Accuracy of prognostic scoring tools for ICU triage and mortality among patients in low- and middle-income countries: a systematic review and meta-analysis Authors: Thong Huu Tran, Hien Thi Thu Nguyen, Trung Huu Tran Published: 2026 ABSTRACT AbstractBackgroundEarly warning and severity scores are widely used to support triage and predict clinical deterioration. Their diagnostic performance in low- and middle-income countries (LMICs) however remains uncertain due to differences in patient populations and healthcare resources compared with high-income settings.ObjectivesThis study aimed to evaluate and compare the diagnostic accuracy of commonly used scoring systems for predicting intensive care unit (ICU) admission and in-hospital mortality among adults presenting to emergency departments in LMICs.MethodsWe conducted a systematic review and meta-analysis of PubMed, Science Direct, Web of Science and the Cochrane Library. Observational studies assessing the accuracy of early warning or severity scores for ICU admission or in-hospital mortality of patient admitted to the emergency department were included. Data were pooled using a bivariate random-effects model to estimate sensitivity, specificity and area under the curve (AUC). Study quality was appraised using QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies 2).ResultsTwenty-six studies comprising over 45 000 patients from LMICs were analysed. For in-hospital mortality, quick Sequential Organ Failure Assessment (qSOFA) showed sensitivity of 0.61 and specificity of 0.73 (AUC=0.67); National Early Warning Score (NEWS)-family had higher sensitivity (0.80) but lower specificity (0.56, AUC=0.68); Modified Early Warning Score (MEWS) provided high specificity (0.83) but lower sensitivity (0.53, AUC=0.68). For ICU admission, NEWS demonstrated balanced performance (sensitivity 0.74, specificity 0.77, AUC=0.76), while Rapid Emergency Medicine Score (REMS) achieved the highest overall accuracy (sensitivity 0.79, specificity 0.85, AUC=0.81). Considerable heterogeneity was observed across studies.ConclusionsEarly warning scores show moderate accuracy in predicting ICU admission and in-hospital mortality in LMIC emergency departments. qSOFA and MEWS may be most useful when high specificity is required, while NEWS and REMS provide more balanced performance, with REMS showing the best overall accuracy. No single score was optimal across settings, underscoring the need for local validation and adaptation in resource-limited environments.Keywords: Palliative Care, Physical Examination, Respiratory Function Test WHAT IS ALREADY KNOWN ON THIS TOPIC Early warning and severity scores (eg, qSOFA (quick Sequential Organ Failure Assessment), NEWS (National Early Warning Score), MEWS (Modified Early Warning Score), REMS (Rapid Emergency Medicine Score)) are commonly used to support intensive care unit (ICU) triage and mortality prediction in emergency departments. Early warning and severity scores (eg, qSOFA (quick Sequential Organ Failure Assessment), NEWS (National Early Warning Score), MEWS (Modified Early Warning Score), REMS (Rapid Emergency Medicine Score)) are commonly used to support intensive care unit (ICU) triage and mortality prediction in emergency departments. Most evidence on their diagnostic performance originates from high-income countries, where patient characteristics and healthcare resources differ substantially from those in low- and middle-income countries (LMICs). Most evidence on their diagnostic performance originates from high-income countries, where patient characteristics and healthcare resources differ substantially from those in low- and middle-income countries (LMICs). WHAT THIS STUDY ADDS This systematic review and meta-analysis synthesise available evidence on the accuracy of prognostic scoring tools for ICU admission and in-hospital mortality in LMIC emergency settings. This systematic review and meta-analysis synthesise available evidence on the accuracy of prognostic scoring tools for ICU admission and in-hospital mortality in LMIC emergency settings. It shows that commonly used scores have only moderate diagnostic accuracy, with marked variation in sensitivity and specificity across tools and settings. It shows that commonly used scores have only moderate diagnostic accuracy, with marked variation in sensitivity and specificity across tools and settings. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY Clinicians in LMICs should apply existing scoring systems with caution and select tools based on local clinical priorities and resource constraints. Clinicians in LMICs should apply existing scoring systems with caution and select tools based on local clinical priorities and resource constraints. The findings support the need for local validation, adaptation or development of context-specific triage tools to improve critical care allocation in resource-limited settings. The findings support the need for local validation, adaptation or development of context-specific triage tools to improve critical care allocation in resource-limited settings. Introduction Early recognition of critically ill patients is a cornerstone of emergency and critical care. In high-income countries, clinical prediction scores, such as the quick Sequential Organ Failure Assessment (qSOFA), Systemic Inflammatory Response Syndrome (SIRS), Modified Early Warning Score (MEWS) and National Early Warning Score (NEWS and NEWS2), are widely applied to guide triage, anticipate deterioration and inform decisions regarding intensive care unit (ICU) admission.1 2 These scoring tools are valued for their simplicity, rapid bedside applicability and potential to support non-specialist clinicians in making timely decisions. In low- and middle-income countries (LMICs), the need for effective risk stratification is even more urgent. These settings face disproportionate burdens of infectious diseases, sepsis and trauma, while simultaneously experiencing resource constraints such as limited ICU capacity, high patient-to-staff ratios and delays in presentation.3 4 Mortality rates among acutely ill patients in LMICs are substantially higher compared with high-income countries, with some reports exceeding 50% in septic or critically ill cohorts.5 6 In such environments, robust and easy-to-use prognostic tools may assist clinicians in prioritising scarce resources and improving outcomes. However, the predictive performance of these scoring systems in LMICs remains inconsistent. qSOFA, while widely adopted, often demonstrates limited sensitivity in identifying high-risk patients.7 In contrast, NEWS and its derivatives tend to have higher sensitivity but reduced specificity, potentially leading to unnecessary ICU admissions where capacity is already constrained.8 9 Other scores such as Rapid Emergency Medicine Score (REMS), Sequential Organ Failure Assessment (SOFA) and the Universal Vital Assessment (UVA) have shown promise in certain contexts, yet results remain heterogeneous across populations and disease profiles.10 11 Previous studies in LMICs have generally been single-centre, limited in sample size or focused on specific patient subgroups, leading to uncertainty about which scoring tools are most reliable for predicting ICU admission and in-hospital mortality across diverse clinical contexts.12 13 Therefore, the objective of this systematic review and meta-analysis is to evaluate and compare the diagnostic accuracy of widely used early warning and severity scores, including qSOFA, SIRS, NEWS, MEWS, SOFA, REMS and UVA, for predicting ICU admission and in-hospital mortality among adult patients in LMIC emergency departments. By synthesising available evidence, this study seeks not only to clarify which tools provide the most reliable predictive performance in resource-limited contexts but also to inform clinical practice and guide policymakers in adopting practical, evidence-based approaches to critical care triage. Method Search strategy A thorough search was carried out in accordance with the Cochrane Handbook for Systematic Reviews of Interventions 2024.14 Multiple databases were used in this review, including PubMed, Cochrane Library, Science Direct and Web of Science. The study adhered to Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines, and the review protocol was pre-registered with PROSPERO (CRD420251118470, available from https://www.crd.york.ac.uk/PROSPERO/view/CRD420251118470) in August 2025. The inclusion and exclusion criteria were organised according to Population, Intervention, Comparison, Outcome and Study (PICOS) criteria (see online supplemental supplement S1 and table 1). bmjresp-13-1-s001.pdf We included full-length reports from peer-reviewed journals, specifically observational studies such as cohort or cross-sectional studies, that focused on adult patients (aged ≥18 years) who were admitted to the emergency department. To be considered, a study had to describe assessment using at least one of the following scores: qSOFA, SIRS, NEWS, MEWS or UVA. We also considered other commonly used scores such as SOFA and NEWS2, as well as modified versions of the original scoring systems where applicable. Crucially, the studies needed to report the relationship between these scoring systems and key outcomes, such as sensitivity and specificity for predicting ICU admission or in-hospital mortality. Additionally, we used the filter proposed by the Cochrane Collaboration based on the World Bank list of low-income, lower-middle-income or upper-middle-income countries.15 No restrictions regarding language of publication were applied in the search strategy or study selection process. We excluded studies not conducted in LMICs, those that lacked sufficient data to calculate score performance metrics (eg, sensitivity or specificity) and those limited to specific patient populations (eg, patients with COVID-19). Study selection Two independent investigators (THT and HTTN) systematically screened all studies identified from the search strategy for eligibility. The screening process was conducted in two phases: first by reviewing titles and abstracts, followed by a full-text review of potentially relevant articles. Disagreements were resolved through discussion and consensus. If an agreement could not be reached, a third invited investigator was consulted to make a final decision. Data were extracted from each eligible study, including but not limited to the first author, publication year, country, study design, patient population and the specific predictive scores and outcomes reported. Any missing information was noted for further follow-up. The search flow is summarised in figure 1. The PRISMA flowchart of the process of searching and selecting studies. ED, emergency department; LMICs, low- and middle-income countries; PRISMA, Preferred Reporting Items for Systematic Reviews and Meta-Analyses. Outcome assessments The primary outcome was short-term mortality. Mortality was defined as either in-hospital mortality or time-specific mortality (including 28-day or 30-day mortality), depending on the definition reported in each included study. In-hospital mortality was defined as all-cause death occurring during the index hospitalisation, whereas time-specific mortality referred to all-cause death within a defined follow-up period after admission or diagnosis. For the meta-analysis, all mortality outcomes were synthesised under a unified construct of short-term mortality. No transformation or recalculation of mortality time points was performed. Given the heterogeneity in follow-up periods, subgroup analyses based on mortality time definitions were planned but were not feasible due to the limited number of studies reporting time-specific outcomes. In all included studies, clinical prediction scores (eg, qSOFA, SIRS, MEWS, UVA, NEWS) were calculated at admission or during the early phase of hospitalisation, and their predictive performance was assessed against observed outcomes. Diagnostic accuracy measures, including sensitivity, specificity, likelihood ratios, predictive values and the area under the receiver operating characteristic curve, were extracted. According to published definitions, standard cut-off values were used for interpretation of each scoring system (eg, qSOFA≥2, SIRS≥2, SOFA≥2, NEWS2≥5, MEWS≥5 and UVA≥5). However, some variability in cut-off thresholds across studies was observed, particularly for NEWS and MEWS. Therefore, all data were extracted using the original thresholds reported in each study without recalibration to a unified cut-off. Pooled estimates were generated using a bivariate random-effects model to account for between-study heterogeneity. No studies were excluded based on differences in cut-off definitions. Data extraction and quality assessment The data extraction form captured key information, including the first author’s name, publication year, country of origin, study design, sample size, mortality rate, patient selection criteria and the specific scores evaluated. Where information was missing, we made an effort to contact the corresponding authors for clarification. The primary outcome of interest for this review was overall mortality, defined as either in-hospital or 28-day/30-day mortality. The secondary outcome was the ICU admission with any reason. We evaluated the risk of bias and methodological quality of included studies using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tool, as recommended by the Cochrane Collaboration.16 Two independent reviewers, THT and NTTH, assessed the risk of bias for each study, resolving any discrepancies through discussion. A comprehensive quality assessment, including details on rejected articles, is provided in the supplemental files. Statistical analysis Statistical analyses were performed using RevMan V.5.4 (Nordic Cochrane Centre, Copenhagen, Denmark) and Meta-DiSc v.2.0 (Ramón y Cajal Research Institute, Spain)17 . For each included study, 2×2 contingency tables (true positives, false positives, false negatives and true negatives) were constructed from reported diagnostic accuracy data. Because diagnostic performance measures depend on the threshold applied, sensitivity and specificity were extracted using the cut-off defined in each original study. Consequently, pooled sensitivity and specificity represent summary estimates across study-specific thresholds rather than performance at a single common cut-off. Heterogeneity was assessed using Cochran’s Q test and the I² statistic, with I² >50% indicating substantial heterogeneity. Pooled sensitivity and specificity were calculated using a random-effects model. Summary receiver operating characteristic (SROC) curves were generated in Meta-DiSc, and the corresponding area under the curve (AUC) with 95% CIs was used to assess overall discriminative performance. The reported pooled AUCs were derived from SROC modelling rather than direct pooling of study-level AUC values. Study-level AUC values extracted from the original reports were used descriptively to summarise overall discrimination of each score and should be interpreted separately from threshold-specific sensitivity and specificity. Each study was treated as an independent dataset, as no overlapping populations were identified. Formal meta-regression to explore heterogeneity was considered; however, because the number of studies within each score–outcome subgroup was limited and study-level characteristics were inconsistently reported, such analyses were considered insufficiently robust for reliable inference and were not presented. Results In total, 1883 initially published articles were identified (785 from PubMed, 146 from Science Direct, 396 from the Cochrane Library and 556 from Web of Science). After removing duplicates, 1362 eligible articles were screened. Of these, 1214 articles were excluded based on titles and abstracts. A total of 148 articles were reviewed in full text. Among them, 122 articles were excluded for reasons presented in figure 1. Ultimately, 26 articles met our inclusion criteria for the systematic review and meta-analysis. All studies were published between 2015 and 2025. The studies were conducted in 11 countries/regions classified as LMICs, including 4 in Asia, 5 in Africa and 2 in the Americas. The highest number of studies were carried out in Turkey (n=7), Brazil (n=3) and Colombia (n=3), while each of the other countries contributed one study. The characteristics of these studies are summarised in table 1. The number of patients in each study ranged from 112 to 2680. Reported mortality rates ranged from 2.36% to 69.7%, and ICU admission rates ranged from approximately 1.25% to 41.18%. Mortality outcomes varied across studies, with most reporting in-hospital mortality (n=23), while a minority reported time-specific outcomes such as 1-day (n=1), 1-week (n=1) or 30-day (n=1) mortality. The most frequently reported conditions across studies included infections, pulmonary embolism, acute exacerbations of chronic obstructive pulmonary disease and drug intoxication (table 1). Table 1Summary of included studiesArticlesCountryNo. of sitesStudy designParticipants diagnosedNo. of patientsOutcome availabilityHospital mortalityMortality time pointICU admissionDundar et al1Turkey1Prospective, single-centred observational studyNot reported671YesIn-hospital mortalityYesWixon-Genack et al2Thailand1Prospective observational studyInfection within the prior 24 hours2.680Yes1-day mortalityYesKayambankadzanja et al3Malawi2Prospective studyPatients with infectious1.117YesIn-hospital mortality (censored at 30 days)NoBonnewell et al4Tanzania1Prospective cohort studyPatients with febrile1132YesIn-hospital mortalityNoHazard et al5Rwanda8Retrospective studyAcute infection573YesIn-hospital mortalityNoÇelik and Karaca6Northern Syria7Cross-sectional studyNot reported1.310YesIn-hospital mortalityYesYeşil et al7Turkey1Single-centre, prospective cohort studyNot reported350YesIn-hospital mortalityYesCasallas-Barrera et al8Colombia1Observational retrospective studySepsis126YesIn-hospital mortalityNoGuzelce et al9Turkey1Cross-sectional studyDecompensated heart failure130YesIn-hospital mortalityNoAbbey et al10Ghana1Retrospective studyNot reported112YesIn-hospital mortalityNoAykol et al11Turkey1Single-centre prospective studyPatients with decreased general condition137YesIn-hospital mortalityYesBidart et al12Brazil1Retrospective cohort studySepsis665Yes30-day mortalityNoSingh et al13India1Cross-sectional studyAcute exacerbate COPD200YesIn-hospital mortalityNoLashin and Sharif18Egypt1Retrospective cohort studyAcute clozapine poisoning156YesIn-hospital mortalityNoSagah and Elmansy19Egypt1Cross-sectional studyacute theophylline poisoning58NoIn-hospital mortalityYesÖzkan and Özdemir20Turkey1Prospective studyPulmonary embolism132YesIn-hospital mortalityYesTaslidere et al21Turkey1Retrospective studyUpper gastrointestinal bleeding133YesIn-hospital mortalityYesİlhan et al22Turkey1Retrospective cohort studyAcute exacerbate COPD577YesIn-hospital mortalityYesGarbero et al23Brazil1Single-centre retrospective studyNot reported184YesIn-hospital mortalityNoYolcu et al25Turkey2Retrospective studyPulmonary embolism245Yes1-week mortalityYesKlinger et al24Rwandan1Prospective observational studySuspected infection with temperature ≤35.0°C or ≥38.0°C647YesIn-hospital mortalityNoMontenegro and Miranda26Brazil1Cross-sectional studyNot reported300YesIn-hospital mortalityNoDundar et al27Turkey1Retrospective observational studyNot reported455YesIn-hospital mortalityNoCarugati et al28Tanzania2Prospective cohort studyFebrile patients870YesIn-hospital mortalityNoDevia Jaramillo and Ibáñez Pinilla29Colombia1Retrospective cohort studyNot reported179YesIn-hospital mortalityNoDevia-Jaramillo et al30Colombia1Retrospective studySuspected sepsis383YesIn-hospital mortalityNoCOPD, chronic obstructive pulmonary disease. Summary of included studies COPD, chronic obstructive pulmonary disease. The studies employed a wide variety of scoring systems, among which the most frequently used were the NEWS scale (8 studies) and its variants (including NEWS2 and NEWS-Lactate), as well as SOFA and qSOFA (17 studies) and MEWS and MEWS-Lactate (8 studies) for predicting in-hospital mortality (see online supplemental table 2). In contrast, NEWS and its variants (six studies) and REMS (five studies) were most used for predicting ICU admission rates (see online supplemental table 3). Only two studies evaluated a single scoring tool, whereas most studies compared the predictive performance of two or more scores. All included studies were descriptive in design, comprising 5 cross-sectional studies, 9 prospective studies and 12 retrospective studies. Overall, the studies were well-designed, and quality assessments indicated a low risk of bias. The QUADAS-2 appraisal is presented in figure 2, with detailed assessments provided in online supplemental table 4. Risk of bias of included studies. Predicting in-hospital mortality Because the included studies applied different cut-off values for the same scoring system, the pooled sensitivity and specificity estimates should be interpreted as summary measures across study-specific thresholds rather than as accuracy at a single uniform cut-off. The pooled analyses showed that the qSOFA family, across 15 studies, had low sensitivity (Se=0.61; 95% CI 0.41 to 0.78) but relatively high specificity (Sp=0.73; 95% CI 0.58 to 0.84), with a pooled AUC of 0.67 (95% CI 0.50 to 0.81). The NEWS family, evaluated in nine studies, demonstrated higher sensitivity (SE=0.80; 95% CI 0.77 to 0.83) but lower specificity (Sp=0.56; 95% CI 0.44 to 0.68), with a pooled AUC of 0.68 (95% CI 0.61 to 0.76). MEWS, assessed in six studies, showed lower sensitivity (Se=0.53; 95% CI 0.33 to 0.71) but higher specificity (Sp=0.83; 95% CI 0.76 to 0.88), with a pooled AUC of 0.68 (95% CI 0.55 to 0.80). Overall, pooled AUCs were very similar across the qSOFA family, NEWS family and MEWS, suggesting broadly comparable overall discriminative performance for in-hospital mortality. The main differences were observed in their sensitivity–specificity profiles, with NEWS favouring sensitivity and qSOFA/MEWS favouring specificity. These findings therefore reflect different threshold-dependent trade-offs rather than clear evidence that one score is globally superior to another (figure 3). Forest plots and SROC curves for in-hospital mortality: qSOFA family (3a, sensitivity; 3b, specificity; 3c, SROC curve), NEWS family (3d, sensitivity; 3e, specificity; 3f, SROC curve) and MEWS (3g, sensitivity; 3h, specificity; 3i, SROC curve). MEWS, Modified Early Warning Score; NEWS, National Early Warning Score; SROC, summary receiver operating characteristic; qSOFA, quick Sequential Organ Failure Assessment. In addition to the commonly used scores, several other early warning tools were reported (online supplemental figure 1). As each was evaluated in fewer than three studies, meta-analysis was not feasible and findings were summarised narratively. REMS and RAPS showed good discrimination, with reported AUCs of 0.91 and 0.93, respectively.6 In the study by Çelik and Karaca, HOTEL (Hypotension, Oxygenation, Temperature, Electrocardiography, Loss of consciousness score) also demonstrated high discrimination (AUC=0.90),6 whereas SIRS showed limited accuracy because of low specificity.4 7 Other scores, including UVA, ViEWS (VitalPAC Early Warning Score), CURB65 (Confusion Urea Respiratory rate Blood pressure age ≥65 score), BAP65 (Blood urea nitrogen, Altered mental status, Pulse, and age ≥65 score), IEWS (Individual Early Warning Score), qNEWS (modified Quick NEWS), qPitt (quick Pitt score) and WPS (Worthing Physiological Score), were evaluated in only one or two studies, with reported AUCs ranging from 0.60 to 0.89. Predicting ICU admission As in the mortality analysis, the pooled estimates for ICU admission reflect summary performance across study-specific thresholds because the included studies used different cut-off values for the same score. The pooled analysis showed that qSOFA, across three studies, had a sensitivity of 0.77 (95% CI 0.49 to 0.92) and specificity of 0.41 (95% CI 0.25 to 0.59), with a pooled AUC of 0.59 (95% CI 0.37 to 0.76). NEWS, across five studies, demonstrated a sensitivity of 0.74 (95% CI 0.67 to 0.79) and specificity of 0.77 (95% CI 0.68 to 0.84), with a pooled AUC of 0.76 (95% CI 0.68 to 0.82). REMS, also across five studies, achieved a sensitivity of 0.79 (95% CI 0.54 to 0.93) and specificity of 0.85 (95% CI 0.77 to 0.91), with the highest pooled AUC of 0.81 (95% CI 0.66 to 0.92). Unlike the mortality analysis, pooled AUCs for ICU admission differed more noticeably across scores, suggesting that REMS and NEWS may provide better overall discrimination than qSOFA for this outcome. However, these comparisons remain descriptive and should not be interpreted as formal statistical differences between scoring systems (figure 4). Forest plots and SROC curves for ICU admission: qSOFA (4a, sensitivity; 4b, specificity; 4c, SROC curve), NEWS (4d, sensitivity; 4e, specificity; 4f, SROC curve) and REMS (4g, sensitivity; 4h, specificity; 4i, SROC curve). ICU, intensive care unit; NEWS, National Early Warning Score; REMS, Rapid Emergency Medicine Score; SROC, summary receiver operating characteristic; qSOFA, quick Sequential Organ Failure Assessment. Several additional early warning and severity scoring systems have also been evaluated for predicting ICU admission in emergency department patients but were not included in the meta-analysis because fewer than three studies were available for each score. REMS, MEWS and PSS showed high discriminative performance in individual studies, with AUCs ranging from 0.88 to 0.98 and sensitivities often exceeding 90% at the reported optimal cut-offs.18 19 The Simple Clinical Score (SCS) score also demonstrated excellent accuracy, with an AUC of 0.96, sensitivity of 92.9%, and specificity of 90.9%.19 Other tools, such as HOTEL, PESI (Pulmonary Embolism Severity Index), WPS and GBS (Glasgow-Blatchford Score), showed moderate predictive ability, with AUCs between 0.70 and 0.90, generally reflecting different trade-offs between sensitivity and specificity depending on the selected threshold.20 21 Modified NEWS-based systems, including CREWS, S-NEWS and qNEWS, also performed well, with reported AUCs of approximately 0.80–0.84 and relatively balanced sensitivity (67–76%) and specificity (80–85%).22 By contrast, SOFA showed very high sensitivity (93.7%) but poor specificity (25.9%).23 Overall, these findings suggest that more complex or composite scores, such as REMS, SCS and PSS, may offer better prognostic discrimination for ICU admission than simpler single-parameter models, although further validation in large multicentre cohorts is required (online supplemental figure 1). Discussion This systematic review and meta-analysis demonstrated substantial variation in the ability of early warning scores to predict in-hospital mortality among adults presenting to emergency departments in LMICs. The pooled results showed that qSOFA, although widely used because of its simplicity and bedside applicability, had relatively low sensitivity (0.61) but higher specificity (0.73), with an AUC of 0.67. This is consistent with prior evidence suggesting that qSOFA may be more suitable for ‘ruling in’ critically ill patients than for identifying all those at risk of deterioration.7 23 In contrast, the NEWS family showed higher sensitivity (0.80) but lower specificity (0.56), with an AUC of 0.68, suggesting better identification of patients at risk of deterioration but with more false positives.8 9 MEWS, evaluated in six studies, had a sensitivity of 0.53 and specificity of 0.83, with an AUC of 0.68, indicating a pattern similar to qSOFA, with better performance for confirming high-risk cases than for broad screening. Comparable findings have been reported in both high- and middle-income settings, particularly among septic populations.1 10 Despite differences in sensitivity and specificity, the overall discriminatory ability of qSOFA, NEWS and MEWS was only moderate, with AUCs clustering around 0.67–0.68. This suggests that no single score achieves both high sensitivity and high specificity in LMIC emergency departments. Variation in patient populations, disease burden and healthcare system constraints may partly explain these findings, particularly given the diversity of emergency presentations in LMICs.3 5 In addition to the commonly used scores, several other early warning tools, including REMS, HOTEL and UVA, were evaluated in fewer studies. REMS, in particular, showed promising discriminatory performance, with AUC values as high as 0.93 in some single-centre cohorts.11 20 These findings suggest that more complex scoring systems incorporating additional physiological or laboratory parameters may improve predictive accuracy. However, their practicality in LMICs remains uncertain, given limited laboratory capacity and overstretched staff. Substantial heterogeneity was observed across included studies, with I² exceeding 80% for several pooled estimates. This likely reflects variation in study design, patient case-mix, disease spectrum, score thresholds and healthcare context. For example, predictive performance often differed between sepsis-dominated cohorts and broader medical or surgical emergency populations.8 12 These findings underscore the need to interpret pooled estimates in context and to validate scoring tools locally before large-scale implementation. Another important consideration is the role of training and system-level support in the effective use of early warning scores. Studies from sub-Saharan Africa and South Asia suggest that even simple scores may underperform when frontline clinicians are inadequately trained or when escalation protocols are lacking.2 24 Thus, although this review highlights differences in score performance, their real-world impact depends on integration into broader emergency and critical care pathways. From a practical perspective, early warning scores should be used as adjuncts to, rather than substitutes for, clinical judgement. In LMICs, scores with more balanced sensitivity and specificity, such as NEWS or REMS, may support earlier recognition of deterioration and escalation of care, whereas highly specific tools such as qSOFA or MEWS may be more useful for prioritising ICU admission when resources are severely constrained. Ultimately, careful score selection and local adaptation, supported by training and healthcare infrastructure, are essential to optimise resource use and reduce preventable in-hospital mortality.2 5 For predicting ICU admission in LMIC emergency departments, the pooled findings showed that no single score achieved both high sensitivity and high specificity; rather, each had distinct strengths and limitations that may influence its usefulness across settings. qSOFA, assessed in three studies, had a sensitivity of 0.77, specificity of 0.41, and an AUC of 0.59, suggesting moderate ability to identify patients requiring escalation of care but limited discrimination due to a high false-positive rate. This is consistent with previous concerns that qSOFA, although simple and easy to apply at the bedside, may be insufficient as a sole predictor of ICU transfer.7 23 In LMICs, where ICU capacity is severely constrained, reliance on qSOFA alone may contribute to unnecessary admissions and misallocation of scarce resources. In contrast, the NEWS family demonstrated more balanced performance, with a sensitivity of 0.74, specificity of 0.77 and an AUC of 0.76, highlighting its value as a practical tool for predicting ICU need in resource-limited settings. Its reliance on routinely available vital signs enhances feasibility, while its relatively high specificity may reduce inappropriate ICU referrals.8 9 Evidence from individual LMIC studies also suggests that NEWS may identify early clinical deterioration more reliably than qSOFA, supporting its use for triage and escalation decisions in busy emergency departments.2 REMS, evaluated in five studies, showed the highest pooled accuracy, with a sensitivity of 0.79, specificity of 0.85 and an AUC of 0.81, suggesting better overall performance than qSOFA and NEWS for predicting ICU admission. Its superior discrimination may reflect the inclusion of additional physiological and laboratory variables.11 19 20 However, its greater complexity may limit routine use in overcrowded or understaffed LMIC emergency settings. Several other tools, including SCS, PSS, HOTEL and modified NEWS-based systems, also demonstrated promising discriminatory performance, often with AUC values above 0.85.18 21 22 However, the small number of studies and greater data requirements for some of these tools limit firm conclusions and may restrict their routine use in resource-limited emergency departments. Overall, these findings suggest that score selection for predicting ICU admission in LMICs should be context-specific. In settings with extremely limited ICU capacity, tools with higher specificity, such as REMS or MEWS, may better support ICU prioritisation, whereas more sensitive tools such as NEWS may be preferable where ward-based monitoring and early escalation of care are feasible. Together, these findings indicate that a one-size-fits-all approach is unlikely to be effective. Substantial heterogeneity was observed across included studies, reflecting differences in disease patterns, patient populations, clinical settings, healthcare infrastructure and score cut-off thresholds. For example, studies involving toxicological emergencies often reported different predictive performance compared with cohorts dominated by infectious diseases.12 19 Variations in patient age, comorbidities and institutional resources may have further contributed to the observed inconsistency. Consequently, pooled estimates should be interpreted cautiously, and local validation remains essential before widespread implementation. The strength of evidence also varied across scoring systems. Findings for commonly studied scores such as qSOFA, NEWS and MEWS are supported by a larger evidence base and may therefore be considered more robust, whereas evidence for alternative tools, including REMS, HOTEL, SCS and modified scores, is derived from relatively few studies and should be regarded as preliminary. From a practical perspective, early warning scores should complement rather than replace clinical judgement and institutional protocols. Integrated into structured triage systems, they may assist clinicians in identifying patients who require escalation of care and help reduce preventable delays in ICU admission. Among the evaluated tools, REMS demonstrated the best overall performance for predicting ICU admission, although trade-offs between sensitivity, specificity and feasibility must be considered in resource-limited settings. Ultimately, context-specific implementation, supported by staff training, system-level improvements and local validation, is essential to optimise triage efficiency and improve patient outcomes in LMIC emergency departments.2 24 Limitations This review has several limitations. First, some scoring systems, including REMS, RAPS, HOTEL, SIRS and UVA, were evaluated in fewer than three studies, precluding meta-analysis and limiting synthesis to narrative summary. Second, substantial heterogeneity was observed across pooled analyses, likely due to differences in study design, patient case mix, clinical setting, outcome definitions and cut-off thresholds. As sensitivity and specificity are threshold-dependent, pooled estimates should be interpreted as average diagnostic performance across study-specific thresholds rather than evidence of a single optimal cut-off. The small number of studies within score–outcome subgroups and inconsistent reporting of study-level covariates also limited formal exploration of heterogeneity. Third, most included studies were single-centre and retrospective, increasing the risk of selection bias and limiting external validity. Fourth, publication bias cannot be excluded despite the comprehensive search strategy. Finally, some scores, such as REMS and SOFA, require laboratory parameters, which may reduce their practicality in time-sensitive emergency settings and in resource-limited low- and middle-income countries. These limitations underscore the need for larger prospective multicentre studies, particularly in LMIC settings. Conclusion This systematic review and meta-analysis evaluated the performance of early warning and severity scores for predicting in-hospital mortality and ICU admission among adults presenting to emergency departments in LMICs. Overall, qSOFA and MEWS showed higher specificity but lower sensitivity, making them more useful for identifying patients at highest risk than for detecting all patients likely to deteriorate. In contrast, NEWS and REMS demonstrated more balanced accuracy, with REMS showing the best overall predictive performance. In resource-limited settings, score selection should be context-specific, depending on whether the priority is ICU triage or early clinical escalation. However, no single score performed optimally across all settings, highlighting the need for local validation and adaptation. These findings should be interpreted cautiously given substantial heterogeneity and limited evidence for several scoring systems. Footnotes Collaborators: None. Contributors: Conceptualisation: Thong Huu Tran, HTTN. Data curation, formal analysis, validation and writing – original draft: Tran Huu Thong, HTT, Tran Huu Trung. Writing – review and editing: Tran Huu Thong, HTTN, Tran Huu Trung. Tran Huu Thong is the guarantor of this work. Artificial intelligence tools (ChatGPT, OpenAI) were used to assist with language editing and improvement of academic writing clarity. The AI was not used for data analysis, interpretation or generation of study results. All scientific content, analyses and conclusions were developed and verified by the authors. Funding: The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors. Competing interests: None declared. Provenance and peer review: Not commissioned; externally peer reviewed. Supplemental material: This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise. Data availability statement All data relevant to the study are included in the article or uploaded as supplementary information. Ethics statements Patient consent for publication Not applicable. Ethics approval This study is a systematic review and meta-analysis of published literature and does not involve the collection or analysis of individual-level patient data. Therefore, ethical approval was not required. In accordance with institutional and national research guidelines, formal approval from an ethics committee was not necessary for this type of study. References Associated Data Supplementary Materials bmjresp-13-1-s001.pdf Data Availability Statement All data relevant to the study are included in the article or uploaded as supplementary information.