Association of an age-albumin-red cell distribution width score with 28-day in-hospital mortality in emergency department patients with suspected sepsis

Abstract

This study evaluates the performance of a simple age-albumin-red cell distribution width (AAR) score for stratifying 28-day in-hospital mortality risk in an operational cohort of adult emergency department (ED) patients with suspected sepsis and in a later-period subgroup. This single-center retrospective cohort study analyzed an operational ED suspected-sepsis cohort of 852 adults defined by a modified early warning score of 5 or higher plus an infection-related ED discharge diagnosis from 2022 to 2025. The AAR score assigned one point each for age ≥ 75 years, albumin < 3.0 g/dL, and red cell distribution width ≥ 14.5%; scores ≥ 2 indicated high risk. Performance was assessed using the area under the receiver operating characteristic curve (AUROC), classification metrics, calibration, and exploratory clinical net benefit overall and across chronological study periods. Overall, 98 patients (11.5%) died within 28 days. The AAR AUROC was 0.787 (95% confidence interval [CI]: 0.740–0.833) overall and 0.778 (95% CI: 0.703–0.852) in 2024 to 2025; paired comparison did not demonstrate greater discrimination than continuous albumin alone (difference, 0.013; 95% CI: −0.074–0.100; P = .770). However, mortality increased stepwise across AAR categories (2.2%, 7.4%, 21.7%, and 49.2%), and this gradient was preserved in the later-period cohort (1.4%, 6.6%, 15.9%, and 41.4%). The earlier-period equation overpredicted absolute mortality risk in the later period. The AAR score combined 3 routinely available variables into 4 ordinal groups with a consistent mortality gradient across study periods. Its simple 0 to 3-point structure supports further external evaluation as an adjunct to early ED risk stratification. Patient-level probability estimation and decision-support use require additional validation and recalibration.

1. Introduction

Sepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection.^[1] It remains a leading cause of morbidity and mortality worldwide, with an estimated 48.9 million incident cases and 11.0 million deaths reported globally in 2017.^[2] In the emergency department (ED), timely risk stratification is essential to guide decisions regarding monitoring intensity, intensive care unit admission, and initiation of aggressive treatment. Because mortality rates in critically ill patients range from 10 to 40%, early identification of high-risk patients allows clinicians to allocate resources appropriately and improve outcomes through time-sensitive interventions.^[1,3]

Several scoring systems have been developed to support the prognostic assessment of sepsis; however, each has notable limitations in ED settings. The quick SOFA (qSOFA) score is simple to apply but lacks sufficient sensitivity as a primary screening tool.^[4] In a meta-analysis, qSOFA had higher specificity but significantly lower sensitivity than systemic inflammatory response syndrome criteria for predicting in-hospital mortality.^[5,6] The full sequential organ failure assessment (SOFA) score provides a comprehensive assessment of organ dysfunction, but its complexity and requirement for multiple physiological parameters limit its immediate use in early ED settings.^[7] Other tools, such as the modified early warning score (MEWS), National Early Warning Score, Acute Physiology and Chronic Health Evaluation II, and Simplified Acute Physiology Score II, incorporate many variables that further constrain bedside usability in time-sensitive environments.^[8,9] Individual biomarkers, such as lactate, albumin, red blood cell distribution width (RDW), blood urea nitrogen (BUN), and the BUN-to-albumin ratio (BAR), have been associated with sepsis prognosis in previous studies.

Assessing baseline vulnerability, acute inflammatory responses, and sepsis-induced functional decline is important for estimating disease severity. Age, albumin, and RDW are useful markers that reflect these dimensions in a simple and clinically accessible manner. Age reflects host vulnerability, immune senescence, reduced physiological reserve, and comorbidity burden, all of which are associated with increased sepsis incidence and mortality in older patients.^[10] Albumin reflects systemic inflammation, capillary leakage, nutritional status, and physiological reserve during acute illness. Low serum albumin is associated with increased mortality risk in patients with severe sepsis.^[11] RDW, widely recognized as a marker of systemic inflammation, oxidative stress, bone marrow responsiveness, and chronic disease burden,^[12] is independently associated with mortality in critically ill patients with sepsis.^[13] Together, these 3 readily available variables may reflect host vulnerability, inflammatory and nutritional status, and hematologic stress.

The objectives of this study were to evaluate the age-albumin-red cell distribution width (AAR) score for stratifying 28-day in-hospital mortality risk in an operational cohort of adult ED patients with suspected sepsis, to compare its performance with established clinical scores and routine biomarkers, and to assess its performance in a later-period cohort.

2. Methods

2.1. Study design and setting

This single-center retrospective study was conducted at a regional emergency medical center that treats approximately 40,000 patients annually. We collected data on age, sex, initial vital signs, laboratory results, vasopressor use, ED discharge diagnosis, length of hospital stay, and discharge and mortality status. Adults meeting the operational ED suspected-sepsis definition were included. This study aimed to assess whether the AAR score, a simple ordinal risk score calculated from early ED data, could stratify patients into groups with differing risks of 28-day in-hospital mortality. This study was approved by the Institutional Review Board (IRB) of Ewha Womans University Mokdong Hospital (IRB No. 2026-06-016) and used data from electronic medical records. The requirement for informed consent was waived by the IRB because of the retrospective design and use of electronic medical record data.

2.2. Study population

The study population included adult patients aged 19 years or older with suspected or confirmed sepsis who presented to the ED between January 1, 2022, and December 31, 2025. Operationally, suspected sepsis in the ED was defined by a MEWS of 5 or higher and an infection-related ED discharge diagnosis. This approach was intended to identify acutely ill adult ED patients screened as being at risk for sepsis using routinely available clinical data, consistent with contemporary screening-oriented recommendations,^[14] rather than by prospectively adjudicating sepsis-3-confirmed sepsis.^[1] Infection-related diagnoses were defined using the International Classification of Diseases, 10th revision. Patients younger than 19 years and those with missing albumin or RDW values were excluded.

This report was prepared in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology statement and applicable items from the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis + Artificial Intelligence statement.^[15,16]

No a priori sample-size calculation was performed; all eligible records during the study period were included. Post hoc scenario analyses using pmsampsize and pmvalsampsize were conducted to contextualize sample-size adequacy under alternative assumptions about the effective number of candidate parameters and desired precision for model evaluation.^[17,18] These scenarios were not used to redefine eligibility and were not interpreted as proof of definitive sample-size adequacy.

2.3. Outcome definition

The outcome was death recorded at the index hospital within 28 days of admission. Patients who did not meet this definition, including those discharged or transferred alive and those who died after day 28, were classified as nonevents. All transfers occurred on or after day 28.

2.4. AAR score definition

The AAR score comprised 3 components: age, albumin, and RDW. Candidate score combinations and cutoffs were explored using the full dataset. The final AAR configuration was retained after exploratory performance assessment, with emphasis on clinical interpretability. One point was assigned to each of the following: age ≥ 75 years, albumin < 3.0 g/dL, and RDW ≥ 14.5%. Age ≥ 75 years was selected to represent a clinically vulnerable older population. Albumin < 3.0 g/dL was selected as a conservative threshold for definite hypoalbuminemia, and RDW ≥ 14.5% was selected based on the upper limit of the local laboratory reference range. In this study, the AAR score was evaluated as an ordinal, population-level risk-stratification tool based on differences in observed event rates across score categories, rather than as a model for estimating a calibrated mortality probability for an individual patient or as a stand-alone clinical decision rule.

2.5. Comparison with existing screening tools

The AAR score was compared with qSOFA, MEWS, SOFA, and laboratory markers, including albumin, RDW, BUN, the BAR, chloride, lactate, arterial pH, and the lactate-to-albumin ratio (LAR). Scores were calculated using the initial ED vital signs and laboratory values available during the early ED evaluation. No missing values were imputed. Patients with missing albumin or RDW were excluded from the AAR analytic cohort, whereas other variables with missing observations were analyzed on an available-case basis. Invalid negative sentinel values for body temperature and oxygen saturation were treated as missing, and variable-specific availability was reported.

2.6. Later-period performance assessment

For period-specific performance assessment, the cohort was divided chronologically into an earlier-period cohort (2022–2023, n = 432) and a later-period cohort (2024–2025, n = 420). AAR discrimination and cutoff-based measures were calculated separately for each period. Because candidate score combinations and chronological splits had been explored using the full dataset, the 2024 to 2025 analysis was interpreted as a later-period performance assessment rather than independent temporal validation.

2.7. Statistical analysis

Baseline characteristics were compared between survivors and patients with 28-day in-hospital mortality. Continuous variables are presented as medians and interquartile ranges and were compared using the Wilcoxon rank-sum test. Categorical variables are presented as n(%) and were compared using the chi-square test or Fisher exact test. We calculated 28-day in-hospital mortality rate for each AAR score and used logistic regression to estimate the odds ratio (OR) per one-point increase in the AAR score. An AAR score of 2 or higher was defined as high risk, and the OR for the high-risk group was calculated. Discrimination was compared using the area under the receiver operating characteristic curve (AUROC). We compared the AUROC of the AAR score with those of existing predictors; 95% confidence intervals (CIs) were calculated using the DeLong method. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated for each predictor. The AUROC, sensitivity, specificity, PPV, NPV, and OR were obtained separately in the 2022 to 2023 and 2024 to 2025 cohorts. Wilson 95% CIs were calculated for the classification metrics. Paired DeLong tests compared the AUROC of the AAR score with continuous albumin overall and in the later-period cohort. We examined alternative cutoffs, including albumin ≤ 3.2 g/dL, RDW ≥ 15%, and age ≥ 70 years, as exploratory sensitivity analyses. All analyses were performed using R (version 4.5.1; R Foundation for Statistical Computing, Vienna, Austria), and a 2-sided P value < .05 was considered statistically significant.

For chronological model evaluation, 5 logistic models were fitted in the 2022 to 2023 cohort and applied to the 2024 to 2025 cohort without refitting: the ordinal AAR score; albumin as a continuous variable; albumin < 3.0 g/dL as a binary variable; the 3 binary AAR components with separately estimated coefficients; and age, albumin, and RDW as continuous variables. Paired DeLong tests compared AUROCs in the same later-period patients. Calibration-in-the-large, calibration slope, Brier score, and score-specific observed versus predicted risks were assessed with the earlier-period equations held fixed. Percentile 95% CIs were obtained from 2000 bootstrap resamples of the later-period cohort.

Decision curve analysis was performed over threshold probabilities of 5 to 30% to compare exploratory net benefit, with paired differences and 95% CIs estimated from 2000 bootstrap resamples.^[19] Apparent full-cohort performance of the final ordinal AAR model was also corrected for model-refitting optimism using 2000 bootstrap samples^[20]; this procedure did not repeat the preceding variable and cutoff search and therefore did not correct selection optimism.

3. Results

3.1. Study population

Among 855 potentially eligible records, 1 patient younger than 19 years, 1 patient with missing albumin, and 1 patient with missing RDW after the albumin exclusion were removed. The final analytic cohort comprised 852 patients. Of these, 754 survived and 98 (11.5%) died in the hospital within 28 days. Period-specific analyses included 432 patients treated in 2022 to 2023 and 420 treated in 2024 to 2025. The study flow and chronological evaluation structure are shown in Figure 1.

Figure 1.

Study flow and chronological evaluation structure. The cohort was defined operationally using a modified early warning score of 5 or higher plus an infection-related emergency department discharge diagnosis. AAR = age-albumin-red cell distribution width, ED = emergency department, RDW = red blood cell distribution width.

The baseline characteristics of the patients are shown in Table 1. Patients who died were older than survivors and had lower albumin levels, higher RDW, and higher BUN, BAR, and lactate levels. Arterial pH was lower in the mortality group. These differences were consistent with greater baseline vulnerability, nutritional and inflammatory derangement, and physiologic severity in patients who died.

Table 1.

Baseline characteristics by 28-day mortality.

Characteristic	Overall (n = 852)	Non-mortality (n = 754)	28-day mortality (n = 98)	P value
Age, yrs	72 (61–82)	70 (59–81)	82 (75–87)	<.001
Systolic blood pressure, mm Hg	120 (97–144)	121 (99–145)	114 (89–139)	.085
Diastolic blood pressure, mm Hg	70 (57–81)	70 (58–81)	64 (52–76)	.015
Heart rate, /min	123 (111–134)	123 (111–135)	123 (108–133)	.482
Respiratory rate, /min	23 (21–26)	23 (21–25)	25 (22–28)	<.001
Body temperature, °C	38.8 (37.8–39.4)	38.9 (38.1–39.4)	37.7 (36.7–38.7)	<.001
Oxygen saturation, %	96 (93–98)	96 (93–98)	93 (87–96)	<.001
qSOFA	1 (1–2)	1 (1–2)	2 (1–2)	<.001
MEWS	6 (5–7)	6 (5–7)	6 (5–7)	.423
SOFA	2 (1–3)	2 (0–3)	3 (1–5)	<.001
Creatinine, mg/dL	1.06 (0.79–1.52)	1.04 (0.79–1.45)	1.25 (0.87–2.38)	<.001
Platelet, ×103/µL	183 (129–258)	183 (130–253)	175 (114–282)	.935
Total bilirubin, mg/dL	0.7 (0.5–1.1)	0.7 (0.5–1.1)	0.7 (0.5–1.1)	.388
Albumin, g/dL	3.4 (2.9–3.8)	3.5 (3.0–3.8)	2.7 (2.3–3.2)	<.001
BUN, mg/dL	21 (14–33)	20 (14–31)	34 (20–51)	<.001
BUN/albumin ratio	6.2 (4.1–10.9)	5.8 (3.9–9.7)	12.4 (7.7–21.6)	<.001
Chloride, mmol/L	100 (97–104)	100 (97–104)	100 (96–105)	.952
RDW, %	13.4 (12.7–14.7)	13.3 (12.6–14.6)	14.6 (13.4–16.3)	<.001
Lactate, mg/dL	23.0 (16.0–35.0)	22.0 (16.0–33.0)	29.0 (19.0–58.0)	<.001
Arterial pH	7.43 (7.38–7.47)	7.43 (7.38–7.47)	7.41 (7.26–7.46)	<.001
Arterial PaO2, mm Hg	67 (49–87)	66 (47–85)	72 (57–100)	.003
Lactate/albumin ratio	6.97 (4.88–11.00)	6.67 (4.74–10.33)	10.00 (6.82–21.67)	<.001
Male sex	450/852 (52.8%)	397/754 (52.7%)	53/98 (54.1%)	.790
Alert mental status	635/852 (74.5%)	589/754 (78.1%)	46/98 (46.9%)	<.001
KTAS level 1–2	498/852 (58.5%)	414/754 (54.9%)	84/98 (85.7%)	<.001
ICU admission from ED	391/852 (45.9%)	332/754 (44.0%)	59/98 (60.2%)	.003
Inotrope use	250/852 (29.3%)	213/754 (28.2%)	37/98 (37.8%)	.052

Note. Values are median (interquartile range) or n/N (%). Body temperature was available in 842 patients, oxygen saturation in 678, total bilirubin in 824, lactate and the lactate-to-albumin ratio in 701, arterial pH in 750, and arterial PaO₂ in 749; all other listed variables were available in the full analytic cohort.

BUN = blood urea nitrogen, ED = emergency department, ICU = intensive care unit, KTAS = Korean Triage and Acuity Scale, MEWS = modified early warning score, PaO₂ = arterial partial pressure of oxygen, qSOFA = quick sequential organ failure assessment, RDW = red cell distribution width, SOFA = sequential organ failure assessment.

3.2. Distribution of the AAR score and mortality

Table 2 and Figure 2 show the distribution of the AAR score and corresponding 28-day in-hospital mortality. Mortality increased steadily as the score increased, and each one-point increase in the AAR score was associated with higher odds of 28-day in-hospital mortality.

Table 2.

Clinical outcomes according to the age-albumin-red cell distribution width score.

AAR score	N	28-day mortality, n (%)	Odds ratio (95% CI) vs score 0
0	277	6 (2.2%)	Reference
1	353	26 (7.4%)	3.59 (1.46–8.85)
2	157	34 (21.7%)	12.49 (5.11–30.51)
3	65	32 (49.2%)	43.80 (17.04–112.55)
Per 1-point increase	852		3.51 (2.72–4.53), P < .001
High risk (score ≥ 2)	222	66 (29.7%)	7.91 (5.00–12.49), P < .001

AAR = age-albumin-red cell distribution width, CI = confidence interval.

Figure 2.

Observed 28-day in-hospital mortality by AAR score. Bars show observed mortality; error bars show exact binomial 95% confidence intervals. Labels show percentages and event counts. AAR = age-albumin-red cell distribution width.

When an AAR score ≥ 2 was defined as high risk, mortality was higher in the high-risk group than in the low-risk group.

3.3. Predictive performance compared with existing scores and biomarkers

Table 3 shows the predictive performances of the AAR score and comparator predictors; Figure 3 shows the receiver operating characteristic curves for the 2024 to 2025 later-period cohort. In the overall cohort, the primary AAR score showed numerically higher discrimination than qSOFA, MEWS, SOFA, BAR, RDW, lactate, arterial pH, and LAR, whereas albumin alone had a similar AUROC. Paired DeLong comparisons did not show incremental discrimination over albumin alone overall (ΔAUROC, −0.001; 95% CI: −0.047–0.046; P = .982) or in the 2024 to 2025 cohort (ΔAUROC, 0.013; 95% CI: −0.074–0.100; P = .770).

Table 3.

Predictive performance of the age-albumin-red cell distribution width score and comparators.

Predictor	Overall AUROC (95% CI)	2024–2025 AUROC (95% CI)	Available, n/N	Cutoff	Sens.	Spec.
Primary AAR score (age ≥ 75 + albumin < 3.0 + RDW ≥ 14.5)	0.787 (0.740–0.833)	0.778 (0.703–0.852)	420/420	≥2	0.622	0.804
qSOFA	0.667 (0.613–0.721)	0.696 (0.611–0.781)	420/420	≥2	0.622	0.684
MEWS	0.523 (0.460–0.587)	0.540 (0.436–0.643)	420/420	≥8	0.243	0.906
SOFA	0.651 (0.591–0.711)	0.678 (0.576–0.781)	420/420	≥2	0.730	0.538
Albumin	0.787 (0.740–0.834)	0.765 (0.677–0.852)	420/420	<3.0	0.568	0.812
RDW	0.682 (0.627–0.737)	0.665 (0.577–0.753)	420/420	≥14.5	0.541	0.713
BUN	0.707 (0.652–0.762)	0.757 (0.666–0.847)	420/420	≥30	0.703	0.747
BAR	0.761 (0.712–0.811)	0.782 (0.691–0.873)	420/420	≥8	0.784	0.695
Abnormal chloride	0.566 (0.514–0.619)	0.509 (0.430–0.589)	420/420	<98 or >107	0.324	0.695
Lactate, mg/dL	0.629 (0.562–0.696)	0.636 (0.522–0.750)	312/420	≥18	0.812	0.339
Arterial pH	0.607 (0.538–0.677)	0.649 (0.536–0.763)	355/420	<7.35	0.444	0.865
Arterial PaO2, mm Hg	0.406 (0.347–0.464)	0.376 (0.280–0.471)	352/420	<60	0.306	0.566
LAR	0.713 (0.654–0.772)	0.719 (0.617–0.821)	312/420	≥7.5	0.750	0.589

Note. AUROC 95% confidence intervals were calculated using the DeLong method. Availability, cutoffs, sensitivity, and specificity refer to the 2024–2025 cohort.

AAR = age-albumin-red cell distribution width, AUROC = area under the receiver operating characteristic curve, BAR = blood urea nitrogen-to-albumin ratio, BUN = blood urea nitrogen, CI = confidence interval, LAR = lactate-to-albumin ratio, MEWS = modified early warning score, PaO₂ = arterial partial pressure of oxygen, qSOFA = quick sequential organ failure assessment, RDW = red cell distribution width, Sens. = sensitivity, SOFA = sequential organ failure assessment, Spec. = specificity.

Figure 3.

Receiver operating characteristic curves in the 2024–2025 later-period cohort. Curves compare the AAR score with selected clinical scores and laboratory biomarkers for 28-day in-hospital mortality. AAR = age-albumin-red cell distribution width, AUROC = area under the receiver operating characteristic curve, BAR = blood urea nitrogen-to-albumin ratio, qSOFA = quick sequential organ failure assessment, SOFA = sequential organ failure assessment.

In the 2024 to 2025 cohort, the AAR AUROC was 0.778 (95% CI: 0.703–0.852), and its point estimate was similar to the 0.791 (95% CI: 0.731–0.851) observed in 2022 to 2023. Comparator predictors showed a similar pattern to that in the overall cohort, while albumin alone had a similar AUROC.

3.4. Period-specific performance

Table 4 reports AAR performance in the 2022 to 2023 and 2024 to 2025 cohorts. AUROC point estimates and cutoff-based measures were similar across periods; detailed classification metrics are provided in Supplementary Table S1, Supplemental Digital Content 1. At the high-risk cutoff of AAR ≥ 2, PPV was 34.7% (95% CI: 26.9–43.4) in 2022 to 2023 and 23.5% (95% CI: 16.2–32.8) in 2024 to 2025; the corresponding NPVs were 94.2% (95% CI: 91.0–96.3) and 95.7% (95% CI: 92.8–97.4), respectively.

Table 4.

Period-specific performance of the age-albumin-red cell distribution width score.

Period, n (events)	AUROC (95% CI)	Sensitivity	Specificity	PPV	NPV	OR (95% CI)
2022–2023, 432 (61)	0.791 (0.731–0.851)	0.705	0.782	0.347	0.942	8.55 (4.68–15.63)
2024–2025, 420 (37)	0.778 (0.703–0.852)	0.622	0.804	0.235	0.957	6.75 (3.31–13.73)

Note. High risk was defined as an AAR score ≥ 2. AUROC 95% confidence intervals were calculated using the DeLong method. The 2024–2025 results represent a later-period internal performance assessment rather than independent temporal validation.

AAR = age-albumin-red cell distribution width, AUROC = area under the receiver operating characteristic curve, CI = confidence interval, NPV = negative predictive value, OR = odds ratio, PPV = positive predictive value.

In an exploratory post hoc probability mapping, estimates based on the 2022-2023 cohort overpredicted mortality in the 2024 to 2025 cohort (mean predicted risk, 12.7%; observed mortality, 8.8%). The calibration intercept was −0.484 (95% CI: −0.883 to −0.139), and the calibration slope was 0.935 (95% CI: 0.652–1.273). The AAR and continuous-albumin models had nearly identical later-period Brier scores (0.07328 and 0.07330, respectively; Supplementary Tables S2 and S3, Supplemental Digital Content 2 and Supplementary Fig. S1, Supplemental Digital Content 3).

3.5. Additional model evaluation

In 2024 to 2025, AAR discrimination exceeded that of binary albumin < 3.0 g/dL (difference in AUROC, 0.088; 95% CI: 0.025–0.151; P = .006) but did not exceed continuous albumin. A model retaining age, albumin, and RDW as continuous predictors had an AUROC of 0.830 (95% CI: 0.765–0.895), which was numerically higher than the AAR AUROC (difference, −0.052; 95% CI: −0.105–0.001; P = .054; Supplementary Tables S4–S6, Supplemental Digital Content 4). Across threshold probabilities of 5 to 30%, AAR did not show a consistent or statistically supported net-benefit advantage over continuous albumin.

For the final ordinal AAR model fitted in the full cohort, the apparent AUROC was 0.7867 and the optimism-corrected AUROC was 0.7863. Post hoc development-sample scenarios were satisfied when the model was treated as 1 ordinal term or 3 prespecified component terms, but not under broader exploratory assumptions of 14 or 22 effective candidate parameters. The later-period subgroup included 37 events, limiting precision for calibration assessment. Bootstrap, sample-size, and decision-curve details are provided in Supplementary Tables S7–S10, Supplemental Digital Content 5 and Supplementary Figure S2, Supplemental Digital Content 6.

3.6. Sensitivity analysis using alternative cutoffs

We applied alternative cutoff values for each component in the overall cohort as exploratory sensitivity analyses (Supplementary Table S11, Supplemental Digital Content 7). Overall discrimination changed only modestly across these alternatives.

4. Discussion

4.1. Summary of main findings

Observed 28-day in-hospital mortality increased stepwise across the 4 AAR categories, with similar gradients and AUROC estimates in both study periods. AAR did not outperform continuous albumin, but it offered an easily interpreted 0 to 3-point framework and discriminated better than albumin dichotomized at 3.0 g/dL in the later period. The earlier-period equation overpredicted absolute mortality in the later period, indicating that recalibration would be required for probability-based use. In the later period, the PPV was 23.5% and the NPV was 95.7% at an event prevalence of 8.8%. These values support interpreting the score as an ordinal risk-stratification aid rather than as an individual mortality prediction or rule-out threshold.

4.2. Clinical implications

The AAR score may serve as a simple adjunct to early ED risk assessment because its 3 components are routinely available and its 0 to 3-point structure conveys a graded mortality pattern. It should complement qSOFA, SOFA, and comprehensive clinical assessment. Its effect on clinical decisions and outcomes requires prospective evaluation.

4.3. Pathophysiological interpretation

Even when the infectious insult is similar, sepsis outcomes can differ substantially depending on baseline health status. Accordingly, several established severity and prognostic scores incorporate age, chronic disease, residential status, and functional dependence. For example, Acute Physiology and Chronic Health Evaluation II includes age and chronic health status, and the Mortality in ED Sepsis score includes variables such as terminal illness and nursing home residence.^[8,21] However, these variables often rely on patient or caregiver reports and can be partly subjective. In this study, age, available immediately in the ED, was selected as a simple marker of baseline vulnerability. Older age may worsen sepsis outcomes through immune senescence, greater comorbidity burden, frailty, reduced physiological reserve, and impaired tolerance to acute inflammatory stress.^[10] We used ≥75 years as the age threshold because it identifies a more clinically vulnerable older adult population and is consistent with severity-scoring frameworks in which ≥75 years represents the highest risk category.^[8]

In addition to age, albumin was used as a screening component. Albumin is influenced by chronic nutritional status, systemic inflammation, vascular permeability, hepatic synthesis, and the severity of acute illness. Hypoalbuminemia is associated with mortality in severe sepsis.^[11] Nutritional risk scores such as the Controlling Nutritional Status score also use albumin categories that differentiate normal albumin from progressively lower levels, with values below approximately 3.0 g/dL indicating more pronounced depletion.^[22] In our study, an albumin threshold of 3.2 g/dL showed the highest discriminative performance for 28-day in-hospital mortality. However, because this value was data-derived and may vary across cohorts, we selected albumin < 3.0 g/dL for the primary AAR score as a conservative and clinically recognizable threshold for definite hypoalbuminemia.

Finally, RDW was included to reflect the inflammatory and hematologic stress of sepsis. RDW is affected by inflammation, oxidative stress, impaired erythropoiesis, nutritional deficiency, and chronic disease burden,^[12] and elevated RDW has been associated with mortality in patients with sepsis.^[13] We used RDW ≥ 14.5%, which corresponded to the upper limit of the local laboratory reference range and was consistent with the direction of risk observed in our cohort. Combining this reference-based threshold with the study results strengthens the clinical interpretability of the AAR score.

4.4. Comparison with previous studies and predictors

Previous BMC Emergency Medicine studies evaluated RDW as a standalone prognostic marker in patients with suspected infection or sepsis and in frail older ED populations.^[23–25] Single markers such as lactate, albumin, RDW, BUN, and BAR have been reported to be associated with sepsis prognosis. In our cohort, albumin alone showed an AUROC similar to that of the AAR score. The AAR score provided a stepwise risk gradient across score categories and integrated age and RDW with albumin, thereby capturing host vulnerability, physiological reserve, inflammation, hematologic stress, and vascular permeability in a single bedside-friendly score. This stepwise risk gradient is a potential advantage, although incremental discrimination beyond albumin alone was not established.

In the later period, AAR discriminated better than albumin dichotomized at 3.0 g/dL but not better than albumin retained as a continuous variable. A three-variable continuous model also showed numerically higher discrimination than the equal-weight AAR score. These results illustrate the expected tradeoff: categorization and equal weighting improve transparency and ease of use but discard information. The potential value of AAR is therefore its reproducible ordinal risk gradient and pathophysiologic interpretability, not statistical superiority over simpler continuous biomarkers or more flexible models.

4.5. Limitations

This study has several limitations. First, its single-center retrospective design carries risks of selection and documentation bias and did not permit prospective assessment of implementation or clinical impact; therefore, prospective external validation in other centers is needed. Second, the outcome was in-hospital mortality rather than 28-day all-cause mortality; therefore, deaths occurring within 28 days after discharge may have been missed. Additionally, body temperature, oxygen saturation, total bilirubin, lactate, arterial pH, arterial partial pressure of oxygen, and the LAR had missing values and were analyzed on an available-case basis. The 2026 Surviving Sepsis Campaign guidelines recommend using the National Early Warning Score, National Early Warning Score 2, MEWS, or systemic inflammatory response syndrome rather than qSOFA alone as screening tools for sepsis in acutely ill hospitalized patients.^[14] Therefore, the use of MEWS in our cohort definition is consistent with contemporary screening-oriented practice. However, MEWS-based screening does not replace formal sepsis-3 adjudication; therefore, our findings should be interpreted as applying to an operational ED suspected-sepsis cohort. Because MEWS was part of the cohort definition, its comparative performance should be interpreted cautiously. Candidate score combinations, cutoffs, and chronological splits were explored using the full dataset; therefore, the 2024 to 2025 analysis was not an independent temporal validation. The exploratory probability mapping overpredicted absolute mortality risk in the later-period cohort and should not be used for individual risk estimation without independent evaluation and recalibration.

The later-period cohort contained only 37 deaths, resulting in substantial uncertainty around calibration and clinical-utility estimates. AAR did not improve discrimination or net benefit over continuous albumin, and clinically actionable risk thresholds were not prespecified. Finally, bootstrap optimism correction addressed refitting of the final ordinal model but did not repeat the earlier variable and cutoff search; selection optimism may therefore remain.

5. Conclusion

The AAR score showed a stepwise association with 28-day in-hospital mortality and similar AUROC point estimates across the 2 study periods. Its simple 0 to 3-point structure offers a framework for ordinal risk stratification using routinely available ED data. Incremental discrimination over continuous albumin was not demonstrated, and prospective external validation is needed to determine its clinical utility.

Author contributions

Conceptualization: Ji Weon Kang, Jae Hee Lee.

Data curation: Ji Weon Kang.

Formal analysis: Jae Hee Lee.

Investigation: Ji Weon Kang.

Methodology: Jae Hee Lee.

Project administration: Jae Hee Lee.

Supervision: Jae Hee Lee.

Validation: Jae Hee Lee.

Visualization: Ji Weon Kang, Jae Hee Lee.

Writing – original draft: Ji Weon Kang.

Writing – review & editing: Jae Hee Lee.