European Journal of Cardiovascular Medicine

Sepsis, defined by the Sepsis-3 consensus as life-threatening organ dysfunction resulting from a dysregulated host response to infection, remains a major global public health challenge [1]. Contemporary estimates indicate that sepsis affects approximately 48 to 50 million individuals annually and contributes to nearly 11 million deaths worldwide, accounting for almost one fifth of global mortality [2,3]. The burden is disproportionately higher in low- and middle-income countries, where delayed recognition and limited healthcare resources exacerbate morbidity and mortality [2]. Children under five years of age represent a particularly vulnerable population, contributing substantially to global sepsis incidence [3]. In high income settings, sepsis also imposes a considerable economic burden. In the United States alone, it is among the most expensive causes of hospitalization, with annual healthcare expenditures exceeding USD 20 billion and average per patient costs surpassing USD 30,000 [2,3]. These data underscore the substantial clinical and economic impact of sepsis across healthcare systems.

Timely recognition and early initiation of appropriate therapy are critical determinants of survival in sepsis, as each hour of delay in treatment is associated with increased mortality [4]. However, early diagnosis remains challenging due to nonspecific early clinical features and heterogeneous disease trajectories [4]. Conventional clinical scoring systems, including the Systemic Inflammatory Response Syndrome, Modified Early Warning Score, Sequential Organ Failure Assessment, and quick Sequential Organ Failure Assessment, are widely used for risk stratification but exhibit important limitations [1,5]. The Systemic Inflammatory Response Syndrome demonstrates high sensitivity but poor specificity, whereas the quick Sequential Organ Failure Assessment favors specificity at the expense of sensitivity, particularly in early disease stages [5]. As a result, sepsis is frequently identified only after established organ dysfunction, limiting opportunities for early intervention. Moreover, traditional biomarkers and imaging modalities have shown limited accuracy for early sepsis prediction [6].

In recent years, artificial intelligence and machine learning approaches have been increasingly investigated as tools for early sepsis detection. By leveraging high dimensional electronic health record data, including vital signs, laboratory values, demographic variables, and temporal trends, machine learning models aim to identify subtle physiological patterns preceding clinical recognition of sepsis [7]. A wide range of algorithms has been explored, including tree-based models, neural networks, recurrent and convolutional architectures, and ensemble methods. Individual studies have reported promising predictive performance, often demonstrating earlier detection compared with conventional clinical scoring systems [7,8].

Despite the rapid growth of this literature, substantial variability exists across studies with respect to patient populations, data sources, model architectures, prediction horizons, outcome definitions, and validation strategies. Concerns regarding generalizability, interpretability, risk of bias, and lack of prospective validation persist [8]. Furthermore, reported performance metrics vary widely, making it difficult to draw definitive conclusions regarding the overall diagnostic accuracy and clinical utility of artificial intelligence-based sepsis prediction models. To date, no consensus has been reached regarding which modeling approaches offer the most robust and generalizable performance.

Accordingly, a systematic review and meta-analysis is warranted to quantitatively synthesize the available evidence on artificial intelligence driven models for early sepsis detection. The objectives of this review are to systematically evaluate the performance of machine learning based sepsis prediction models, pool diagnostic accuracy metrics where feasible, assess between study heterogeneity, and critically appraise study quality and risk of bias. By providing a comprehensive and quantitative synthesis of existing evidence, this review aims to clarify the current state of artificial intelligence-based sepsis prediction and identify key gaps to inform future model development, validation, and clinical implementation.

Literature Screening and Study Selection The literature screening and study selection process was conducted in accordance with PRISMA guidelines [9]. Following database searches, all retrieved records were imported into a reference management software and duplicates were removed. Titles and abstracts were independently screened by two reviewers to assess eligibility based on predefined inclusion and exclusion criteria. Full text articles were subsequently reviewed for studies meeting the criteria, including use of artificial intelligence or machine learning models for early sepsis prediction in human populations. Disagreements were resolved through discussion or consultation with a third reviewer. The final set of studies was included for qualitative synthesis and quantitative meta-analysis. Registered with the Prospero with number CRD420251276030. Data Extraction and Statistical Analysis Data extraction was independently performed by two reviewers using a standardized and pre piloted extraction form. Extracted variables included study design, setting, population characteristics, sample size, sepsis definition, type of artificial intelligence or machine learning model, input features, prediction horizon, validation strategy, and reported performance metrics such as area under the receiver operating characteristic curve, sensitivity, specificity, and accuracy. Discrepancies were resolved by consensus. For meta-analysis, pooled estimates of diagnostic accuracy were calculated using random effects models. Between study heterogeneity was assessed using the I² statistic, and subgroup analyses were performed where feasible. Risk of Bias Assessment Risk of bias was independently assessed by two reviewers using the Risk Of Bias In Non randomized Studies of Interventions tool (Robins-I) [10]. The following domains were evaluated: bias due to confounding, participant selection, classification of interventions, deviations from intended interventions, missing data, outcome measurement, and selection of the reported result. Disagreements were resolved by consensus or third reviewer adjudication, and overall risk of bias was classified as low, moderate, serious, or critical.

Literature Screening and Study Selection

The literature screening and study selection process was conducted in accordance with PRISMA guidelines [9]. Following database searches, all retrieved records were imported into a reference management software and duplicates were removed. Titles and abstracts were independently screened by two reviewers to assess eligibility based on predefined inclusion and exclusion criteria. Full text articles were subsequently reviewed for studies meeting the criteria, including use of artificial intelligence or machine learning models for early sepsis prediction in human populations. Disagreements were resolved through discussion or consultation with a third reviewer. The final set of studies was included for qualitative synthesis and quantitative meta-analysis. Registered with the Prospero with number CRD420251276030.

Data Extraction and Statistical Analysis

Data extraction was independently performed by two reviewers using a standardized and pre piloted extraction form. Extracted variables included study design, setting, population characteristics, sample size, sepsis definition, type of artificial intelligence or machine learning model, input features, prediction horizon, validation strategy, and reported performance metrics such as area under the receiver operating characteristic curve, sensitivity, specificity, and accuracy. Discrepancies were resolved by consensus. For meta-analysis, pooled estimates of diagnostic accuracy were calculated using random effects models. Between study heterogeneity was assessed using the I² statistic, and subgroup analyses were performed where feasible.

Risk of Bias Assessment

Risk of bias was independently assessed by two reviewers using the Risk Of Bias In Non randomized Studies of Interventions tool (Robins-I) [10]. The following domains were evaluated: bias due to confounding, participant selection, classification of interventions, deviations from intended interventions, missing data, outcome measurement, and selection of the reported result. Disagreements were resolved by consensus or third reviewer adjudication, and overall risk of bias was classified as low, moderate, serious, or critical.

In this systematic review and meta-analysis, we comprehensively evaluated the diagnostic performance of machine learning (ML)–based early warning systems for the early prediction of sepsis in hospitalized adult populations. Across 30 included studies encompassing diverse clinical settings and ML architectures, these systems demonstrated high pooled sensitivity, specificity, discriminative ability, and overall accuracy, supporting their potential role as clinically valuable tools for early sepsis detection.

Principal Findings

The pooled sensitivity of ML-based models was high, indicating a strong ability to correctly identify patients who subsequently developed sepsis. This finding is clinically important, as delayed recognition remains a major contributor to sepsis-related morbidity and mortality. Prior studies have consistently shown that each hour of delay in sepsis recognition and treatment is associated with worse outcomes, underscoring the need for sensitive early warning systems [4,21,24]. The observed sensitivity values in our analysis exceed those typically reported for traditional rule-based tools such as qSOFA or SIRS, which have been shown to trade sensitivity for simplicity [4,26].

Similarly, pooled specificity was robust, suggesting that ML-based systems are capable of correctly identifying non-sepsis cases and thereby limiting unnecessary alerts. This is particularly relevant in real-world settings where alarm fatigue remains a major barrier to adoption of clinical decision support systems [8,26]. Several studies included in this meta-analysis explicitly incorporated interpretability or uncertainty estimation frameworks, enabling models to abstain from predictions when confidence was low, which may further improve clinician trust and specificity in practice [8,14].

The pooled AUC of 0.93 reflects excellent overall discrimination, highlighting the ability of ML models to distinguish between patients who will and will not develop sepsis across a wide range of thresholds. This level of discrimination compares favorably with previously reported AUCs for conventional scoring systems, which often range between 0.65 and 0.75 in heterogeneous hospital populations [4,24]. Deep learning architectures, particularly recurrent and temporal convolutional networks, appeared to contribute substantially to this performance by capturing longitudinal physiological trends rather than relying on static snapshots of clinical data [1,9,11,16].

Accuracy, Precision, and Clinical Trade-offs

The pooled diagnostic accuracy further reinforces the consistency of ML-based models in correctly classifying patients. However, precision (positive predictive value) showed the greatest variability across studies, with substantial heterogeneity observed. This finding is expected, as precision is highly sensitive to sepsis prevalence, alert thresholds, and case mix within individual cohorts. Studies conducted in high-acuity ICU populations generally reported higher precision compared with those in emergency department or mixed inpatient settings [3,10,22]. This variability emphasizes that while ML models may generalize well in terms of discrimination, local calibration remains essential to optimize alert burden and clinical utility [13,26].

The pooled F1 score, integrating sensitivity and precision, demonstrated balanced performance overall, supporting the notion that ML-based early warning systems can achieve a favorable trade-off between early detection and false-positive reduction. This balance is critical for real-world implementation, where overly sensitive systems may overwhelm clinicians, whereas overly specific systems risk missing early sepsis cases.

Heterogeneity and Sources of Variability

Across all major effect sizes, substantial heterogeneity was observed. This heterogeneity likely reflects differences in study design, patient populations, sepsis definitions, prediction horizons, data sources, and ML methodologies. Earlier studies frequently relied on Sepsis-2 or ICD-based definitions, whereas more recent investigations predominantly adopted Sepsis-3 criteria, which emphasize organ dysfunction and may alter case labeling [4,14,24]. Additionally, models varied in their use of structured EHR data alone versus multimodal inputs incorporating laboratory trends, waveform data, or unstructured clinical notes [1,12,18].

Importantly, despite high heterogeneity, pooled estimates remained stable and statistically significant across all analyses, suggesting that the overall direction and magnitude of effect are robust. Nonetheless, these findings support the need for subgroup analyses and prospective validation to better understand context-specific performance.

Clinical Implications

Collectively, the findings of this meta-analysis indicate that ML-based early warning systems outperform traditional clinical scoring tools in terms of sensitivity and discrimination, while maintaining acceptable specificity and accuracy. These systems hold promise for enabling earlier recognition of sepsis, facilitating timely initiation of antibiotics and supportive care. However, successful clinical translation will depend on workflow integration, clinician engagement, explainability, and continuous performance monitoring [8,26,30].

Limitations and Future Directions

Several limitations warrant consideration. Most included studies were retrospective and relied on EHR-derived data, raising concerns regarding residual confounding, data quality, and label noise. External validation was inconsistently performed, and few studies assessed downstream clinical outcomes following model implementation. Future research should prioritize prospective, multicenter evaluations, standardized reporting of performance metrics, and assessment of patient-centered outcomes, including mortality and resource utilization [30].

This systematic review and meta-analysis demonstrate that machine learning–based early warning systems provide high diagnostic performance for early sepsis prediction, with consistently strong sensitivity, specificity, discriminative ability, and overall accuracy across diverse clinical settings. Compared with conventional clinical scoring systems, these models show superior capacity for early risk stratification while maintaining acceptable false-positive rates. Despite substantial between-study heterogeneity, the robustness of pooled estimates supports the potential clinical value of machine learning–driven approaches. Future efforts should focus on prospective validation, local calibration, and seamless clinical integration to optimize real-world implementation and improve patient outcomes in sepsis care. Conflict of Interest The authors certify that there is no conflict of interest with any financial organization regarding the material discussed in the manuscript. Funding The authors report no involvement in the research by the sponsor that could have influenced the outcome of this work. Authors’ contributions. All authors contributed equally to the manuscript and read and approved the final version of the manuscript. Acknowledgement This paper is the collaborative work of all authors under the mentorship for the research work from BIR (Biomedical and International Research). We all authors acknowledge this mentorship for this meta-analysis.

[1]           Singer M, Deutschman CS, Seymour CW, Shankar-Hari M, Annane D, Bauer M, et al. The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA. 2016;315(8):801-810. doi:10.1001/jama.2016.0287

[2]           Rudd KE, Johnson SC, Agesa KM, Shackelford KA, Tsoi D, Kievlan DR, et al. Global, regional, and national sepsis incidence and mortality, 1990–2017: analysis for the Global Burden of Disease Study. Lancet. 2020;395(10219):200-211. doi:10.1016/S0140-6736(19)32989-7

[3]           Fleischmann C, Scherag A, Adhikari NKJ, Hartog CS, Tsaganos T, Schlattmann P, et al. Assessment of global incidence and mortality of hospital-treated sepsis. Am J Respir Crit Care Med. 2016;193(3):259-272. doi:10.1164/rccm.201504-0781OC

[4]           Kumar A, Roberts D, Wood KE, Light B, Parrillo JE, Sharma S, et al. Duration of hypotension before initiation of effective antimicrobial therapy is the critical determinant of survival in human septic shock. Crit Care Med. 2006;34(6):1589-1596. doi:10.1097/01.CCM.0000217961.75225.E9

[5]           Seymour CW, Liu VX, Iwashyna TJ, Brunkhorst FM, Rea TD, Scherag A, et al. Assessment of clinical criteria for sepsis: for the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA. 2016;315(8):762-774. doi:10.1001/jama.2016.0288

[6]           Pierrakos C, Vincent JL. Sepsis biomarkers: a review. Crit Care. 2010;14(1):R15. doi:10.1186/cc8872

[7]           Fleuren LM, Klausch TLT, Zwager CL, Schoonmade LJ, Guo T, Roggeveen LF, et al. Machine learning for the prediction of sepsis: a systematic review and meta-analysis of diagnostic test accuracy. Intensive Care Med. 2020;46(3):383-400. doi:10.1007/s00134-019-05872-y

[8]           Sendak MP, Ratliff W, Sarro D, Alderton E, Futoma J, Gao M, et al. Real-world integration of a sepsis deep learning technology into routine clinical care: implementation study. JMIR Med Inform. 2020;8(7):e15182. doi:10.2196/15182

[9]           Lauritsen SM, Kalør ME, Kongsgaard EL, et al. Early detection of sepsis utilizing deep learning on electronic health record event sequences. Artif Intell Med. 2020;104:101820. doi:10.1016/j.artmed.2020.101820

[10]         Wang D, Li J, Sun Y, et al. A machine learning model for accurate prediction of sepsis in ICU patients. Front Public Health. 2021;9:754348. doi:10.3389/fpubh.2021.754348

[11]         Kijpaisalratana N, Sanglertsinlapachai D, Techaratsami S, et al. Machine learning algorithms for early sepsis detection in the emergency department: a retrospective study. Int J Med Inform. 2022;160:104689. doi:10.1016/j.ijmedinf.2022.104689

[12]         Nemati S, Holder A, Razmi F, et al. An interpretable machine learning model for accurate prediction of sepsis in the ICU. Crit Care Med. 2018;46(4):547–553. doi:10.1097/CCM.0000000000002936

[13]         Duan Y, Huo J, Chen M, et al. Early prediction of sepsis using double fusion of deep features and handcrafted features. Appl Intell (Dordr). 2023;53:1–17. doi:10.1007/s10489-023-04621-9

[14]         Rosnati M, Fortuin V. MGP-AttTCN: an interpretable machine learning model for the prediction of sepsis. PLoS One. 2021;16(4):e0251248. doi:10.1371/journal.pone.0251248

[15]         Zhang D, Yin C, Hunold KM, et al. An interpretable deep-learning model for early prediction of sepsis in the emergency department. Patterns (N Y). 2021;2(2):100196. doi:10.1016/j.patter.2021.100196

[16]         Shashikumar SP, Wardi G, Malhotra A, et al. Artificial intelligence sepsis prediction algorithm learns to say “I don’t know.” NPJ Digit Med. 2021;4(1):134. doi:10.1038/s41746-021-00504-0

[17]         Aşuroğlu T, Oğul H. A deep learning approach for sepsis monitoring via severity score estimation. Comput Methods Programs Biomed. 2021;198:105816. doi:10.1016/j.cmpb.2020.105816

[18]         Oei SP, van Sloun RJG, van der Ven M, et al. Towards early sepsis detection from measurements at the general ward through deep learning. Intell Based Med. 2021;5:100042. doi:10.1016/j.ibmed.2021.100042

[19]         Rafiei A, Rezaee A, Hajati F, et al. SSP: early prediction of sepsis using fully connected LSTM-CNN model. Comput Biol Med. 2021;128:104110. doi:10.1016/j.compbiomed.2020.104110

[20]         Goh KH, Wang L, Yeow AYK, et al. Artificial intelligence in sepsis early prediction and diagnosis using unstructured data in healthcare. Nat Commun. 2021;12(1):711. doi:10.1038/s41467-021-20910-4

[21]         Bedoya AD, Futoma J, Clement ME, et al. Machine learning for early detection of sepsis: an internal and temporal validation study. JAMIA Open. 2020;3(2):252–260. doi:10.1093/jamiaopen/ooaa017

[22]         Yang M, Liu C, Wang X, et al. An explainable artificial intelligence predictor for early detection of sepsis. Crit Care Med. 2020;48(11):e1091–e1096. doi:10.1097/CCM.0000000000004544

[23]         Yuan KC, Tsai LW, Lee KH, et al. The development of an artificial intelligence algorithm for early sepsis diagnosis in the intensive care unit. Int J Med Inform. 2020;141:104176. doi:10.1016/j.ijmedinf.2020.104176

[24]         Kok C, Jahmunah V, Oh SL, et al. Automated prediction of sepsis using temporal convolutional network. Comput Biol Med. 2020;127:103957. doi:10.1016/j.compbiomed.2020.103957

[25]         Reyna MA, Josef CS, Jeter R, et al. Early prediction of sepsis from clinical data: the PhysioNet/Computing in Cardiology Challenge 2019. Crit Care Med. 2020;48(2):210–217. doi:10.1097/CCM.0000000000004145

[26]         Ibrahim ZM, Wu H, Hamoud A, et al. On classifying sepsis heterogeneity in the ICU: insight using machine learning. J Am Med Inform Assoc. 2020;27(3):437–445. doi:10.1093/jamia/ocz205

[27]         Fagerström J, Bång M, Wilhelms D, et al. LiSep LSTM: a machine learning algorithm for early detection of septic shock. Sci Rep. 2019;9(1):15132. doi:10.1038/s41598-019-51639-6

[28]         Kaji DA, Zech JR, Kim JS, et al. An attention-based deep learning model of clinical events in the intensive care unit. PLoS One. 2019;14(2):e0211057. doi:10.1371/journal.pone.0211057

[29]         Giannini HM, Ginestra JC, Chivers C, et al. A machine learning algorithm to predict severe sepsis and septic shock. Crit Care Med. 2019;47(11):1485–1492. doi:10.1097/CCM.0000000000003891

[30]         Barton C, Chettipally U, Zhou Y, et al. Evaluation of a machine learning algorithm for up to 48-hour advance prediction of sepsis. Comput Biol Med. 2019;109:79–84. doi:10.1016/j.compbiomed.2019.04.006

[31]         Delahanty RJ, Alvarez J, Flynn LM, et al. Development and evaluation of a machine learning model for early identification of patients at risk for sepsis. Ann Emerg Med. 2019;73(4):334–344. doi:10.1016/j.annemergmed.2018.11.036

[32]         Mao Q, Jay M, Hoffman JL, et al. Multicentre validation of a sepsis prediction algorithm using only vital sign data. BMJ Open. 2018;8:e017833. doi:10.1136/bmjopen-2017-017833

[33]         Taneja I, Damhorst GL, Lopez-Espina C, et al. Diagnostic and prognostic capabilities of a biomarker and EMR-based machine learning algorithm for sepsis. Clin Transl Sci. 2021;14(4):1578–1589. doi:10.1111/cts.13021

[34]         Horng S, Sontag DA, Halpern Y, et al. Creating an automated trigger for sepsis clinical decision support at emergency department triage. PLoS One. 2017;12(4):e0174708. doi:10.1371/journal.pone.0174708

[35]         Calvert JS, Price DA, Chettipally UK, et al. A computational approach to early sepsis detection. Comput Biol Med. 2016;74:69–73. doi:10.1016/j.compbiomed.2016.05.003

[36]         Desautels T, Calvert J, Hoffman J, et al. Prediction of sepsis in the ICU with minimal electronic health record data. JMIR Med Inform. 2016;4(3):e28. doi:10.2196/medinform.5909

[37]         Brown SM, Jones J, Kuttler KG, et al. Prospective evaluation of an automated method to identify patients with severe sepsis or septic shock. BMC Emerg Med. 2016;16:31. doi:10.1186/s12873-016-0092-7

[38]         Adams R, Henry KE, Sridharan A, et al. Prospective, multisite study of patient outcomes after implementation of a machine learning-based early warning system for sepsis. Nat Med. 2022;28(7):1455–1460. doi:10.1038/s41591-022-01836-8

[39]         Henry KE, Hager DN, Pronovost PJ, Saria S. A targeted real-time early warning score (TREWScore) for septic shock. Sci Transl Med. 2015;7(299):299ra122. doi:10.1126/scitranslmed.aab3719

[40]         Fleuren LM, Klausch TLT, Zwager CL, et al. Machine learning for the prediction of sepsis: a systematic review and meta-analysis of diagnostic test accuracy. Intensive Care Med. 2020;46(3):383–400. doi:10.1007/s00134-019-05872-y