classification and Regression Tree , CatB= Categorical Boosting , CNN= Convolutional Neural Network, CoxPh= Cox Proportional Hazards , DT= Decision Tree , DNN= Deep Neural Network , FNN= Feedforward Neural Network , GBM= Gradient Boosting Machine , GNB= Gaussian Naive Bayes , KNN= K-Nearest Neighbors , LGBM= Light Gradient Boosting Machine , LSTM= Long Short-Term Memory , LASSO= Least Absolute Shrinkage and Selection Operator , MLP= Multi-layer Perceptron , MARS= Multivariate Adaptive Regression Splines , MGP= Multi-output Gaussian Process , NB= Naïve Bayes , PLR= Penalized Logistic Regression , RNN= Recurrent Neural Network , RR= Ridge Regression , SVM= Support Vector Machine , XGB= Extreme Gradient Boosting The 29 most frequently used predictive features were categorized as vital signs, demographics, and laboratory results (Fig. 4 ). The top five commonly reported features across studies (> 50%) were body temperature, respiratory rate, heart rate, systolic blood pressure, and age. Fig. 4. Open in a new tab Common features used for sepsis prediction models Quality of Reporting Across 36 included studies appraised with TRIPOD + AI (Fig. 5 ), basic descriptive and methodological items were well-reported; whereas major deficiencies remained in open-science practices and in the level of result completeness needed for external validation (full item-level ratings were provided in Table S8 ) Fig. 5. Open in a new tab Quality assessment results of included studies Title, Abstract, Introduction The assessment criteria for abstract, healthcare background, targeted population, and study objectives were fulfilled in all studies. In comparison, explicit identification of prediction model development in the title was not frequently presented ( n = 21, 58%), and health-inequality considerations across sociodemographic groups were rarely recognized ( n = 1, 3%). Methods Information on cohort definition and data provenance was reported in nearly all reports, including data sources, study dates and settings, eligibility criteria, outcome definitions, and preprocessing pipelines. Predictors were typically defined with missingness handling and data partitioning, and model development with performance measures and outputs was presented. Ethical approval was also commonly reported. By contrast, several safeguards and reproducibility enablers were inconsistently reported. Training-evaluation differences were described in around half of studies ( n = 16, 44%); details of treatments received that could influence sepsis prediction outcomes were reported in only 7 studies (19%); model updating, fairness assessment, sample-size justification and calculated prediction were seldom summarized (no more than 11%). Open Science and Patient/Public Involvement Open-science fidelity was limited. Protocols, registrations, and patient/public involvement were unavailable, and data/code availability was uncommon (each n = 4, 11%). However, funding and conflicts of interest were more often disclosed (each n = 29, 83%). Results Most studies shared similar trends in their resulting outputs, and participant flow and analytical cohort were clearly presented. The dataset characteristics ( n = 35, 97%) were frequently outlined, participant flows, predictor distributions, model development steps, and overall performances were reported in over 50%. As a result, while headline performance was often visible, transparency needed for recalculation, audit, or external validation remained limited, full model specification was rarely sufficient for independent reconstruction, and forward-looking plans for monitoring in use were almost never provided (each n = 1, 3%). Discussion The interpretation of findings was clarified across included studies ( n = 36, 100%), with limitations and future directions considered ( n = 33, 92%). However, practice-facing details needed for ED adoption were hardly addressed. Approaches for managing input-data quality at the point of use were rarely specified ( n = 2, 6%), and practical requirements for end-user interaction were almost never described ( n = 1, 3%). Therefore, operational readiness and guidance for safe deployment remain underdeveloped despite thoughtful reflections. Meta-Analyses Pooled Prediction Performance The random effects meta-analysis was used to integrate performance metrics from 98 AI models across 36 included studies given the observed high heterogeneity ( > 50%). The pooled AUROC for algorithmic sepsis prediction in EDs was 0.87 (95% CI: 0.86–0.88, = 99.94%). Corresponding pooled estimates of accuracy, sensitivity, and specificity were respectively 0.81 (95% CI: 0.79–0.83, = 99.92%), 0.77 (95% CI: 0.74–0.81, = 99.98%), and 0.85 (95% CI: 0.83–0.87, = 99.99%). By comparison, reported traditional rule-based methods (SIRS, SOFA, qSOFA, MEWS, NEWS, REMS, MEDS) showed lower pooled AUROCs (0.66–0.74), pooled accuracy (0.60–0.78), and pooled sensitivity (0.36–0.77) in 23 included studies (Fig. 6 and Table S9 ). Although the specificity of AI models was comparable with some traditional tools (e.g., qSOFA at 0.93, MEWS at 0.89, and NEWS at 0.84), it remained consistently high across evaluations. Pooled estimates for conventional scores were based on the subset of studies reporting each metric, and the AI-score comparisons should be interpreted with this in mind. Fig. 6. Open in a new tab Comparison of pooled performance between AI-based models and traditional rule-based scores. Multivariate Imputation by Chained Equations Subgroup Analysis In subgroup analyses (Fig. 7 ), higher pooled AUROC values were observed in models targeting early-onset sepsis (0.90; 95%CI: 0.88–0.92), using International Classification of Diseases (ICD) based definitions (0.93; 95% CI: 0.89–0.97), private medical databases (0.87; 95% CI: 0.86–0.89), prospective data collection (0.88; 95% CI: 0.85–0.90), larger samples (≥ 10,000) (0.88; 95% CI: 0.86–0.90) with higher prevalence (≥ 20%) (0.88; 95% CI: 0.86–0.90), fewer features (< 20) (0.89; 95% CI: 0.87–0.92), and smaller training proportions (< 70%) (0.90; 95% CI: 0.87–0.92). Among the algorithms covered, boosting-based models exhibited the best pooled AUROC (0.90; 95% CI: 0.87–0.93), followed by KNN (0.90; 95% CI: 0.83–0.98), ensemble methods (0.89; 95% CI: 0.84–0.94), and SVM (0.88; 95% CI: 0.83–0.93). Fig. 7. Open in a new tab Pooled performance measures in subgroups. Prediction outcome significantly moderated AUROC (QM(3)=21.18, p < 0.001; mainly mortality vs early-onset). The prognosis subgroup was based on a single study/model (k=1) and is presented descriptively. Accuracy was not reported by any septic shock prediction studies and is shown as not available Modeling Methodological Predictors 80 identified predictors are visualized in a circular plot (Fig. 8 ). Beyond descriptive co-occurrence and frequency counts, pairwise statistically significant positive interactions among predictors that appeared in at least five studies were harmonized to prespecified classifications (data collection, data preprocessing, feature engineering, and model characteristics). Fig. 8. Open in a new tab Circle plot of predictor classifications, frequency distributions, and significant positive interactions among high-frequency predictors in AI-based sepsis prediction models. Simple imputation: missing values were replaced using fixed or rule-based methods, such as mean/median substitution or carry-forward filling. Advanced imputation: missing values were estimated using model-based approaches, such as Multivariate Imputation by Chained Equations (MICE) or random forest–based imputation Explicit reporting of these predictors revealed model development frameworks and facilitated transparency and comparability across studies. As the foundation for subsequent modeling, specific aspects of data collection were subdivided into prediction outcomes, sepsis definitions, data sources, collection methods, sample sizes and prevalence; data preprocessing covered cleaning for outliers and missing values, data transformation, encoding, balancing and splitting; feature engineering involved feature selection, extraction, generation [ 30 ], counts and types; and model characteristics included algorithm selection, hyperparameter tuning, and a series of training and validation techniques. Interaction motivated the later regression work by showing recurring methodological combinations across studies. These co-occurrence patterns were used to reduce dimensionality and narrow down candidate predictors, rather than to support pairwise performance claims. Low-prevalence or imbalanced settings were repeatedly linked to strategies that strengthened minority-class signals, such as embedded feature selection and boosting-type learners. Common physiological markers (Fig. 4 ), including heart rate, blood pressure, and oxygen saturation, were frequently involved in positive interactions with feature selection, hyperparameter tuning, and boosting algorithms. Data scale and splitting design also interacted with feature-engineering decisions, with larger cohorts generally well-suited for embedded feature selection, while smaller feature groups often required higher training sets for full learning. Three-Step Regression Analysis Meta-regression was used to investigate potential sources of heterogeneity and to identify which study-level variables were associated with model predictiveness, while recognizing that these comparisons reflected methodological combinations rather than isolated causal effects. To ensure data comparability across studies, 40 of the 80 high-frequency predictors (Fig. 8 ) were selected for regression analyses. Considering the substantial heterogeneity, mixed-effects meta-regressions on the logit scale of AUROC were undertaken (Table 1 ). Table 1. Univariate and multivariate outcomes of meta-regressions Variables Univariate Model Multivariate Model Significant Variable Model β p 95% CI β p 95% CI β p 95% CI Prediction Outcome: Early-onset 0.269 0.701 [−1.102, 1.640] −0.102 0.893 [−1.591, 1.387] Prediction Outcome: Septic shock 0.066 0.930 [−1.413, 1.545] 0.198 0.582 [−0.508, 0.905] Prediction Outcome: Mortality 0.403 0.564 [−0.968, 1.774] 0.033 0.965 [−1.456, 1.522] Sepsis definition: Sepsis-3 based −0.140 0.521 [−0.567, 0.287] −0.677 0.031* [−1.293, −0.062] −0.042 0.839 [−0.463, 0.380] Sepsis definition: Infection + SIRS/SOFA −0.228 0.362 [−0.718, 0.262] −0.605 0.086 [−1.294, 0.085] Data Source: Private database −0.009 0.975 [−0.582, 0.564] −0.008 0.979 [−0.622, 0.606] Collection Method: Retrospective −0.041 0.883 [−0.582, 0.501] −0.532 0.114 [−1.191, 0.127] Sample Size ≥ 10000 0.355 0.092 [−0.057, 0.767] 0.820 0.004* [0.263, 1.378] 0.449 0.011 [0.116, 0.782] Sepsis Prevalence 0.285 0.670 [−1.027, 1.597] 1.642 0.051 [−0.006, 3.289] Data cleaning: Simple imputation 0.091 0.685 [−0.348, 0.530] −0.297 0.260 [−0.815, 0.220] Data cleaning: Advanced imputation −0.361 0.158 [−0.864, 0.141] −0.138 0.609 [−0.669, 0.392] Data cleaning: Deletion 0.387 < 0.001* [0.373, 0.401] 0.387 < 0.001* [0.373, 0.401] 0.409 0.003* [0.286, 0.531] Data Transformation: Normalization −0.524 0.034* [−1.010, −0.039] −0.428 0.100 [−0.938, 0.082] −0.234 0.393 [−0.841, 0.372] Data Transformation: Standardization 0.221 0.424 [−0.320, 0.762] 0.653 0.026* [0.079, 1.228] 0.318 0.199 [−0.201, 0.837] Data Balancing: Oversampling 0.086 0.772 [−0.494, 0.666] 0.497 0.080 [−0.060, 1.055] Data Splitting: Training Set ≥ 70% −1.641 0.125 [−3.739, 0.458] −0.738 0.015* [−1.331, −0.146] −0.284 0.371 [−0.950, 0.381] Feature Selection: Filter −0.421 0.141 [−0.982, 0.140] −0.651 0.035* [−1.257, −0.045] −0.182 0.381 [−0.639, 0.275] Feature Selection: Embedded −0.213 0.344 [−0.653, 0.228] −0.238 0.311 [−0.697, 0.222] Feature Extraction: Auto-extraction 0.061 < 0.001* [0.052, 0.070] 0.061 < 0.001* [0.052, 0.070] 0.172 0.522 [−1.186, 1.530] Feature Number −0.006 0.125 [−0.014, 0.002] −0.006 0.157 [−0.014, 0.002] Feature Type: Demographic −0.886 0.019* [−1.626, −0.146] −1.000 0.042* [−1.963, −0.036] −0.995 0.035* [−1.861, −0.128] Feature Type: Vital −0.529 0.421 [−1.818, 0.760] 0.862 0.263 [−0.646, 2.370] Feature Type: Lab 0.085 0.771 [−0.491, 0.662] 0.359 0.263 [−0.269, 0.986] Feature: Body Temperature 0.047 0.843 [−0.418, 0.512] 0.101 0.752 [−0.523, 0.724] Feature: Respiratory Rate −0.054 0.813 [−0.509, 0.400] 0.143 0.654 [−0.483, 0.770] Feature: Age −0.424 0.051 [−0.849, 0.002] −0.487 0.092 [−1.054, 0.080] Feature: Heart Rate −0.102 0.649 [−0.540, 0.337] −0.098 0.715 [−0.626, 0.429] Feature: Systolic Blood Pressure −0.074 0.743 [−0.514, 0.367] 0.107 0.784 [−0.656, 0.869] Feature: Oxygen Saturation −0.023 0.918 [−0.455, 0.410] 0.034 0.889 [−0.447, 0.516] Feature: Diastolic Blood Pressure −0.096 0.664 [−0.530, 0.338] −0.122 0.723 [−0.797, 0.553] Feature: White Blood Cell 0.271 0.216 [−0.159, 0.701] 0.182 0.464 [−0.305, 0.669] Algorithm: Logistic Regression 0.128 < 0.001* [0.118, 0.139] 0.112 < 0.001* [0.101, 0.123] 0.138 0.624 [−1.317, 1.593] Algorithm: Random Forest 0.204 < 0.001* [0.194, 0.215] 0.213 < 0.001* [0.203, 0.224] 0.261 0.379 [−0.928, 1.450] Algorithm: Support Vector Machines 0.335 < 0.001* [0.316, 0.355] 0.336 < 0.001* [0.317, 0.356] 0.369 0.147 [−0.283, 1.021] Algorithm: Boosting 0.180 < 0.001* [0.168, 0.192] 0.188 < 0.001* [0.177, 0.200] 0.208 0.430 [−0.990, 1.405] Algorithm: Neural Networks 0.261 < 0.001* [0.247, 0.274] 0.276 < 0.001* [0.262, 0.289] 0.045 0.523 [−0.327, 0.416] Regularization 0.191 0.491 [−0.352, 0.733] 0.226 0.409 [−0.310, 0.761] K-fold Cross-validation 0.265 0.257 [−0.193, 0.723] 0.247 0.300 [−0.221, 0.716] Hyperparameter Tuning: Grid search 0.073 < 0.001* [0.049, 0.097] 0.186 < 0.001* [0.160, 0.212] 0.163 0.131 [−0.178, 0.504] External Validation −0.166 0.526 [−0.678, 0.347] −0.026 0.922 [−0.545, 0.493] Open in a new tab Values are reported to three decimal places; * denotes statistically significant correlations with AUROC (p < 0.05). Coefficients are contrasts relative to the reference on the logit-AUROC scale; sign changes when the reference changes. Independent associations were first explored in univariate regressions. Higher AUROC was observed in studies reporting deletion for data cleaning, automated feature extraction, grid-search for hyperparameter tuning, and the use of mainstream algorithms (LR, RF, SVM, NN, and boosting). In contrast, normalization for data transformation and inclusion of demographic features were associated with lower AUROCs in unadjusted models. Multivariate meta-regressions were fitted within prespecified methodological blocks (data collection, data preprocessing, feature engineering with types, and model characteristics; see Fig. 8 ) to limit collinearity and reduce overfitting. Several univariate variables attenuated after mutual adjustment, while the overall direction of associations remained consistent. Higher AUROC was basically associated with larger sample size (≥ 10,000), deletion-based and standardization-based data processing, feature auto-extraction, and grid-search tuning. Sepsis-3 definition, filter-based feature selection, demographic inclusion, and larger training allocation (≥ 70%) were linked to lower performance. An integrated model was then used to estimate the joint effects of significant variables identified from univariate and multivariate screens. The integrated model showed an improvement in fit (AICc reduced by 5,827.71) and a reduction in between-study variance ( from 0.38 to 0.25; pseudo- = 33.83%) compared to the baseline random-effects model, indicating that these study-level variables explained a meaningful proportion of between-study heterogeneity. Statistically supported associations were retained for sample size, deletion-based data cleaning, and use of demographic features. Although algorithm families exhibited significant positive associations in the earlier two-step regressions, their contrasts under the integrated model remained directionally positive but were estimated with uncertainty. Publication Bias and Sensitivity Analysis The result of Egger’s test indicated no significant publication bias in this research ( = 0.04, P -value = 0.96 > 0.05), confirmed by a symmetrical funnel plot (Fig S1 ). Sensitivity analysis demonstrated stable pooled estimates, almost unaffected by individual studies, affirming the robustness of our findings. Discussion Principal Findings This systematic review and meta-analysis provides a comprehensive evaluation of recent AI-based sepsis prediction studies in emergency contexts, including 36 studies with 98 models from the past five years. Consistent with the prior research [ 11 , 29 , 30 , 73 ], this work suggests that AI-assisted models achieved higher performance than widely-adopted traditional clinical screening tools. Specifically, the pooled AUROC for AI models (0.87) exceeded that of traditional methods (0.66–0.74). Conventional rule-based scores are typically developed for broad assessment of infection-related illness rather than for explicit sepsis prediction [ 74 ]. By comparison, AI-based models can incorporate large-scale, multi-source, and high-dimensional clinical data, which may allow them to capture subtle and dynamic physiological changes that threshold-based criteria do not fully detect [ 12 ]. However, the significant heterogeneity observed in this review, consistent with previous meta-analyses [ 11 , 25 , 34 ], necessitating deeper consideration of underlying methodological predictors, as well as a cautious explanation of pooled results. Sepsis-prediction model development involves complex processes, and predictors for data preparation and model characterization undeniably affect performance to varying degrees. Methodological Interpretation Compared with existing reviews that mainly summarized predictive performance [ 32 , 75 ], this study expands the scope through subgroup and regression analyses to identify methodological sources of heterogeneity and inform both model development and implementation. Some methodological choices may still be useful, but their independent contribution becomes harder to isolate when accounting for cohort scale and data preprocessing variables. Meanwhile, the negative associations also have interpretive value in that a lower AUROC does not necessarily imply poorer methodological quality [ 52 , 66 ]. Outcome Definition and Prediction Task Subgroup analyses suggested higher pooled discrimination under ICD-based labels, whereas regression analyses showed a negative association for Sepsis-3-based definitions. These findings indicate that sepsis definitions are interpreted as different prediction targets rather than interchangeable labels. ICD-based definitions are more established in routine administrative records, corresponding more closely to later clinician-confirmed states. In comparison, Sepsis-3 is closer to earlier physiologic identification, and lower performance under it reflects a stricter prediction task, which cannot be directly explained as evidence of inferior modeling. Cohort and Data Context At the cohort and data level, higher performance tended to be reported in studies using private databases, prospective collection, larger samples, and higher prevalence. Regression findings further examined that sample size remained one of the most reproducible independent variables in the integrated model. Larger sample sizes potentially reduce estimation instability and better represent heterogeneous ED presentations. These characteristics are more consistent with differences in data fidelity, cohort scale, and class balance than with any single optimal design rule [ 23 , 73 ]. Data Preprocessing and Feature Representation Findings related to data preprocessing and feature representation indicated that model performance depended not only on how much information was available, but also on how that information was cleaned and structured. Subgroups suggested that models using fewer clinical features could perform comparably to those using more high-dimensional sets [ 30 , 43 ]. In regressions, higher AUROCs were associated with deletion-based cleaning, standardized transformation, and automated feature extraction, while larger training allocation, demographic inclusion, and filter-based selection were linked to lower performance. The advantage of smaller training-to-validation ratios can be understood as the presence of broader validation cohorts to produce more reliable estimates. It is not a universally preferable split or an inherent benefit of allocating less data to model development [ 76 , 77 ]. Similarly, the inverse correlation for filter-based feature selection may involve missing interaction-driven or time-dependent information common in ED sepsis progression [ 30 ]. For preprocessing-related variables, standardization showed a stronger association after adjustments, which is reasonable when features measured on different scales are jointly analyzed [ 78 ]. In contrast, the earlier univariate pattern for normalization weakened in the block-wise multivariate model, implying that it is largely related to the modeling context in which normalization is applied than to an adverse effect of normalization itself. In the integrated model, deletion-based cleaning and demographic features remained statistically supported. The positive signal for deletion may reflect the efficiency and simplicity compared to inappropriate imputation methods. Simultaneously, consistent with previous findings, when dealing with larger sample sizes, lower baseline missing rates, and simpler cohort structures, deletion methods can ensure high-quality training data free from artificial value interference. The negative relationship for demographics does not represent that such variables are clinically unimportant. It reveals that static demographics add limited short-term discrimination once dynamic physiological information (e.g., vital signs and laboratory data) is encoded, while still remaining relevant for patient calibration, subgroup reporting, and fairness assessment [ 43 , 48 , 52 , 68 ]. Model-Building Choices The independent contribution of methods used in model construction was less stable than that of cohort and preprocessing variables. Subgroup analyses examined relatively strong pooled AUROC for boosting-based models, and early regression phases also showed positive correlations for several commonly used algorithms and grid-search tuning. Nevertheless, these contrasts attenuated in the integrated model, suggesting that part of their advantages may stem from context-dependent design, data, and feature-engineering decisions. Methodological Synthesis These findings indicate that heterogeneity in ED sepsis prediction studies is shaped by interactions among target definitions, cohort structures, preprocessing strategies, feature representations, and model-development choices. Methodological benchmarks should therefore prioritize clear specification of prediction targets, index time, prediction windows; representative and temporally consistent cohorts [ 42 , 61 , 66 ]; transparent handling of missingness, outliers, and feature scaling [ 77 , 79 , 80 ]; and systematic tunings that enhance model fitting and validation strategies that preserve adequate evaluation scopes. The considerations are reinforced by TRIPOD + AI appraisal, which clarified why many studies lacked details required for independent reconstruction and clinical translation, including open-science practices, reproducible model specification, operating thresholds, calibration outputs, and external/temporal validation. Consequently, methodological benchmarks should be framed as a pathway to improve cross-study comparability, as well as a prerequisite for reconstruction-level standardization and deployment-oriented evaluation. From Theory to Practice: Implementation Benchmarks Standardization and Clinical Impacts The implication of establishing benchmarks extends beyond model development to the credible evaluation and effective deployment of sepsis prediction tools in routine emergency care. Despite demonstrating obvious numerical strengths in predictive accuracy, limited external validations and actual implementations revealed in TRIPOD + AI highlight major challenges to adoption within emergency medical services. It is widely acknowledged in current studies that establishing standardized methods is a foundational step towards transparency [ 30 , 74 , 81 ]. Furthermore, this also encourages emergency professionals to evaluate and incorporate AI-based systems into clinical workflows with clearer operational expectations [ 5 – 9 , 80 ]. Standardized methodologies may also improve clinical accountability through systematic monitoring of model outcomes [ 82 ], including the effectiveness of implementing clinical interventions based on predictive results. Clear standards allow tracking of unintended consequences, such as alert fatigue or inefficient resource use [ 20 , 83 , 84 ]. Explicit reporting requirements can further support regulatory oversight and compliance with healthcare quality standards, ensuring AI-driven clinical decision support systems uphold patient safety and ethical responsibility in practice. Implementation Strategies The analysis supports an implementation-focused approach instead of treating model performance as the ultimate goal. One of the most critical practical barriers is to translate model outputs into reproducible clinical pathways. Reliability, interpretability and actionability should be emphasized during deployment. Sepsis prediction models should be embedded at decision points with prespecified use cases, as reported performance is highly sensitive to target definitions and evaluation designs. It is important to state the prediction window ranges, select operational thresholds that align with local system capacity, and provide calibration results so that risk scores have consistent clinical meaning [ 76 , 85 ]. Sustainable benefits also depend on whether model performance remains stable under routine ED data constraints. Because the more reproducible performance differences in this review were linked to cohort scale and data processing, implementation should prioritize reliable input streams based on routinely available dynamic signals, together with transparent rules for missing, abnormal, or delayed measurements. Ongoing monitoring should assess whether predicted risks continue to match observed event rates and whether alert burden is acceptable for ED workflow [ 44 ]. In addition, responses to performance drifts should be defined in advance and managed through recalibration and reassessment, which has opportunities to avoid repeated algorithm modifications. In theory, future work needs to explore why strategically simplified models may at time outperform more complex approaches that blindly increase feature sizes or training scales. Direct comparisons between parsimonious and high-dimensional models are also required to identify whether the advantages of simpler designs represent reduced overfitting, enhanced calibration, or effective transportability across settings. Research should focus on temporal effects, feature interactions, and target definitions influencing model behaviors, particularly across different sepsis criteria, workflow stages, and patient subgroups. Moreover, fairness and stability should be evaluated more systematically to determine if predictive performance is maintained across clinically and demographically diverse populations [ 25 , 50 ]. Future Directions In practice, further studies should plan prospective external validations and multicentre implementation-oriented research to validate whether performance can be sustained outside of single-site retrospective datasets. Evaluation should move beyond AUROC measures to include calibration accuracy, threshold-based clinical operability, workflow integration, resource use [ 20 , 30 ], and patient-centred outcomes. To improve real-world readiness, models should be developed under ethical and governance requirements that specify intended use, support transparent reporting, and enable post-deployment monitoring [ 86 ]. Greater emphasis on reproducibility, transportability, and clinically embedded evaluation will be important for strengthening AI-based sepsis prediction from promising technical performance to a more accountable emergency care application. Limitations This review inevitably has several limitations. The heterogeneity across included studies remained due to differences ranging from data collection to modeling phases, despite attempts to dissect the algorithm development behind the outstanding performance. Evidence certainty assessment based on prediction outcome categories using GRADEpro GDT shows overall low certainty, with a very low level for prognostic assessment due to methodological limitations and heterogeneity (Table S10). As a result, the pooled estimates should be interpreted with caution. Incomplete reporting of accuracy, sensitivity and specificity reduces the precision of secondary pooled analyses, whereas the representativeness and appropriateness of AUROC as a primary result metric remain controversial. Conclusion This systematic review and meta-analysis indicates that AI-based models achieve higher pooled predictive performance for sepsis compared with traditional screening systems in emergency departments. However, substantial methodological heterogeneity and limited external validation constrain generalizability. The standardized methodological and reporting benchmarks have the potential to improve cross-study reproducibility, comparability, and readiness towards clinical implementation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary File 1 (DOCX 206 KB) (206.1KB, docx) Acknowledgements The authors thank A/Prof. Amith Shetty and Dr. Mostafa Shaikh for their inputs into this project. Author Contributions Yinan Zhang (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Visualization, Writing - Original Draft & Editing); Tim Kirchler (Data curation, Investigation, Writing - Review); Audrey P Wang (Conceptualization, Methodology, Writing - Review, Supervision, Project Administration). Funding Open Access funding enabled and organized by CAUL and its Member Institutions. Audrey P Wang acknowledges funding support for her Westmead Fellowship from Research Education Network, Western Sydney Local Health District. Data Availability No datasets were generated or analysed during the current study. Declarations Ethics and Consent to Participate Declarations Not applicable. Clinical Trial Number Not applicable. Competing Interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Singer M, Deutschman CS, Seymour CW, et al (2016) The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA 315:801. 10.1001/jama.2016.0287 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 2. Rudd KE, Johnson SC, Agesa KM, et al (2020) Global, regional, and national sepsis incidence and mortality, 1990–2017: analysis for the Global Burden of Disease Study. The Lancet 395:200–211. 10.1016/S0140-6736(19)32989-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 3. Goh KH, Wang L, Yeow AYK, et al (2021) Artificial intelligence in sepsis early prediction and diagnosis using unstructured data in healthcare. Nat Commun 12:711. 10.1038/s41467-021-20910-4 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 4. Lam JY, Boussina A, Shashikumar SP, et al (2024) The impact of laboratory data missingness on sepsis diagnosis timeliness. JAMIA Open 7:ooae085. 10.1093/jamiaopen/ooae085 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Rhee C, Kadri SS, Danner RL, et al (2016) Diagnosing sepsis is subjective and highly variable: a survey of intensivists using case vignettes. Crit Care 20:89. 10.1186/s13054-016-1266-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 6. Seymour CW, Gesten F, Prescott HC, et al (2017) Time to Treatment and Mortality during Mandated Emergency Care for Sepsis. N Engl J Med 376:2235–2244. 10.1056/NEJMoa1703058 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Kareemi H, Vaillancourt C, Rosenberg H, et al (2021) Machine Learning Versus Usual Care for Diagnostic and Prognostic Prediction in the Emergency Department: A Systematic Review. Acad Emerg Med 28:184–196. 10.1111/acem.14190 [ DOI ] [ PubMed ] [ Google Scholar ] 8. Morr M, Lukasz A, Rübig E, et al (2016) Sepsis recognition in the emergency department – impact on quality of care and outcome? BMC Emerg Med 17:11. 10.1186/s12873-017-0122-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Husabø G, Nilsen RM, Flaatten H, et al (2020) Early diagnosis of sepsis in emergency departments, time to treatment, and association with mortality: An observational study. PLOS ONE 15:e0227652. 10.1371/journal.pone.0227652 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Schinkel M, Paranjape K, Nannan Panday RS, et al (2019) Clinical applications of artificial intelligence in sepsis: A narrative review. Comput Biol Med 115:103488. 10.1016/j.compbiomed.2019.103488 [ DOI ] [ PubMed ] [ Google Scholar ] 11. Yadgarov MY, Landoni G, Berikashvili LB, et al (2024) Early detection of sepsis using machine learning algorithms: a systematic review and network meta-analysis. Front Med 11:1491358. 10.3389/fmed.2024.1491358 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 12. Tran A, Topp R, Tarshizi E, Shao A (2023) Predicting the Onset of Sepsis Using Vital Signs Data: A Machine Learning Approach. Clin Nurs Res 32:1000–1009. 10.1177/10547738231183207 [ DOI ] [ PubMed ] [ Google Scholar ] 13. Sun B, Lei M, Wang L, et al (2025) Prediction of sepsis among patients with major trauma using artificial intelligence: a multicenter validated cohort study. Int J Surg 111:467–480. 10.1097/JS9.0000000000001866 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Moor M, Bennett N, Plečko D, et al (2023) Predicting sepsis using deep learning across international sites: a retrospective development and validation study. eClinicalMedicine 62:102124. 10.1016/j.eclinm.2023.102124 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Brann F, Sterling NW, Frisch SO, Schrager JD (2024) Sepsis Prediction at Emergency Department Triage Using Natural Language Processing: Retrospective Cohort Study. JMIR AI 3:e49784. 10.2196/49784 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Giacobbe DR, Signori A, Del Puente F, et al (2021) Early Detection of Sepsis With Machine Learning Techniques: A Brief Clinical Perspective. Front Med 8:617486. 10.3389/fmed.2021.617486 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Kausch SL, Moorman JR, Lake DE, Keim-Malpass J (2021) Physiological machine learning models for prediction of sepsis in hospitalized adults: An integrative review. Intensive Crit Care Nurs 65:103035. 10.1016/j.iccn.2021.103035 [ DOI ] [ PubMed ] [ Google Scholar ] 18. Koçak B, Cuocolo R, Santos DPD, et al (2023) Must-have Qualities of Clinical Research on Artificial Intelligence and Machine Learning. Balk Med J 40:3–12. 10.4274/balkanmedj.galenos.2022.2022-11-51 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 19. Park SH, Han K (2018) Methodologic Guide for Evaluating Clinical Performance and Effect of Artificial Intelligence Technology for Medical Diagnosis and Prediction. Radiology 286:800–809. 10.1148/radiol.2017171920 [ DOI ] [ PubMed ] [ Google Scholar ] 20. Cross JL, Choma MA, Onofrey JA (2024) Bias in medical AI: Implications for clinical decision-making. PLOS Digit Health 3:e0000651. 10.1371/journal.pdig.0000651 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 21. Ueda D, Kakinuma T, Fujita S, et al (2024) Fairness of artificial intelligence in healthcare: review and recommendations. Jpn J Radiol 42:3–15. 10.1007/s11604-023-01474-3 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 22. Teng AK, Wilcox AB (2020) A Review of Predictive Analytics Solutions for Sepsis Patients. Appl Clin Inform 11:387–398. 10.1055/s-0040-1710525 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 23. Deng H-F, Sun M-W, Wang Y, et al (2022) Evaluating machine learning models for sepsis prediction: A systematic review of methodologies. iScience 25:103651. 10.1016/j.isci.2021.103651 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 24. Yan MY, Gustad LT, Nytrø Ø (2022) Sepsis prediction, early detection, and identification using clinical text for machine learning: a systematic review. J Am Med Inform Assoc 29:559–575. 10.1093/jamia/ocab236 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 25. Fleuren LM, Klausch TLT, Zwager CL, et al (2020) Machine learning for the prediction of sepsis: a systematic review and meta-analysis of diagnostic test accuracy. Intensive Care Med 46:383–400. 10.1007/s00134-019-05872-y [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 26. Agnello L, Vidali M, Padoan A, et al (2024) Machine learning algorithms in sepsis. Clin Chim Acta 553:117738. 10.1016/j.cca.2023.117738 [ DOI ] [ PubMed ] [ Google Scholar ] 27. Hassan N, Slight R, Weiand D, et al (2021) Preventing sepsis; how can artificial intelligence inform the clinical decision-making process? A systematic review. Int J Med Inf 150:104457. 10.1016/j.ijmedinf.2021.104457 [ DOI ] [ PubMed ] [ Google Scholar ] 28. Desai MD, Tootooni MS, Bobay KL (2022) Can Prehospital Data Improve Early Identification of Sepsis in Emergency Department? An Integrative Review of Machine Learning Approaches. Appl Clin Inform 13:189–202. 10.1055/s-0042-1742369 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 29. Zhang Y, Xu W, Yang P, Zhang A (2023) Machine learning for the prediction of sepsis-related death: a systematic review and meta-analysis. BMC Med Inform Decis Mak 23:283. 10.1186/s12911-023-02383-1 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 30. Bomrah S, Uddin M, Upadhyay U, et al (2024) A scoping review of machine learning for sepsis prediction- feature engineering strategies and model performance: a step towards explainability. Crit Care 28:180. 10.1186/s13054-024-04948-6 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 31. Islam KR, Prithula J, Kumar J, et al (2023) Machine Learning-Based Early Prediction of Sepsis Using Electronic Health Records: A Systematic Review. J Clin Med 12:5658. 10.3390/jcm12175658 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 32. Gao Y, Wang C, Shen J, et al (2024) Systematic review and network meta-analysis of machine learning algorithms in sepsis prediction. Expert Syst Appl 245:122982. 10.1016/j.eswa.2023.122982 [ Google Scholar ] 33. G. A, K.L. N, M.S. AS (2024) Improving sepsis classification performance with artificial intelligence algorithms: A comprehensive overview of healthcare applications. J Crit Care 83:154815. 10.1016/j.jcrc.2024.154815 [ DOI ] [ PubMed ] [ Google Scholar ] 34. Islam MdM, Nasrin T, Walther BA, et al (2019) Prediction of sepsis patients using machine learning approach: A meta-analysis. Comput Methods Programs Biomed 170:1–9. 10.1016/j.cmpb.2018.12.027 [ DOI ] [ PubMed ] [ Google Scholar ] 35. Page MJ, McKenzie JE, Bossuyt PM, et al (2021) The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 372:. 10.1136/bmj.n71 [ DOI ] [ PMC free article ] [ PubMed ] 36. Morgan SR, Chang AM, Alqatari M, Pines JM (2013) Non–Emergency Department Interventions to Reduce ED Utilization: A Systematic Review. Acad Emerg Med 20:969–985. 10.1111/acem.12219 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 37. Collins GS, Moons KGM, Dhiman P, et al (2024) TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 385: e078378. 10.1136/bmj-2023-078378 [ DOI ] [ PMC free article ] [ PubMed ] 38. Delahanty RJ, Alvarez J, Flynn LM, et al (2019) Development and Evaluation of a Machine Learning Model for the Early Identification of Patients at Risk for Sepsis. Ann Emerg Med 73:334–344. 10.1016/j.annemergmed.2018.11.036 [ DOI ] [ PubMed ] [ Google Scholar ] 39. Mohamed A, Ying H, Sherwin R (2020) Electronic-Medical-Record-Based Identification of Sepsis Patients in Emergency Department: A Machine Learning Perspective. In: 2020 International Conference on Contemporary Computing and Applications (IC3A). pp 336–340 40. Bedoya AD, Futoma J, Clement ME, et al (2020) Machine learning for early detection of sepsis: an internal and temporal validation study. JAMIA Open 3:252–260. 10.1093/jamiaopen/ooaa006 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 41. Zhang D, Yin C, Hunold KM, et al (2021) An interpretable deep-learning model for early prediction of sepsis in the emergency department. Patterns 2:100196. 10.1016/j.patter.2020.100196 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 42. Taneja I, Damhorst GL, Lopez-Espina C, et al (2021) Diagnostic and prognostic capabilities of a biomarker and EMR-based machine learning algorithm for sepsis. Clin Transl Sci 14:1578–1589. 10.1111/cts.13030 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 43. Ehwerhemuepha L, Heyming T, Marano R, et al (2021) Development and validation of an early warning tool for sepsis and decompensation in children during emergency department triage. Sci Rep 11:8578. 10.1038/s41598-021-87595-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 44. Shashikumar SP, Wardi G, Malhotra A, Nemati S (2021) Artificial intelligence sepsis prediction algorithm learns to say “I don’t know.” Npj Digit Med 4:134. 10.1038/s41746-021-00504-6 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 45. Lin P-C, Chen K-T, Chen H-C, et al (2021) Machine Learning Model to Identify Sepsis Patients in the Emergency Department: Algorithm Development and Validation. J Pers Med 11:. 10.3390/jpm11111055 [ DOI ] [ PMC free article ] [ PubMed ] 46. Kijpaisalratana N, Sanglertsinlapachai D, Techaratsami S, et al (2022) Machine learning algorithms for early sepsis detection in the emergency department: A retrospective study. Int J Med Inf 160:104689. 10.1016/j.ijmedinf.2022.104689 [ DOI ] [ PubMed ] [ Google Scholar ] 47. Aguirre U, Urrechaga E (2023) Diagnostic performance of machine learning models using cell population data for the detection of sepsis: a comparative study. Clin Chem Lab Med CCLM 61:356–365. 10.1515/cclm-2022-0713 [ DOI ] [ PubMed ] [ Google Scholar ] 48. Mercurio L, Pou S, Duffy S, Eickhoff C (2023) Risk Factors for Pediatric Sepsis in the Emergency Department: A Machine Learning Pilot Study. Pediatr Emerg Care 39:e48–e56. 10.1097/PEC.0000000000002893 [ DOI ] [ PubMed ] [ Google Scholar ] 49. Guo S, Guo Z, Ren Q, et al (2023) A Prediction Model for Sepsis in Infected Patients: Early Assessment of Sepsis Engagement. Shock 60(2) :214–220. 10.1097/SHK.0000000000002170 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 50. Prasad V, Aydemir B, Kehoe IE, et al (2023) Diagnostic suspicion bias and machine learning: Breaking the awareness deadlock for sepsis detection. PLOS Digit Health 2:e0000365. 10.1371/journal.pdig.0000365 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 51. Aygun U, Yagin FH, Yagin B, et al (2024) Assessment of Sepsis Risk at Admission to the Emergency Department: Clinical Interpretable Prediction Model. Diagnostics 14:457. 10.3390/diagnostics14050457 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 52. Xia Y, Xu L, Lai Q, et al (2024) Construction and validation of machine learning models based on bedside parameters for identifying sepsis in acute pancreatitis patients. Signa Vitae 20:60–68 [ Google Scholar ] 53. Hou Y-T, Wu M-Y, Chen Y-L, et al (2024) Efficacy of a sepsis clinical decision support system in identifying patients with sepsis in the emergency department. Shock 62:480–487. 10.1097/SHK.0000000000002394 [ DOI ] [ PubMed ] [ Google Scholar ] 54. Xie J, Gao J, Yang M, et al (2024) Prediction of sepsis within 24 hours at the triage stage in emergency departments using machine learning. World J Emerg Med 15:379. 10.5847/wjem.j.1920-8642.2024.074 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 55. Song Y, Huang H, Ma J, et al (2025) Early prediction of sepsis in emergency department patients using various methods and scoring systems. Nurs Crit Care 30(3):e13201. 10.1111/nicc.13201 [ DOI ] [ PubMed ] 56. Perng J-W, Kao I-H, Kung C-T, et al (2019) Mortality Prediction of Septic Patients in the Emergency Department Based on Machine Learning. J Clin Med 8:1906. 10.3390/jcm8111906 [ DOI ] [ PMC free article ] [ PubMed ] 57. Zhao C, Wei Y, Chen D, et al (2020) Prognostic value of an inflammatory biomarker-based clinical algorithm in septic patients in the emergency department: An observational study. Int Immunopharmacol 80:106145. 10.1016/j.intimp.2019.106145 [ DOI ] [ PubMed ] [ Google Scholar ] 58. Kwon YS, Baek MS (2020) Development and Validation of a Quick Sepsis-Related Organ Failure Assessment-Based Machine-Learning Model for Mortality Prediction in Patients with Suspected Infection in the Emergency Department. J Clin Med 9:875. 10.3390/jcm9030875 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 59. Van Doorn WPTM, Stassen PM, Borggreve HF, et al (2021) A comparison of machine learning models versus clinical evaluation for mortality prediction in patients with sepsis. PLoS One 16:e0245157. 10.1371/journal.pone.0245157 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 60. Karlsson A, Stassen W, Loutfi A, et al (2021) Predicting mortality among septic patients presenting to the emergency department–a cross sectional analysis using machine learning. BMC Emerg Med 21:84. 10.1186/s12873-021-00475-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 61. Chao H-Y, Wu C-C, Singh A, et al (2022) Using Machine Learning to Develop and Validate an In-Hospital Mortality Prediction Model for Patients with Suspected Sepsis. Biomedicines 10:802. 10.3390/biomedicines10040802 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 62. Cheng C-Y, Kung C-T, Chen F-C, et al (2022) Machine learning models for predicting in-hospital mortality in patient with sepsis: Analysis of vital sign dynamics. Front Med 9:964667. 10.3389/fmed.2022.964667 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 63. Greco M, Caruso PF, Spano S, et al (2023) Machine Learning for Early Outcome Prediction in Septic Patients in the Emergency Department. Algorithms 16:76. 10.3390/a16020076 [ Google Scholar ] 64. Jeon E-T, Song J, Park DW, et al (2023) Mortality prediction of patients with sepsis in the emergency department using machine learning models: a retrospective cohort study according to the Sepsis-3 definitions. Signa Vitae Vol.19:pp.112–124. 10.22514/sv.2023.046 [ Google Scholar ] 65. Wong BPK, Lam RPK, Ip CYT, et al (2023) Applying artificial neural network in predicting sepsis mortality in the emergency department based on clinical features and complete blood count parameters. Sci Rep 13:21463. 10.1038/s41598-023-48797-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 66. Park SW, Yeo NY, Kang S, et al (2024) Early Prediction of Mortality for Septic Patients Visiting Emergency Room Based on Explainable Machine Learning: A Real-World Multicenter Study. J Korean Med Sci 39:e53. 10.3346/jkms.2024.39.e53 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 67. Kim J, Chang H, Kim D, et al (2020) Machine learning for prediction of septic shock at initial triage in emergency department. J Crit Care 55:163–170. 10.1016/j.jcrc.2019.09.024 [ DOI ] [ PubMed ] [ Google Scholar ] 68. Scott HF, Colborn KL, Sevick CJ, Bajaj L, Deakyne Davies SJ, Fairclough D, Kissoon N, Kempe A (2021) Development and Validation of a Model to Predict Pediatric Septic Shock Using Data Known 2 Hours After Hospital Arrival. Pediatr Crit Care Med 22(1):16–26. 10.1097/PCC.0000000000002589 [ DOI ] [ PMC free article ] [ PubMed ] 69. Wardi G, Carlile M, Holder A, et al (2021) Predicting Progression to Septic Shock in the Emergency Department Using an Externally Generalizable Machine-Learning Algorithm. Ann Emerg Med 77:395–406. 10.1016/j.annemergmed.2020.11.007 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 70. Yun H, Park JH, Choi DH, et al (2021) Enhancement in Performance of Septic Shock Prediction Using National Early Warning Score, Initial Triage Information, and Machine Learning Analysis. J Emerg Med 61:1–11. 10.1016/j.jemermed.2021.01.038 [ DOI ] [ PubMed ] [ Google Scholar ] 71. Choi A, Chung K, Chung SP, et al (2022) Advantage of Vital Sign Monitoring Using a Wireless Wearable Device for Predicting Septic Shock in Febrile Patients in the Emergency Department: A Machine Learning-Based Analysis. Sensors 22:7054. 10.3390/s22187054 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 72. Chiu I-M, Chuang Y-P, Cheng C-Y, Lin C-HR (2022) Development and Validation of an Explainable Deep Learning Model to Predict Adverse Event During Hospital Admission in Patients with Sepsis. In: 2022 IEEE/ACIS 23rd International Conference on Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing (SNPD). IEEE, Taichung, Taiwan, pp 8–13 73. Chua MT, Boon Y, Lee ZY, et al (2025) The role of artificial intelligence in sepsis in the Emergency Department: a narrative review. Ann Transl Med 13:4–4. 10.21037/atm-24-150 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 74. Liu Z, Shu W, Li T, et al (2025) Interpretable machine learning for predicting sepsis risk in emergency triage patients. Sci Rep 15:887. 10.1038/s41598-025-85121-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 75. Yang Z, Cui X, Song Z (2023) Predicting sepsis onset in ICU using machine learning models: a systematic review and meta-analysis. BMC Infect Dis 23:635. 10.1186/s12879-023-08614-0 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 76. Stenwig E, Salvi G, Rossi PS, Skjærvold NK (2022) Comparative analysis of explainable machine learning prediction models for hospital mortality. BMC Med Res Methodol 22:53. 10.1186/s12874-022-01540-w [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 77. Feng J, Phillips RV, Malenica I, et al (2022) Clinical artificial intelligence quality improvement: towards continual monitoring and updating of AI algorithms in healthcare. Npj Digit Med 5:66. 10.1038/s41746-022-00611-y [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 78. Rashidi HH, Pantanowitz J, Hanna MG, et al (2025) Introduction to Artificial Intelligence and Machine Learning in Pathology and Medicine: Generative and Nongenerative Artificial Intelligence Basics. Mod Pathol 38:100688. 10.1016/j.modpat.2024.100688 [ DOI ] [ PubMed ] [ Google Scholar ] 79. Shashikumar SP, Stanley MD, Sadiq I, et al (2017) Early sepsis detection in critical care patients using multiscale blood pressure and heart rate dynamics. J Electrocardiol 50:739–743. 10.1016/j.jelectrocard.2017.08.013 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 80. Reddy S (2024) Generative AI in healthcare: an implementation science informed translational path on application, integration and governance. Implement Sci 19:27. 10.1186/s13012-024-01357-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 81. Stylianides C, Nicolaou A, Sulaiman WA, et al (2025) AI Advances in ICU with an Emphasis on Sepsis Prediction: An Overview. Mach Learn Knowl Extr 7:6. 10.3390/make7010006 [ Google Scholar ] 82. Díaz-Rodríguez N, Del Ser J, Coeckelbergh M, et al (2023) Connecting the dots in trustworthy Artificial Intelligence: From AI principles, ethics, and key requirements to responsible AI systems and regulation. Inf Fusion 99:101896. 10.1016/j.inffus.2023.101896 [ Google Scholar ] 83. Ackermann K, Baker J, Green M, et al (2022) Computerized Clinical Decision Support Systems for the Early Detection of Sepsis Among Adult Inpatients: Scoping Review. J Med Internet Res 24:e31083. 10.2196/31083 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 84. Tyler S, Olis M, Aust N, et al (2024) Use of Artificial Intelligence in Triage in Hospital Emergency Departments: A Scoping Review. Cureus. 10.7759/cureus.59906 [ DOI ] [ PMC free article ] [ PubMed ] 85. Valik JK, Ward L, Tanushi H, et al (2023) Predicting sepsis onset using a machine learned causal probabilistic network algorithm based on electronic health records data. Sci Rep 13:11760. 10.1038/s41598-023-38858-4 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 86. Williamson SM, Prybutok V (2024) Balancing Privacy and Progress: A Review of Privacy Challenges, Systemic Oversight, and Patient Perceptions in AI-Driven Healthcare. Appl Sci 14:675. 10.3390/app14020675 [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary File 1 (DOCX 206 KB) (206.1KB, docx) Data Availability Statement No datasets were generated or analysed during the current study. Articles from Journal of Medical Systems are provided here courtesy of Springer ACTIONS View on publisher site PDF (3.8 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top