Development and internal validation of the AGATE hybrid score for pre-endoscopic risk stratification in acute upper gastrointestinal bleeding: a single-centre, retrospective cohort study Authors: Yavuz Özden, Ömer Yüzügülen, Ferhat Omurca Published: 2026 ABSTRACT AbstractObjectiveTo compare five established pre-endoscopic risk scores in acute upper gastrointestinal bleeding (AUGIB) and to derive a hybrid score (Acute Gastrointestinal Bleeding ABC-GBS Triage Evaluation (AGATE)) integrating mortality prediction and triage utility.MethodsWe conducted a single-centre retrospective cohort study at Kayseri City Hospital, Kayseri, Türkiye, including 5000 consecutive adults with endoscopically confirmed AUGIB (January 2020–January 2025). The primary outcome was 30-day mortality; secondary outcomes were therapeutic intervention and 30-day rebleeding. Age-Blood tests-Comorbidities (ABC), Mental status-ASA class-Pulse-Albumin-Systolic blood pressure-Haemoglobin (MAP(ASH)), Haematemesis-Heart rate-Haemoglobin-Blood pressure-Blood urea nitrogen, Glasgow-Blatchford score (GBS) and Albumin, INR, Mental status, Systolic blood pressure, Age ≥65 were calculated at presentation. AGATE was derived using multivariable logistic regression for the composite end point (mortality or intervention) and internally validated with 1000 bootstrap resamples.ResultsMortality was 9.2% (460/5000), intervention was required in 71.0% (3550/5000) and rebleeding occurred in 14.5% (725/5000). For mortality, ABC (area under the receiver operating characteristic curve (AUROC) 0.83, 95% CI 0.80 to 0.86) and MAP(ASH) (AUROC 0.81, 95% CI 0.78 to 0.84) demonstrated the highest discrimination; for intervention, GBS (AUROC 0.78, 95% CI 0.74 to 0.82) and MAP(ASH) (AUROC 0.77, 95% CI 0.73 to 0.80) demonstrated the highest discrimination. AGATE achieved AUROCs of 0.85 (95% CI 0.82 to 0.87) for mortality and 0.80 (95% CI 0.77 to 0.82) for intervention. For the composite end point, AGATE outperformed GBS (AUROC 0.86, 95% CI 0.83 to 0.88 vs 0.77 95% CI 0.74 to 0.80; p<0.001), with an optimism-corrected AUROC of 0.843. Using a low-risk cut-off (≤5), AGATE classified 38% (1900/5000) as low risk (0.3% mortality, 4.0% intervention), compared with 15% (750/5000) for GBS ≤1.ConclusionAGATE provides balanced pre-endoscopic triage by integrating predictors from ABC and GBS, expanding the low-risk group while maintaining very low short-term mortality. Prospective external validation in independent multicentre settings is required. WHAT IS ALREADY KNOWN ON THIS TOPIC Existing pre-endoscopic risk scores in acute upper gastrointestinal bleeding (AUGIB) are typically optimised either for mortality prediction (eg, Age-Blood tests-Comorbidities) or for identifying very low-risk patients suitable for discharge (eg, Glasgow-Blatchford score), but not both. Existing pre-endoscopic risk scores in acute upper gastrointestinal bleeding (AUGIB) are typically optimised either for mortality prediction (eg, Age-Blood tests-Comorbidities) or for identifying very low-risk patients suitable for discharge (eg, Glasgow-Blatchford score), but not both. WHAT THIS STUDY ADDS In a large single-centre cohort of 5000 patients, the Acute Gastrointestinal Bleeding ABC-GBS Triage Evaluation score demonstrated balanced discrimination for 30-day mortality and need for therapeutic intervention while identifying a substantially larger low-risk group with very low short-term event rates. In a large single-centre cohort of 5000 patients, the Acute Gastrointestinal Bleeding ABC-GBS Triage Evaluation score demonstrated balanced discrimination for 30-day mortality and need for therapeutic intervention while identifying a substantially larger low-risk group with very low short-term event rates. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY A unified hybrid score may streamline emergency triage, reduce unnecessary hospital admissions and support safe outpatient management pathways, pending prospective external validation. A unified hybrid score may streamline emergency triage, reduce unnecessary hospital admissions and support safe outpatient management pathways, pending prospective external validation. Introduction Acute upper gastrointestinal bleeding (AUGIB) remains a common and potentially life-threatening emergency associated with substantial morbidity, mortality and healthcare utilisation. Despite advances in endoscopic haemostasis and pharmacological management, the global incidence of AUGIB is estimated at approximately 97–150 per 100 000 adults and short-term mortality has remained between 8% and 10% over the past two decades, even in high-resource settings.1 2 This persistent mortality burden underscores the need for accurate early risk stratification to guide timely triage, optimise resource allocation and improve patient outcomes.3 Risk assessment at initial presentation serves two distinct but complementary clinical objectives.4 First, it should reliably identify patients at high risk who require urgent intervention, intensive monitoring and aggressive resuscitation to prevent death or major complications. Second, it should safely recognise a low-risk subgroup suitable for early discharge or ambulatory management, thereby reducing unnecessary hospital admissions and associated healthcare costs. Several pre-endoscopic prognostic models have been developed to address these objectives and are incorporated into contemporary clinical guidelines.57 Current international recommendations acknowledge that no single risk score optimally fulfils both purposes, often leading to the parallel use of multiple tools in routine practice.8 9 The Glasgow-Blatchford Score (GBS) was originally developed to predict the need for hospital-based intervention or death4 and remains widely used to identify very low-risk patients unlikely to require inpatient management.8 9 However, its discrimination for mortality alone is modest, as mortality prediction was not its primary design objective.3 In contrast, the Age-Blood tests-Comorbidities (ABC) score integrates age, haemoglobin, urea, albumin, creatinine and comorbidity burden and has demonstrated stronger discrimination for 30-day mortality.10 11 The Albumin, INR, Mental status, Systolic blood pressure, Age ≥65 (AIMS65) score offers simplicity and objectivity but performs less well for intervention-related outcomes.12 More recently developed models, including Mental status-ASA class-Pulse-Albumin-Systolic blood pressure-Haemoglobin (MAP(ASH)) and Haematemesis-Heart rate-Haemoglobin-Blood pressure-Blood urea nitrogen (H3B2) (also known as the Cologne-WATCH score), attempt to balance acute physiological derangement with chronic vulnerability; nevertheless, their performance varies across end points and populations.13 14 Recent comparative analyses have further highlighted variability in discrimination across clinical outcomes and patient populations.15 16 In real-world emergency settings, calculating separate scores to rule out low risk (eg, GBS) and rule in high mortality risk (eg, ABC) may increase cognitive load and workflow complexity at the point of care.15 A unified pre-endoscopic tool capable of providing balanced prognostic information for both mortality and the need for therapeutic intervention could therefore offer practical advantages. Importantly, the incremental value of such a hybrid score must be demonstrated relative to established tools, particularly for the composite end point historically associated with GBS. Despite the expanding number of available risk models, rigorous head-to-head comparisons within the same large, well-characterised cohort are scarce. Most existing scores were developed to optimise a single prognostic domain, leaving uncertainty as to whether a unified approach can achieve balanced discrimination while maintaining safety in low-risk patients. Furthermore, transparent evaluation against clinically meaningful composite end points is required to determine whether any new score offers true added value over existing standards. Accordingly, we aimed (i) to compare the performance of five established pre-endoscopic prognostic models—ABC, MAP(ASH), H3B2, GBS and AIMS65—in a large, real-world, single-centre cohort of patients with endoscopically confirmed AUGIB; (ii) to derive a pragmatic hybrid score, the Acute Gastrointestinal Bleeding ABC-GBS Triage Evaluation (AGATE) score, integrating predictors from ABC and GBS and (iii) to internally validate its discrimination and clinical utility for 30-day mortality, need for therapeutic intervention and rebleeding, with specific evaluation of its performance relative to the composite end point traditionally associated with GBS. Methods Study design and setting We conducted a single-centre retrospective cohort study to derive and internally validate a hybrid prognostic model for AUGIB. The study was performed at Kayseri City Hospital, University of Health Sciences, Kayseri, Türkiye, a large tertiary referral centre providing 24-hour emergency endoscopy services. All consecutive adult patients (aged ≥18 years) presenting to the emergency department or admitted under gastroenterology with suspected AUGIB between 1 January 2020 and 31 January 2025 were screened. Patients were included only if an endoscopically confirmed source of upper gastrointestinal bleeding was identified. The study is reported in accordance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis plus Artificial Intelligence reporting guideline (online supplemental file).17 Study population Patients were eligible if they met the following criteria: An endoscopically confirmed source of upper gastrointestinal bleeding. An endoscopically confirmed source of upper gastrointestinal bleeding. Complete baseline clinical and laboratory data required for calculation of established risk scores and candidate predictors. Complete baseline clinical and laboratory data required for calculation of established risk scores and candidate predictors. Both variceal and non-variceal bleeding aetiologies were included to reflect routine clinical practice. Patients were excluded if they met any of the following criteria: Terminal-stage malignancy precluding endoscopic evaluation. Terminal-stage malignancy precluding endoscopic evaluation. Major trauma-related gastrointestinal bleeding. Major trauma-related gastrointestinal bleeding. Transfer from another institution after initial resuscitation or endoscopy. Transfer from another institution after initial resuscitation or endoscopy. Missing data for any candidate predictor or required score component. Missing data for any candidate predictor or required score component. Patients without an identifiable bleeding source at endoscopy were excluded, as confirmation of the bleeding source was necessary to ensure consistent outcome adjudication and model derivation. Of 5800 screened patients, 800 were excluded, yielding a final cohort of 5000 individuals. Follow-up procedures Thirty-day follow-up was obtained through electronic medical records, hospital administrative databases and structured telephone interviews when outpatient documentation was incomplete. Telephone contact was attempted in 412 patients (8.2%); successful contact was achieved in 389 (94.4%). For the remaining 23 patients (0.5%), survival status was confirmed through national death registry linkage. No patient was missing 30-day mortality status and no patient was lost to follow-up for the primary outcome. Data collection Clinical and laboratory variables were extracted from the institutional electronic health record using a standardised data collection template. Data abstraction was performed independently by two investigators and discrepancies were resolved by a third adjudicator to ensure data accuracy. Vital signs were recorded at initial emergency department triage prior to definitive resuscitative interventions. Comorbidity burden was defined as the presence of two or more of the following conditions recorded at presentation: hypertension, diabetes mellitus, coronary artery disease, liver cirrhosis or active malignancy. These conditions were selected based on prior associations with adverse outcomes in bleeding cohorts, their inclusion in established models (eg, ABC and MAP(ASH)) and consistent availability in routine emergency practice. This simplified definition was used to enhance reproducibility in emergency settings. Calculation of established risk scores During the recruitment period, GBS and AIMS65 were variably used in routine clinical practice for triage decisions; however, management was not protocolised according to any single score. Derivation of the AGATE score Model construction Predictors with p<0.10 in univariable analysis or with strong clinical justification were entered into the multivariable logistic regression model, followed by backward elimination (p<0.05) to derive a parsimonious final model. Clinical relevance was defined a priori based on published evidence and consensus among three senior gastroenterologists experienced in AUGIB management.18 The number of outcome events relative to candidate predictors yielded an events-per-variable ratio exceeding 20, supporting model stability and reducing the risk of overfitting. Continuous variables were assessed for non-linearity using restricted cubic splines with three knots placed at the 10th, 50th and 90th percentiles.19 Non-linearity was evaluated through visual inspection of spline plots and comparison of model fit. No significant deviations from linearity were detected; continuous predictors were therefore retained as linear terms. Systolic blood pressure was modelled as a binary variable (<100 mm Hg vs ≥100 mm Hg), reflecting a widely accepted threshold for haemodynamic instability in resuscitation guidelines and prior risk models. Internal validation was performed using 1000 bootstrap resamples to estimate optimism-corrected model discrimination.20 Decision-curve analysis was subsequently performed using the net benefit framework to evaluate clinical utility across a range of threshold probabilities.21 Outcome definitions The primary outcome was 30-day all-cause mortality. Secondary outcomes were: Need for therapeutic intervention. Need for therapeutic intervention. 30-day rebleeding. 30-day rebleeding. Rebleeding was defined as recurrent haematemesis or melaena accompanied by haemodynamic instability (systolic blood pressure <100 mmHg or heart rate >100 beats/min), a haemoglobin decrease ≥20 g/L within 24 hours or the requirement for repeat therapeutic intervention within 30 days.22 23 Early rebleeding was defined as recurrence within 7 days of the index event. Model performance Discrimination was quantified using the area under the receiver operating characteristic curve (AUROC). AUROCs were compared using the DeLong test.24 Cut-off values were determined using receiver operating characteristic (ROC) analysis guided by the Youden index and by a predefined clinical requirement to achieve a negative predictive value exceeding 99% for mortality in the low-risk category. Decision-curve analysis was used to evaluate clinical utility and was not used to define cut-offs.25 Statistical analysis Continuous variables are presented as medians (IQR) or means (SD), as appropriate according to distribution. Categorical variables are presented as counts (percentages). Statistical analyses were performed using IBM SPSS Statistics V.25.026 and R V.4.3.2 (rms, package for receiver operating characteristic curve analysis and rmda packages).27 A complete-case analysis was performed. Patients with missing baseline candidate predictor data were excluded prior to model derivation, as detailed in figure 1 and online supplemental table 1. Study flow diagram. Flow diagram of patient selection. Of 5800 adults screened with suspected acute upper gastrointestinal bleeding (AUGIB), 800 were excluded (missing baseline data, n=342; absence of endoscopic confirmation, n=281; interhospital transfer after initial management, n=105; other reasons, n=72), resulting in a final cohort of 5000 patients included in the analysis. Results Patient characteristics and study flow Of the 5800 adults screened for suspected AUGIB, 800 were excluded, resulting in a final cohort of 5000 patients with endoscopically confirmed bleeding and complete baseline data. Figure 1 presents the study flow and reasons for exclusion, including missing baseline data (n=342), absence of endoscopic confirmation (n=281), interhospital transfer after initial management (n=105) and other reasons (n=72). Baseline characteristics of excluded patients are summarised in online supplemental table 1. Baseline demographic and clinical characteristics are presented in table 1. Median age was 62 years (IQR 49–75), and 67% were male. The most common comorbidities were hypertension (41%), diabetes mellitus (24%), coronary artery disease (16%) and liver cirrhosis (17%). Table 1Baseline characteristics of the study population (n=5000)VariableValueAge, years, median (IQR)62 (49–75)Male sex, n (%)3350 (67%)Comorbidities, n (%) Hypertension2050 (41%) Diabetes mellitus1200 (24%) Coronary artery disease800 (16%) Liver cirrhosis850 (17%)Type of bleeding, n (%) Non-variceal3950 (79%) Duodenal ulcer1250 (25%) Gastric ulcer900 (18%) Erosive gastritis/duodenitis750 (15%) Mallory-Weiss tear350 (7%) Oesophagitis350 (7%) Malignancy350 (7%) Variceal1050 (21%)Presenting symptom, n (%) Haematemesis (±melaena)2300 (46%) Isolated melaena2700 (54%)Admission vitals, median (IQR) Systolic blood pressure, mm Hg108 (90–128) Heart rate, beats/min99 (82–116)Laboratory values, median (IQR) Haemoglobin, g/L91 (76–112) Blood urea nitrogen, mg/dL61 (38–89) Serum albumin, g/dL3.0 (2.6–3.5) Creatinine, mg/dL1.1 (0.8–1.6) INR1.3 (1.1–1.7)Clinical severity, n (%) Intubation and active resuscitation650 (13%) ICU admission700 (14%)Hospital course Length of stay, days, median (IQR)5 (3–7) 30-day mortality, n (%)460 (9.2%) 30-day rebleeding, n (%)725 (14.5%)ICU, intensive care unit; INR, international normalised ratio. Baseline characteristics of the study population (n=5000) ICU, intensive care unit; INR, international normalised ratio. Non-variceal aetiologies accounted for 79% of bleeding episodes, while variceal bleeding was identified in 21%. Among non-variceal causes, peptic ulcer disease predominated (duodenal ulcer 25%, gastric ulcer 18%), followed by erosive gastroduodenitis (15%), Mallory-Weiss tear (7%), oesophagitis (7%) and upper gastrointestinal malignancy (7%). Haematemesis (with or without melaena) was reported in 46% of patients, whereas 54% presented with isolated melaena. At presentation, median systolic blood pressure was 108 mm Hg (IQR 90–128) and median heart rate was 99 beats/min (IQR 82–116). Median haemoglobin was 91 g/L (IQR 76–112), blood urea nitrogen 61 mg/dL (IQR 38–89), serum albumin 3.0 g/dL (IQR 2.6–3.5), creatinine 1.1 mg/dL (IQR 0.8–1.6) and INR 1.3 (IQR 1.1–1.7). Haemodynamic instability or altered mental status requiring aggressive resuscitation occurred in 13% of patients. Intensive care unit admission occurred in 14% and median length of hospital stay was 5 days (IQR 3–7). Follow-up completeness Thirty-day outcomes were ascertained from electronic records and administrative databases, with structured telephone follow-up used when outpatient documentation was incomplete. Telephone contact was attempted for 412 patients (8.2%); 389 (94.4%) were successfully contacted and survival status for the remaining 23 patients (0.5% of the cohort) was confirmed through national death registry linkage. No patient was missing 30-day mortality status. Clinical outcomes Thirty-day all-cause mortality was 9.2% (n=460). Most deaths occurred during the index hospitalisation; 58% were attributed to progression of underlying comorbid conditions or multiorgan failure and 42% to uncontrolled bleeding or early rebleeding despite intervention. Overall, 71.0% of patients (n=3550) required at least one therapeutic intervention. These included transfusion of ≥1 unit of packed red blood cells (52.0%, n=2600), endoscopic haemostasis (31.0%, n=1550), angiographic embolisation (5.0%, n=250) and emergency surgery (2.0%, n=100). Because some patients received more than one type of intervention, the sum of individual intervention categories exceeds the number of unique patients requiring intervention. Among transfused patients, the median number of units transfused was 2 (IQR 1–4). Rebleeding within 30 days occurred in 14.5% (n=725), with 68.0% (n=493) occurring within the first 7 days. Patients with rebleeding had higher 30-day mortality than those without rebleeding (27% vs 6%, p<0.001), higher ICU utilisation (41% vs 9%, p<0.001) and longer hospital stay (median 12 days (IQR 7–18) vs 4 days (IQR 3–6), p<0.001). Performance of established pre-endoscopic risk scores Discriminatory performance of the five established scores is shown in table 2. ROC curves for 30-day mortality and need for therapeutic intervention are shown in figure 2A,B. Receiver operating characteristic (ROC) curves. (A) ROC curves for prediction of 30-day all-cause mortality. (B) ROC curves for prediction of need for therapeutic intervention. Curves are shown for Acute Gastrointestinal Bleeding ABC-GBS Triage Evaluation (AGATE), Age-Blood tests-Comorbidities (ABC), Mental status-ASA class-Pulse-Albumin-Systolic blood pressure-Haemoglobin (MAP(ASH)), Haematemesis-Heart rate-Haemoglobin-Blood pressure-Blood urea nitrogen (H3B2), Glasgow-Blatchford score (GBS) and Albumin, INR, Mental status, Systolic blood pressure, Age ≥65 (AIMS65). Areas under the receiver operating characteristic curves with 95% CIs are reported in table 2. AUC, area under the curve. Table 2Discriminatory performance (AUROC with 95% CI)Score30-day mortalityNeed for intervention30-day rebleedingComposite end point*AGATE0.85 (0.82 to 0.87)0.80 (0.77 to 0.82)0.65 (0.61 to 0.68)0.86 (0.83 to 0.88)ABC0.83 (0.80 to 0.86)0.73 (0.69 to 0.77)0.60 (0.56 to 0.64)0.79 (0.76 to 0.82)MAP(ASH)0.81 (0.78 to 0.84)0.77 (0.73 to 0.80)0.66 (0.62 to 0.70)0.81 (0.78 to 0.84)H3B20.78 (0.74 to 0.82)0.75 (0.71 to 0.79)0.64 (0.60 to 0.68)0.78 (0.75 to 0.81)AIMS650.75 (0.71 to 0.79)0.69 (0.65 to 0.73)0.60 (0.56 to 0.64)0.73 (0.70 to 0.76)GBS0.70 (0.66 to 0.74)0.78 (0.74 to 0.82)0.62 (0.58 to 0.66)0.77 (0.74 to 0.80)*Composite end point: 30-day mortality or therapeutic intervention.ABC, Age-Blood tests-Comorbidities; AGATE, Acute Gastrointestinal Bleeding ABC-GBS Triage Evaluation; AIMS65, Albumin, INR, Mental status, Systolic blood pressure, Age ≥65; GBS, Glasgow-Blatchford score; H3B2, Haematemesis-Heart rate-Haemoglobin-Blood pressure-Blood urea nitrogen; MAP(ASH), Mental status-ASA class-Pulse-Albumin-Systolic blood pressure-Haemoglobin. Discriminatory performance (AUROC with 95% CI) Composite end point: 30-day mortality or therapeutic intervention. ABC, Age-Blood tests-Comorbidities; AGATE, Acute Gastrointestinal Bleeding ABC-GBS Triage Evaluation; AIMS65, Albumin, INR, Mental status, Systolic blood pressure, Age ≥65; GBS, Glasgow-Blatchford score; H3B2, Haematemesis-Heart rate-Haemoglobin-Blood pressure-Blood urea nitrogen; MAP(ASH), Mental status-ASA class-Pulse-Albumin-Systolic blood pressure-Haemoglobin. For 30-day mortality, ABC showed the highest discrimination (AUROC 0.83, 95% CI 0.80 to 0.86), followed by MAP(ASH) (AUROC 0.81, 95% CI 0.78 to 0.84). H3B2, AIMS65 and GBS yielded AUROCs of 0.78 (95% CI 0.74 to 0.82), 0.75 (95% CI 0.71 to 0.79) and 0.70 (95% CI 0.66 to 0.74), respectively. ABC and MAP(ASH) both outperformed GBS (p<0.01), while no significant difference was observed between ABC and MAP(ASH) (p=0.47). For therapeutic intervention, GBS and MAP(ASH) demonstrated the strongest discrimination (AUROC 0.78, 95% CI 0.74 to 0.82 and 0.77, 95% CI 0.73 to 0.80, respectively), followed by H3B2 (0.75, 95% CI 0.71 to 0.79), ABC (0.73, 95% CI 0.69 to 0.77) and AIMS65 (0.69, 95% CI 0.65 to 0.73). GBS, MAP(ASH) and H3B2 performed significantly better than AIMS65 (all p<0.05). All evaluated scores demonstrated modest discrimination for 30-day rebleeding (AUROC range 0.60–0.66), with reduced discrimination in patients aged ≥65 years (AUROC range 0.55–0.59). Identification of low-risk patients by established scores Performance at predefined low-risk thresholds is summarised in table 3. Table 3Performance of established scores at predefined low-risk thresholdsScoreThresholdLow-risk patients, n (%)30-day mortality, n (%)Intervention, n (%)Endoscopic therapy, n (%)Transfusion only, n (%)ABC≤32150 (43%)17 (0.8%)45 (2.1%)12 (0.6%)33 (1.5%)MAP(ASH)≤22050 (41%)2 (0.1%)31 (1.5%)8 (0.4%)23 (1.1%)GBS≤1750 (15%)0 (0%)0 (0%)00AIMS650450 (9%)3 (0.7%)8 (1.8%)2 (0.4%)6 (1.3%)H3B2≤1950 (19%)0 (0%)18 (1.9%)5 (0.5%)13 (1.4%)ABC, Age-Blood tests-Comorbidities; AIMS65, Albumin, INR, Mental status, Systolic blood pressure, Age ≥65; GBS, Glasgow-Blatchford score; H3B2, Haematemesis-Heart rate-Haemoglobin-Blood pressure-Blood urea nitrogen; MAP(ASH), Mental status-ASA class-Pulse-Albumin-Systolic blood pressure-Haemoglobin. Performance of established scores at predefined low-risk thresholds ABC, Age-Blood tests-Comorbidities; AIMS65, Albumin, INR, Mental status, Systolic blood pressure, Age ≥65; GBS, Glasgow-Blatchford score; H3B2, Haematemesis-Heart rate-Haemoglobin-Blood pressure-Blood urea nitrogen; MAP(ASH), Mental status-ASA class-Pulse-Albumin-Systolic blood pressure-Haemoglobin. ABC ≤3 classified 43% (n=2150) as low risk, with 30-day mortality of 0.8% (n=17) and intervention of 2.1% (n=45). MAP(ASH) ≤2 classified 41% (n=2050) as low risk, with mortality of 0.1% (n=2) and intervention of 1.5% (n=31). GBS ≤1 classified 15% (n=750), with no deaths and no therapeutic interventions. AIMS65=0 identified 9% (n=450) with mortality of 0.7% (n=3) and intervention of 1.8% (n=8). H3B2 ≤1 classified 19% (n=950), with no deaths and intervention of 1.9% (n=18). Derivation and performance of the AGATE score AGATE was derived using multivariable logistic regression for the composite end point of 30-day mortality or therapeutic intervention. Final model coefficients and point assignments are shown in table 4. A simplified bedside scoring table to facilitate clinical application of AGATE is provided in online supplemental table 2. Table 4Final multivariable logistic regression model for AGATEPredictorβ-CoefficientAdjusted OR (95% CI)P valuePointsLiver cirrhosis1.173.22 (2.71 to 3.89)<0.0014Albumin <3.0 g/dL1.032.80 (2.38 to 3.28)<0.0013Comorbidity burden ≥20.962.61 (2.22 to 3.08)<0.0013SBP <100 mm Hg0.922.50 (2.13 to 2.95)<0.0013Haemoglobin (per 10 g/L decrease)0.691.99 (1.76 to 2.24)<0.0012BUN (per 10 mg/dL increase)0.661.94 (1.71 to 2.19)<0.0012HR ≥100 /min0.581.79 (1.56 to 2.04)<0.0012Haematemesis0.531.70 (1.47 to 1.96)<0.0012Syncope0.511.66 (1.42 to 1.93)<0.0012Creatinine (per 0.5 mg/dL increase)0.411.52 (1.32 to 1.74)<0.0011Age (per decade increase)0.381.46 (1.31 to 1.63)<0.0011Total score range: 0–25.AGATE, Acute Gastrointestinal Bleeding ABC-GBS Triage Evaluation; BUN, blood urea nitrogen; HR, heart rate; SBP, systolic blood pressure. Final multivariable logistic regression model for AGATE Total score range: 0–25. AGATE, Acute Gastrointestinal Bleeding ABC-GBS Triage Evaluation; BUN, blood urea nitrogen; HR, heart rate; SBP, systolic blood pressure. Discrimination For 30-day mortality, AGATE achieved an AUROC of 0.85 (95% CI 0.82 to 0.87), significantly higher than GBS (p<0.001) and AIMS65 (p<0.001) and numerically higher than ABC (p=0.12) and MAP(ASH) (p=0.08). For therapeutic intervention, AGATE achieved an AUROC of 0.80 (95% CI 0.77 to 0.82), higher than AIMS65 (p<0.001) and ABC (p=0.01), with no statistically significant difference compared with GBS (p=0.34) or MAP(ASH) (p=0.29). For the composite end point used for model derivation, AGATE achieved an AUROC of 0.86 (95% CI 0.83 to 0.88), compared with GBS (AUROC 0.77, 95% CI 0.74 to 0.80; DeLong p<0.001). For 30-day rebleeding, AGATE showed modest discrimination (AUROC 0.65, 95% CI 0.61 to 0.68), comparable to MAP(ASH) (p=0.71). Internal validation Bootstrap internal validation (1000 resamples) yielded an optimism-corrected AUROC of 0.843 for the composite end point. The bootstrap-corrected calibration slope was 0.97 and calibration showed no evidence of lack of fit (Hosmer-Lemeshow p=0.34). Sensitivity analysis excluding transfusion When therapeutic intervention was restricted to endoscopic, angiographic or surgical therapy (excluding transfusion), AGATE maintained discrimination (AUROC 0.78, 95% CI 0.75 to 0.81), with no statistically significant difference compared with GBS (AUROC 0.77, 95% CI 0.74 to 0.80; p=0.52). Risk stratification Using predefined cut-offs, AGATE stratified patients into low-risk (≤5), intermediate-risk610 and high-risk (≥11) groups (table 5). Table 5AGATE risk categories and outcomesRisk categoryScore rangePatients, n (%)30-day mortality, n (%)Intervention, n (%)Intervention type (low-risk only)NPV for mortality (95% CI)Low≤51900 (38%)6 (0.3%)76 (4.0%)62 transfusion only (82%)14 endoscopic (18%)0 surgery/embolisation99.7% (99.3 to 99.9%)Intermediate6–102250 (45%)115 (5.1%)1755 (78%)——High≥11850 (17%)270 (31.8%)816 (96%)——AGATE, Acute Gastrointestinal Bleeding ABC-GBS Triage Evaluation; NPV, negative predictive value. AGATE risk categories and outcomes AGATE, Acute Gastrointestinal Bleeding ABC-GBS Triage Evaluation; NPV, negative predictive value. Low-risk patients comprised 38% (n=1900), with 30-day mortality of 0.3% (n=6) and intervention of 4.0% (n=76). Among these 76 patients, 62 (82%) received transfusion only and 14 (18%) underwent endoscopic therapy; none required surgery or embolisation and no deaths occurred in this subgroup. Intermediate-risk patients comprised 45% (n=2250), with mortality of 5.1% (n=115) and intervention of 78% (n=1755). High-risk patients comprised 17% (n=850), with mortality of 31.8% (n=270) and intervention of 96% (n=816). Predictive values for predefined AGATE thresholds are presented in online supplemental table 3. Decision-curve analysis Decision-curve analyses for 30-day mortality and therapeutic intervention are shown in figure 3A,B. Across clinically relevant threshold probabilities (0%–40% for mortality; 0%–60% for intervention), AGATE demonstrated the highest net benefit compared with established scores and treat-all/treat-none strategies. Decision-curve analysis. (A) Decision-curve analysis for 30-day mortality. (B) Decision-curve analysis for need for therapeutic intervention. The x-axis represents threshold probability and the y-axis represents net benefit. Curves are shown for AGATE and the five established risk scores. The light-grey dashed line represents a ‘treat-all’ strategy and the dark-grey horizontal line represents a ‘treat-none’ strategy. ABC, Age-Blood tests-Comorbidities; AGATE, Acute Gastrointestinal Bleeding ABC-GBS Triage Evaluation; AIMS65, Albumin, INR, Mental status, Systolic blood pressure, Age ≥65; GBS, Glasgow-Blatchford score; H3B2, Haematemesis-Heart rate-Haemoglobin-Blood pressure-Blood urea nitrogen; MAP(ASH), Mental status-ASA class-Pulse-Albumin-Systolic blood pressure-Haemoglobin. Discussion Principal findings In this large single-centre cohort of 5000 patients with AUGIB, we performed a comprehensive head-to-head comparison of five widely used pre-endoscopic risk scores and derived a hybrid score, AGATE, designed to integrate the complementary strengths of ABC (mortality prediction) and GBS (low-risk identification). Our findings support the differential performance profiles of existing models and suggest that a unified score can provide balanced discrimination across clinically relevant end points. Comparative performance of established scores Consistent with prior multicentre studies,10 13 ABC and MAP(ASH) demonstrated the strongest discrimination for 30-day mortality (AUROC 0.83 and 0.81, respectively). The performance of ABC likely reflects its integration of acute physiological markers (haemoglobin, urea) with indicators of baseline vulnerability (age, albumin, comorbidity burden).10 In our cohort, the ABC low-risk threshold (≤3) classified 43% of patients as low risk, with mortality below 1%, supporting its potential utility in discharge pathways.8 MAP(ASH), which similarly incorporates acute and chronic risk factors, also performed well across outcomes, with decision-curve analysis supporting its clinical utility.13 As expected, GBS excelled in identifying patients at very low risk of requiring therapeutic intervention. The established cut-off of ≤1 identified 15% of patients with no observed mortality or therapeutic interventions, reaffirming its value for safe discharge decisions.4 9 However, its more limited discrimination for mortality (AUROC 0.70) reflects the fact that mortality prediction was not its primary design objective.3 H3B2 demonstrated reasonable discrimination for intervention but performed less well for mortality, particularly among older patients, likely reflecting the absence of age-related and comorbidity-related variables.14 AIMS65 showed the weakest overall performance, supporting previous observations that simplicity may come at the expense of reduced prognostic granularity.12 The AGATE score: rationale and performance The divergent strengths of existing scores create practical challenges in emergency settings, where clinicians may need to calculate and interpret multiple tools simultaneously.15 AGATE was developed to address this limitation by combining predictors derived from ABC and GBS within a single framework. For 30-day mortality, AGATE achieved an AUROC of 0.85, higher than GBS and AIMS65 and numerically higher than ABC and MAP(ASH), although these differences were not statistically significant. For prediction of therapeutic intervention, AGATE (AUROC 0.80) demonstrated discrimination comparable to GBS (0.78) and superior to AIMS65 and ABC. Importantly, when evaluated against the composite end point of mortality or intervention—aligned with the original design objective of GBS4—AGATE demonstrated higher discrimination than GBS (AUROC 0.86 vs 0.77; p<0.001). However, this finding should be interpreted cautiously, as derivation and validation were performed within the same dataset. Clinical utility and risk stratification At the predefined low-risk threshold (≤5), AGATE classified 38% of patients as low risk with 0.3% 30-day mortality. Although 4% required therapeutic intervention, most of these interventions consisted of transfusion alone (82%) and no patient required surgery, embolisation or died within 30 days. These findings indicate that expansion of the low-risk group was not associated with excess short-term mortality in this cohort. Intermediate-risk and high-risk categories demonstrated progressively increasing mortality (5.1% and 31.8%) and intervention rates (78% and 96%), supporting their potential relevance for risk-based resource allocation and prioritisation of urgent endoscopy. Subgroup performance and generalisability All evaluated scores, including AGATE, showed modest discrimination for rebleeding, particularly among patients aged ≥65 years (AUROC range 0.55–0.59). This likely reflects the multifactorial and dynamic nature of rebleeding, which may depend on evolving clinical and endoscopic factors not captured by baseline variables.15 Future models incorporating time-dependent or procedure-related characteristics may improve prediction in this domain. The proportion of variceal bleeding (21%) was higher than reported in many general cohorts and reflects the tertiary hepatology referral profile of our centre, which includes a high-volume liver transplantation programme. This case-mix may influence model performance and limits generalisability to non-tertiary settings.28 Comparison with emerging machine learning models Recent studies have explored machine learning (ML) approaches for risk stratification in AUGIB, with some reporting high discrimination for mortality and need for intervention.29 30 However, ML models typically require larger variable sets, external computational infrastructure and reduced interpretability, which may limit bedside applicability. AGATE was intentionally derived from routinely available clinical and laboratory variables to preserve transparency and clinical usability. Whether ML approaches provide meaningful incremental benefit over parsimonious regression-based models in external validation cohorts remains to be determined. Strengths and limitations Strengths of this study include a large consecutive cohort, complete 30-day mortality ascertainment, direct head-to-head comparison of major pre-endoscopic scores, adherence to TRIPOD+AI reporting standards, internal validation using bootstrap resampling and a planned sensitivity analysis excluding transfusion to assess the potential influence of local practice patterns. Several limitations should be acknowledged. First, the retrospective single-centre design limits generalisability. External validation in independent multicentre cohorts is required. Second, derivation and validation were performed within the same dataset; although optimism-corrected performance was estimated, discrimination may attenuate in external settings. Third, the composite end point included transfusion, which may reflect local transfusion thresholds. Restrictive transfusion strategies have been shown to influence outcomes in AUGIB, underscoring the importance of considering transfusion practices when interpreting composite end points.31 Sensitivity analysis excluding transfusion yielded similar discrimination (AGATE AUROC 0.78 vs GBS 0.77), suggesting that performance was not solely driven by transfusion practices. Fourth, inclusion was restricted to patients with endoscopically confirmed bleeding. While necessary for consistent outcome adjudication, this limits applicability to patients with suspected AUGIB who do not undergo endoscopy. Fifth, comorbidity burden was defined using selected routinely recorded conditions. Although these were chosen based on prior literature and inclusion in established scores, alternative indices such as the Charlson Comorbidity Index may yield different results. Sixth, because risk scores were calculated retrospectively, clinicians were not blinded to them during care. Although this may introduce management-related bias, findings were broadly consistent with prior prospective evaluations.3 10 Finally, the study period included the COVID-19 pandemic. Although no temporal shift in mortality or intervention rates was observed in sensitivity analyses, pandemic-related effects on healthcare delivery cannot be entirely excluded. Conclusion In this large retrospective cohort of 5000 patients with endoscopically confirmed AUGIB, established pre-endoscopic scores demonstrated complementary strengths: ABC and MAP(ASH) showed the strongest discrimination for mortality, whereas GBS remained effective for identifying very low-risk patients. The AGATE hybrid score integrated predictors from ABC and GBS and achieved balanced discrimination for mortality and therapeutic intervention, demonstrating higher discrimination than GBS for the composite end point within this cohort. AGATE identified a substantially larger low-risk group than GBS ≤1 while maintaining very low short-term mortality. Before routine clinical implementation, prospective external validation in independent multicentre cohorts with varying case-mix profiles and clinical management strategies is required. Supplementary material