Log-transformed systemic inflammation response index improves prediction of 28-day ICU mortality in critically ill patients with cirrhosis and atrial fibrillation: development in MIMIC-IV and external validation in eICU-CRD

Abstract

Background

Short-term mortality is high among critically ill patients with cirrhosis and atrial fibrillation (AF). Existing scores such as SOFA, MELD-Na, and CHA₂DS₂-VASc do not directly reflect systemic inflammation. We investigated whether blood count-derived inflammatory indices could improve prediction of 28-day ICU mortality in this high-risk population.

Methods

Adult ICU patients with cirrhosis and AF were identified from MIMIC-IV v3.1 (single-center development cohort) and eICU-CRD v2.0 (multicenter external validation cohort). Five inflammatory indices, including SIRI, log-SIRI, SII, log-SII, and NLR, were compared by adding each marker to a base model of age, sex, SOFA, and MELD-Na. The final model was developed using multivariable logistic regression. Internal validation used 1000-resample bootstrap, and performance was assessed by discrimination, calibration, reclassification, and decision curve analysis.

Results

A total of 454 patients from MIMIC-IV and 121 from eICU-CRD were included. The 28-day ICU mortality rates were 19.6% and 16.5%, respectively. Among the five candidate markers, SII achieved the highest AUC (0.758) and lowest AIC, but its per-unit odds ratio (1.0001) made point-based nomogram construction impractical. log-SIRI was selected for its interpretable per-unit odds ratio (1.48) and near-identical discrimination to SII (ΔAUC − 0.007), together with the biological argument that its three-lineage formulation avoids platelet count, which is unreliable in cirrhosis. Adding log-SIRI to the base model increased the AUC from 0.721 to 0.751 (ΔAUC + 0.030, DeLong P = 0.026) and improved model fit on likelihood ratio testing (χ² = 9.18, P = 0.002), with concordant improvement in continuous NRI (0.241, P = 0.037) and calibration (Hosmer–Lemeshow P 0.002 to 0.729). The optimism-corrected C-index was 0.734. External validation in eICU-CRD yielded an AUC of 0.818.

Conclusion

Log-SIRI was associated with 28-day ICU mortality after adjustment for age, sex, SOFA, and MELD-Na. The improvement in AUC over the base clinical model was modest but statistically significant, corroborated by concordant improvements in likelihood ratio, continuous NRI, and calibration. Prospective external calibration is needed before individual clinical use.

Introduction

Liver cirrhosis remains a major source of morbidity worldwide. Its prevalence has been rising, with complications driven by hepatic decompensation, portal hypertension, and progressive immune dysfunction [1–3]. Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia [4]. In cirrhotic patients, it is reported more frequently than in age-matched populations without liver disease, with published series describing a prevalence of 5% to over 10%. When the two conditions coexist, several issues compound at once [4,5]. Coagulopathy makes anticoagulation difficult, even though thromboembolic risk does not disappear. Hemodynamic instability is more frequent, and altered hepatic metabolism often impairs the clearance of cardiac and anticoagulant medications [6,7].

Among patients with both cirrhosis and AF, ICU admission is associated with poor short-term prognosis [8,9]. Luo et al. reported that AF was independently associated with higher in-hospital and 4-year all-cause mortality in critically ill cirrhotic patients [8]. Darrat et al. linked AF in hospitalized cirrhotic patients to worse inpatient mortality, more stroke and acute kidney injury, and longer and more expensive admissions [9]. The clinical scores commonly used in this setting all have origins outside the cirrhosis-AF combination. The Sequential Organ Failure Assessment (SOFA) score quantifies organ dysfunction across six systems, which makes it useful in general critical care, but it does not include liver-disease-specific markers or measures of inflammation [10]. The Model for End-Stage Liver Disease–Sodium (MELD-Na) score was developed for liver transplant waitlist prioritization and captures hepatic and renal failure, but its time horizon is months rather than days, and it does not adjust for concurrent acute critical illness [11]. The CHA₂DS₂-VASc score is designed to estimate stroke risk in non-valvular AF and is not intended for short-term ICU mortality prediction. It is included in our analyses only as a secondary reference to illustrate that an AF-specific score alone does not capture short-term ICU risk in this population; the primary comparators for a mortality prediction model in this setting are SOFA and MELD-Na [12]. Inflammation is the common omission across all three. Inflammatory activation in advanced cirrhosis drives bacterial translocation and acute decompensation, and is closely tied to acute chronic liver failure progression and mortality.

Cirrhosis is now understood to involve chronic systemic inflammation that interacts with cardiac function and drives multiorgan failure during acute insults [13]. Several composite indices derived from routine complete blood count parameters have been proposed as accessible biomarkers of this inflammatory state [14]. One of these, the systemic inflammation response index (SIRI = neutrophils × monocytes / lymphocytes), draws on three immune cell lineages at once. Neutrophilia reflects innate immune activation [14]. Monocyte counts respond to reticuloendothelial stress and bridge innate and adaptive responses [15]. A related index, the systemic immune-inflammation index (SII = platelets × neutrophils / lymphocytes), incorporates platelet count and has shown prognostic value in oncology and cardiovascular populations. In cirrhosis, however, platelet counts are systematically reduced by portal hypertension and hypersplenism, and their prognostic signal is confounded by this pathophysiology rather than by inflammation. SIRI, which replaces platelets with monocytes, avoids this issue and was therefore favored on a priori grounds for use in cirrhotic populations. Lymphopenia signals adaptive immune exhaustion, which is itself a known mortality predictor in cirrhosis [16]. SIRI has shown prognostic value in sepsis [17], malignancy [18], and cardiovascular disease [19], but whether it adds prognostic information beyond established severity scores in critically ill cirrhotic patients with concurrent AF has not been formally tested.

We asked whether a simple inflammation index from the blood count could add prognostic information on top of SOFA and MELD-Na in this ICU population. To answer this, we compared five candidate composite inflammation indices and selected the one with the best balance of discrimination and clinical interpretability for predicting 28-day ICU mortality. We then built a nomogram combining the chosen marker with standard severity scores, and validated it in an independent multicenter cohort.

Methods

Study design and data sources

We conducted a retrospective cohort study using two large publicly available critical care databases. The Medical Information Mart for Intensive Care IV (MIMIC-IV), version 3.1, developed at Beth Israel Deaconess Medical Center, served as the development cohort [20]. The eICU Collaborative Research Database (eICU-CRD), version 2.0, which compiles de-identified records from 208 hospitals across the United States, served as the external validation cohort [21]. Use of both datasets was approved by the institutional review boards of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center, with waiver of informed consent and approved data-sharing protocols. A member of the study team completed the Collaborative Institutional Training Initiative (CITI) “Data or Specimens Only Research” certification (CITI ID 63749646) prior to data access. Patient-level data were extracted from both databases using SQL queries executed in pgAdmin 4. The study followed the ethical principles of the Declaration of Helsinki and was reported in accordance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) guideline [22].

Study population

Adult patients (≥18 years) admitted to an ICU with concurrent ICD-9/10 diagnoses of liver cirrhosis and atrial fibrillation were eligible for inclusion. The specific ICD-9 and ICD-10 codes used for case identification are provided in Supplementary Table S5. For patients with multiple ICU admissions, only the first stay was analyzed. Patients with missing outcome data or missing data for every candidate inflammation marker were excluded. The final analytic cohorts comprised 454 patients in MIMIC-IV and 121 patients in eICU-CRD (Fig. 1).

Fig. 1.

Study flow diagram. Sequential patient selection from the MIMIC-IV (development cohort) and eICU-CRD (external validation cohort) databases, with the exact number of patients excluded at each step, followed by the four-step statistical analysis pipeline

Outcome

The primary outcome was 28-day ICU mortality, defined as death in the ICU within 28 days of admission. Patients discharged from the ICU alive or still in the ICU at day 28 were classified as survivors.

Candidate predictors

Candidate predictors were extracted from data recorded within the first 24 hours of ICU admission. Demographic variables included age, sex, race, and body mass index (BMI). Disease severity at admission was captured by the Sequential Organ Failure Assessment (SOFA), Oxford Acute Severity of Illness Score (OASIS), Glasgow Coma Scale (GCS), Model for End-Stage Liver Disease–Sodium (MELD-Na), and CHA₂DS₂-VASc score. Inflammation and nutrition composite indices included the systemic inflammation response index (SIRI = neutrophils × monocytes / lymphocytes), log-SIRI = ln(SIRI + 1), the systemic immune-inflammation index (SII = platelets × neutrophils / lymphocytes), log-SII, neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), and prognostic nutritional index (PNI = albumin [g/L] + 5 × lymphocyte count [×10⁹/L]). Laboratory values comprised albumin, total bilirubin, international normalized ratio (INR), creatinine, sodium, white blood cell count, and platelets. Vital signs included heart rate, mean arterial pressure, peripheral oxygen saturation (SpO₂), temperature, and respiratory rate. Comorbidities were derived from Charlson coding and included congestive heart failure, diabetes mellitus, hypertension, myocardial infarction, peripheral vascular disease, and stroke or transient ischemic attack. In addition to SOFA, OASIS, MELD-Na, and CHA₂DS₂-VASc, the SAPS II score was extracted for comparison against the nomogram. Treatment variables (proton pump inhibitors, anticoagulants, beta-blockers, diuretics, vasoactive agents, and mechanical ventilation) were extracted for descriptive cohort characterization only. These were not considered candidate predictors for the mortality model because they reflect clinical decisions made in response to disease trajectory and are not available at the prediction time point (ICU admission). Because predictors were extracted from the first 24 hours of ICU admission, the nomogram provides a risk estimate that is applicable 24 hours after ICU admission, once the initial clinical data have been collected. It is not intended as an at-admission model.

Missing data

Variables with more than 30% missing data were excluded from analysis. For variables with missingness below this threshold, missing values were imputed using k-nearest neighbor (KNN) imputation with k = 5, implemented in the VIM package (Table S10). The outcome variable and all inflammation markers were excluded from the distance metric to prevent outcome leakage. In the eICU-CRD external validation cohort, KNN imputation was performed independently within the eICU-CRD cohort itself (k = 5), rather than by projecting eICU-CRD observations onto the MIMIC-IV development cohort. This choice reflects the different variable definitions, laboratory reference ranges, and missingness patterns between the two databases, and preserves the missing-at-random assumption within each cohort. A limitation of this approach is that it does not perfectly reflect real-world deployment, in which a single new patient would be imputed against a fixed reference dataset.

Statistical analysis

Univariable logistic regression characterized the unadjusted association of each candidate predictor with 28-day ICU mortality. Predictors for the multivariable model were selected on clinical and structural grounds rather than by univariable significance testing, to avoid data-driven selection. The base clinical model comprised four prespecified variables, age, sex, SOFA, and MELD-Na, representing the minimum non-redundant set of established prognostic anchors against which candidate inflammatory markers were tested. Age was included as a near-universal ICU outcome confounder; sex was included as a demographic covariate for transportability. SOFA was chosen over OASIS as the general ICU severity score because the two scores are designed to serve the same purpose and their component overlap would introduce collinearity if both were included; SOFA is also more widely used internationally and is a recognized component of the EASL-CLIF and NACSELD acute-on-chronic liver failure definitions. GCS was not entered as a separate predictor because it is a component of SOFA (the neurological subscore) and is therefore captured implicitly. MELD-Na was retained as the liver-specific severity anchor because it carries information (INR, bilirubin, sodium) that SOFA does not incorporate. Other univariably significant variables were not added to the base model to avoid redundancy: INR and total bilirubin are components of MELD-Na, and respiratory rate is a component of SOFA. Retention of prespecified variables regardless of joint-model significance is standard practice for prediction models and preserves transportability. Multicollinearity among predictors was assessed using variance inflation factors (VIF); all values were below 1.5, indicating no problematic collinearity between SOFA and MELD-Na. Sensitivity analyses using alternative parameterizations of organ dysfunction confirmed that the log-SIRI effect was stable (Table S13). Five candidate inflammation indices (SIRI, log-SIRI, SII, log-SII, NLR) were then compared head-to-head, with each added individually to a base clinical model containing age, sex, SOFA, and MELD-Na, and ranked by AUC, Akaike information criterion (AIC), and the clinical interpretability of the per-unit odds ratio. The functional form of the selected marker was examined with restricted cubic splines using four knots at the 5th, 35th, 65th, and 95th percentiles, with a likelihood-ratio test for nonlinearity. A multivariable logistic regression model was then fitted in the MIMIC-IV cohort with log-SIRI, age, sex, SOFA, and MELD-Na as predictors; the events-per-variable ratio was 17.8 (89 events for 5 predictors), exceeding the conventional minimum of 10. A nomogram was constructed from the model coefficients using the rms package. To assess robustness, four progressively adjusted models were fitted: Model 1 (crude), Model 2 (adjusted for age and sex), Model 3 (additionally adjusted for SOFA and MELD-Na, equivalent to the final nomogram model), and Model 4 (further adjusted for BMI and six comorbidities). Individual vital signs (heart rate, mean arterial pressure, respiratory rate, temperature) were not entered as separate predictors because they are already captured within the composite severity scores in the base model: respiratory rate and Glasgow Coma Scale are components of SOFA, and heart rate, mean arterial pressure, and temperature are components of OASIS.

Internal validation was performed by 1000-resample bootstrap to obtain optimism-corrected estimates of the C-index, calibration slope, calibration intercept, and Brier score. Discrimination was further evaluated by AUC with DeLong 95% confidence intervals; calibration by the Hosmer-Lemeshow goodness-of-fit test and graphical calibration plots; and clinical utility by decision curve analysis. The reclassification improvement of the nomogram over SOFA, MELD-Na, and CHA₂DS₂-VASc was quantified by the integrated discrimination improvement (IDI) and the continuous net reclassification index (cNRI), following Pencina et al. The locked nomogram was then applied to the eICU-CRD cohort, where performance was assessed by AUC and decision curve analysis.

Continuous variables were summarized as median [interquartile range] and compared using the Wilcoxon rank-sum test. Categorical variables were summarized as count (percent) and compared using the χ² test or Fisher’s exact test as appropriate. A two-sided P value < 0.05 was considered statistically significant. All analyses were performed in R version 4.4 using the rms, VIM, pROC, openxlsx, and ggplot2 packages.

Results

Basic characteristics

The MIMIC-IV development cohort comprised 454 adult patients with concurrent diagnoses of liver cirrhosis and atrial fibrillation, of whom 89 (19.6%) died in the ICU within 28 days. Median age was 67 years (IQR 59–75), 66.5% were male, and the cohort was predominantly Caucasian (65.9%) (Table 1). Non-survivors had higher disease-severity scores at admission than survivors (SOFA 10 vs 8, MELD-Na 28.5 vs 23.3, OASIS 40 vs 34; all P < 0.001), higher systemic inflammation indices (log-SIRI 2.56 vs 1.92, NLR 15.1 vs 8.9; both P < 0.001), and greater coagulopathy (INR 2.00 vs 1.60, P < 0.001). Vasoactive support was used in 91.0% of non-survivors and 55.9% of survivors (P < 0.001). The eICU-CRD external validation cohort comprised 121 patients with comparable demographic and clinical profiles, of whom 20 (16.5%) died in the ICU within 28 days (Supplementary Table S7). A side-by-side comparison of baseline characteristics between the two cohorts is provided in Supplementary Table S6.

Table 1.

Baseline characteristics of the MIMIC-IV training cohort stratified by 28-day ICU mortality

Characteristic	Total (n = 454)	Survivors (n = 365)	Non-survivors (n = 89)	P value
Number of patients	454	365	89
Demographics
Age, years	67 [59,75]	67 [59,75]	66 [60,75]	0.420
Sex, male, n (%)	302 (66.5)	243 (66.6)	59 (66.3)	1.000
Race, n (%)				0.361
Caucasian	299 (65.9)	246 (67.4)	53 (59.6)
African American	27 (5.9)	23 (6.3)	4 (4.5)
Asian	8 (1.8)	5 (1.4)	3 (3.4)
Hispanic	11 (2.4)	8 (2.2)	3 (3.4)
Other/Unknown	109 (24.0)	83 (22.7)	26 (29.2)
Body mass index, kg/m²	29.3 [25.1, 35.3]	28.8 [24.8, 35.2]	31.4 [25.7, 36.5]	0.103
Severity scores
SOFA score	8 [6,11]	8 [5,10]	10 [8,15]	<0.001
OASIS score	35 [29,42]	34 [28,41]	40 [34,47]	<0.001
Glasgow Coma Scale	15 [13,15]	15 [14,15]	14 [11,15]	0.003
MELD-Na	24.1 [17.4, 29.9]	23.3 [16.5, 28.4]	28.5 [21.8, 33.4]	<0.001
CHA₂DS₂-VASc	2 [1,4]	2 [1,4]	2 [1,3]	0.237
Inflammation and nutrition markers
SIRI	6.4 [2.8, 14.3]	5.8 [2.6, 11.9]	12.0 [4.4, 24.6]	<0.001
log-SIRI	2.00 [1.33, 2.73]	1.92 [1.28, 2.56]	2.56 [1.68, 3.24]	<0.001
NLR	9.3 [5.6, 18.4]	8.8 [5.3, 17.0]	15.1 [7.5, 24.7]	<0.001
SII	1230 [556, 2396]	1180 [548, 2222]	1600 [590, 4176]	0.017
PLR	147 [80, 256]	145 [80, 247]	157 [83, 315]	0.436
PNI	35.2 [30.0, 40.7]	35.8 [31.0, 40.5]	32.5 [27.1, 42.0]	0.091
Laboratory values
Albumin, g/dL	3.0 [2.5, 3.5]	3.0 [2.6, 3.5]	2.9 [2.5, 3.7]	0.343
Total bilirubin, mg/dL	2.4 [1.1, 5.9]	2.1 [1.1, 5.2]	3.6 [1.3, 9.4]	0.007
INR	1.70 [1.40, 2.14]	1.60 [1.38, 2.03]	2.00 [1.62, 2.53]	<0.001
Creatinine, mg/dL	1.5 [1.0, 2.6]	1.4 [0.9, 2.5]	1.9 [1.3, 3.2]	0.001
Sodium, mmol/L	137 [134, 140]	137 [134, 140]	136 [132, 140]	0.787
White blood cells, ×10⁹/L	10.6 [7.2, 15.5]	10.0 [7.0, 14.6]	13.7 [7.9, 18.0]	0.001
Platelets, ×10⁹/L	116 [77, 192]	118 [80, 190]	112 [69, 194]	0.404
Vital signs at admission
Heart rate, bpm	87 [75, 102]	86 [75, 101]	93 [77, 107]	0.033
Mean arterial pressure, mmHg	73 [67,80]	74 [67,81]	71 [65,79]	0.092
SpO₂, %	97.0 [95.3, 98.3]	97.1 [95.5, 98.4]	96.4 [95.0, 98.0]	0.025
Temperature, °C	36.7 [36.5, 37.0]	36.7 [36.6, 37.0]	36.7 [36.4, 36.9]	0.051
Respiratory rate, /min	18.9 [16.6, 21.7]	18.6 [16.4, 21.0]	20.7 [17.9, 24.0]	<0.001
Comorbidities, n (%)
Congestive heart failure	192 (42.3)	164 (44.9)	28 (31.5)	0.029
Diabetes mellitus	176 (38.8)	140 (38.4)	36 (40.4)	0.809
Hypertension	133 (29.3)	103 (28.2)	30 (33.7)	0.373
Myocardial infarction	43 (9.5)	36 (9.9)	7 (7.9)	0.707
Peripheral vascular disease	19 (4.2)	16 (4.4)	3 (3.4)	0.894
Stroke	30 (6.6)	27 (7.4)	3 (3.4)	0.257
Treatments during ICU stay, n (%)
Proton pump inhibitor	355 (78.2)	285 (78.1)	70 (78.7)	1.000
Anticoagulant	428 (94.3)	346 (94.8)	82 (92.1)	0.475
Beta-blocker	324 (71.4)	275 (75.3)	49 (55.1)	<0.001
Diuretic	344 (75.8)	291 (79.7)	53 (59.6)	<0.001
Vasoactive agent	285 (62.8)	204 (55.9)	81 (91.0)	<0.001
Mechanical ventilation	375 (82.6)	295 (80.8)	80 (89.9)	0.062
Clinical outcomes
ICU length of stay, days	4.1 [2.2, 7.9]	3.9 [2.1, 7.4]	6.2 [3.3, 10.9]	0.002

Univariable predictors of 28-day ICU mortality

Univariable logistic regression identified significant associations between 28-day ICU mortality and disease-severity scores (SOFA, OASIS, GCS, MELD-Na; all P < 0.001), inflammation markers (SIRI, log-SIRI, NLR, SII; all P ≤ 0.003), coagulation indices (INR, P < 0.001; total bilirubin, P = 0.001), and respiratory rate (P < 0.001) (Table S1). Treatment variables were not entered into univariable prediction analyses because they represent post-baseline clinical decisions rather than admission-time predictors; their frequencies in survivors and non-survivors are shown in the baseline characteristics table (Table 1) for descriptive purposes only.

Selection and modeling of log-SIRI

Among the five candidate markers added individually to the base clinical model, SII achieved the highest discrimination (AUC 0.758) and the lowest AIC (390.2), with log-SII (AIC 392.7) and log-SIRI (AIC 394.5) closely behind. Raw SIRI and NLR ranked lower on both criteria (Table S2). We selected log-SIRI as the final marker not because it was the best-performing marker statistically, but for a combination of reasons: (i) its three-lineage formulation avoids platelet count, which is unreliable in cirrhosis due to portal hypertension and hypersplenism, so its biological signal is more specifically inflammatory in this population; (ii) its odds ratio is directly interpretable on the nomogram scale (OR 1.48 per unit, or 1.73 per SD), unlike SII whose per-unit odds ratio (1.0001) required rescaling to be interpretable; and (iii) its discrimination is nearly identical to SII (ΔAUC = -0.007, well within sampling variability at this sample size). We acknowledge that a rescaled SII or log-SII would be a defensible alternative, and this trade-off is discussed further in the Discussion. Restricted cubic spline analysis confirmed an approximately linear association between log-SIRI and 28-day ICU mortality (likelihood-ratio test for nonlinearity P = 0.456; ΔAIC = +2.44 favoring the linear specification). Raw SIRI showed a saturating dose-response pattern, supporting the logarithmic transformation (Figure S1).

Nomogram development

The final multivariable logistic regression model included log-SIRI, age, sex, SOFA, and MELD-Na. After mutual adjustment, log-SIRI (OR 1.48, 95% CI 1.14–1.90; P = 0.003) and SOFA (OR 1.21, 95% CI 1.12–1.30; P < 0.001) remained independent predictors of 28-day ICU mortality, while age, sex, and MELD-Na were not (Table 2). MELD-Na’s univariable association with mortality was therefore largely captured by SOFA in the joint model, consistent with overlap in the renal severity component (creatinine appears in both scores). MELD-Na was nevertheless retained because it also carries INR, bilirubin, and sodium information that SOFA does not incorporate, and because it serves as the standard liver-specific severity anchor in cirrhotic populations. Age, sex, and MELD-Na were retained as prespecified structural predictors rather than as independent significant predictors; we do not interpret the model as demonstrating independent prognostic value for each of the five variables. Across four hierarchical models, the odds ratio for log-SIRI remained statistically significant from the crude value (1.73) to the fully adjusted value (1.53), with all P ≤ 0.003 (Table 3). Discrimination increased from AUC 0.648 in Model 2 to 0.751 in Model 3 when SOFA and MELD-Na were added. Model 3 had the lowest AIC of the four models (394.5) and was selected as the final nomogram model. The nomogram is shown in Fig. 2. Compared with ICU-specific severity scores (Table S14), the nomogram (AUC 0.751) discriminated significantly better than OASIS (0.682, P = 0.027) and numerically better than SAPS II (0.707, P = 0.083) and SOFA (0.718, P = 0.054). Adding log-SIRI to the base model improved discrimination at least as much as adding raw white blood cell count (0.751 vs 0.746).

Table 2.

Final multivariable logistic regression model for predicting 28-day ICU mortality

Variable	β	SE	OR (95% CI)	P value
(Intercept)	-5.124	1.093	—	<0.001
log-SIRI	0.389	0.130	1.48 (1.14, 1.90)	0.003
Age (years)	0.007	0.012	1.01 (0.98, 1.03)	0.549
Sex (Male vs Female)	0.116	0.273	1.12 (0.66, 1.92)	0.670
SOFA	0.187	0.037	1.21 (1.12, 1.30)	<0.001
MELD-Na	0.022	0.019	1.02 (0.98, 1.06)	0.252

Table 3.

Multivariable logistic regression of log-SIRI for 28-day ICU mortality

Model	Adjusted for	No. of variables	EPV	log-SIRI
OR (95% CI)	P value	AUC	AIC
Model 1 (Crude)	Unadjusted	1	89.0	1.73 (1.36, 2.20)	<0.001	0.643	432.0
Model 2 (+ Age, Sex)	Age + Sex	3	29.7	1.75 (1.37, 2.22)	<0.001	0.648	434.2
Model 3 (final nomogram)	+ SOFA + MELD-Na	5	17.8	1.48 (1.14, 1.90)	0.003	0.751	394.5
Model 4 (Fully adjusted)	+ BMI + comorbidities	12	7.4	1.53 (1.17, 1.99)	0.002	0.766	399.5

Fig. 2.

Nomogram for predicting 28-day ICU mortality. For each predictor, the contribution to risk is read on the “Points” axis. Summed across all five variables, the “Total Points” map to the predicted probability of 28-day ICU mortality on the bottom axis

Incremental value of log-SIRI over the base clinical model

To directly quantify the value added by log-SIRI, we compared a base clinical model containing age, sex, SOFA, and MELD-Na with the same model plus log-SIRI (Table 4). The base model achieved an AUC of 0.721 (95% CI 0.660–0.783); adding log-SIRI increased the AUC to 0.751 (95% CI 0.695–0.808), a difference of +0.030 (bootstrap 95% CI + 0.005 to +0.057; DeLong P = 0.026). The likelihood ratio test showed that adding log-SIRI significantly improved model fit (χ² = 9.18, df = 1, P = 0.002). Calibration also improved: the base model showed evidence of miscalibration (Hosmer–Lemeshow P = 0.002), whereas the model with log-SIRI was well calibrated (Hosmer–Lemeshow P = 0.729). The Brier score decreased slightly from 0.134 to 0.132. Continuous NRI was 0.241 (95% CI 0.006 to 0.469, P = 0.037), favoring the model with log-SIRI; IDI was 0.018 (95% CI − 0.001 to 0.036, P = 0.059), showing a borderline improvement. Decision curve analysis showed higher net benefit for the model with log-SIRI at clinically relevant thresholds (at threshold 0.20: 0.067 vs 0.064; at threshold 0.30: 0.047 vs 0.041; Figure S2). The gain in AUC was modest in magnitude, but the significant likelihood ratio test, cNRI, DeLong test, and calibration improvement all supported an incremental prognostic contribution of log-SIRI beyond the four base clinical variables. To provide operational metrics that do not rely on IDI or cNRI, sensitivity, specificity, PPV, and NPV at prespecified thresholds (10%–30%) for both models are reported in Table S11; adding log-SIRI consistently improved sensitivity and PPV while maintaining specificity.

Table 4.

Incremental value of log-SIRI over the base clinical model for predicting 28-day ICU mortality in the MIMIC-IV development cohort

Metric	Base clinical model	Base + log-SIRI (final nomogram)
Model composition
Predictors	age, sex, SOFA, MELD-Na	age, sex, SOFA, MELD-Na, log-SIRI
Discrimination
AUC (95% CI)	0.721 (0.660–0.783)	0.751 (0.695–0.808); ΔAUC = +0.030
ΔAUC bootstrap 95% CI (2000 replicates)	(reference)	+0.005 to +0.057
DeLong test	(reference)	P = 0.026
Likelihood ratio test	(reference)	χ² = 9.18, df = 1; P = 0.002
Overall accuracy
Brier score (bootstrap 95% CI)	0.134 (0.113–0.156)	0.132 (0.112–0.154); Δ = −0.002
Calibration
Hosmer–Lemeshow test (10 groups)	P = 0.002	P = 0.729
Bootstrap-corrected calibration slope	—	0.926
Bootstrap-corrected calibration intercept	—	−0.084
Reclassification
Continuous NRI (95% CI)	(reference)	0.241 (0.006 to 0.469); P = 0.037
IDI (95% CI)	(reference)	0.018 ( − 0.001 to 0.036); P = 0.059
Clinical utility (decision curve analysis)
Net benefit at threshold 0.20	0.064	0.067
Net benefit at threshold 0.30	0.041	0.047

Internal validation

In the MIMIC-IV development cohort, the nomogram achieved an AUC of 0.751 (95% CI 0.695–0.808), significantly higher than MELD-Na (AUC 0.676, DeLong P = 0.002) and CHA₂DS₂-VASc (AUC 0.540, DeLong P < 0.001). The comparison against SOFA (AUC 0.718) did not reach statistical significance (DeLong P = 0.054) (Fig. 3A; Table S4). Calibration was good, with predicted and observed probabilities aligning across deciles of risk and a non-significant Hosmer–Lemeshow test (P = 0.729; Fig. 3B). Decision curve analysis showed greater net benefit for the nomogram across threshold probabilities of 0.1 to 0.5 (Fig. 3C). Bootstrap resampling (B = 1000) gave an optimism-corrected C-index of 0.734 (optimism = 0.017), a calibration slope of 0.926, and an intercept of −0.084, consistent with minimal overfitting (Table S3). To isolate the contribution of log-SIRI beyond organ dysfunction, we compared SOFA alone with SOFA plus log-SIRI (Table S15). Adding log-SIRI significantly improved discrimination (AUC 0.718 to 0.751; ΔAUC + 0.033, 95% CI + 0.003 to +0.063; DeLong P = 0.038), with significant reclassification (IDI 0.023, P = 0.029; continuous NRI 0.391, P = 0.001) and a likelihood ratio test P = 0.0007.

Fig. 3.

Internal validation of the nomogram in the MIMIC-IV development cohort (n = 454; 89 events). A ROC curves comparing the nomogram with three reference scores. B Calibration plot with 10 deciles, 95% CIs, and a LOESS-smoothed curve (Hosmer–Lemeshow P = 0.729). C Decision curve analysis

Reclassification analyses

The nomogram improved risk classification over the two primary severity scores (SOFA, MELD-Na) and, as expected, over CHA₂DS₂-VASc, which is a stroke risk score rather than a mortality predictor (Table 5). The integrated discrimination improvement (IDI) versus SOFA, MELD-Na, and CHA₂DS₂-VASc was 0.029 (P = 0.009), 0.097 (P < 0.001), and 0.158 (P < 0.001), respectively. The continuous net reclassification index (cNRI) followed the same pattern: 0.389, 0.690, and 0.711 (all P < 0.001). The DeLong test for AUC versus SOFA did not reach statistical significance (P = 0.054). Reclassification metrics (IDI, cNRI) and decision curve analysis favored the nomogram, but the incremental gain over SOFA on discrimination alone is modest and should be interpreted with caution.

Table 5.

Reclassification performance of the nomogram versus existing prognostic scores

Comparison	IDI (95% CI)	P value	Continuous NRI (95% CI)	P value
Nomogram vs SOFA score	0.029 (0.007, 0.050)	0.009	0.389 (0.161, 0.618)	<0.001
Nomogram vs MELD-Na	0.097 (0.065, 0.129)	<0.001	0.690 (0.472, 0.909)	<0.001
Nomogram vs CHA₂DS₂-VASc	0.159 (0.114, 0.203)	<0.001	0.711 (0.491, 0.931)	<0.001

External validation

When the locked nomogram was applied to the eICU-CRD cohort, the point estimate of discrimination was AUC 0.818 (95% CI 0.722–0.914) (Fig. 4A). The wide confidence interval, driven by only 20 outcome events in 121 patients, means this external estimate should be interpreted as preliminary rather than as a stable measure of external performance. In comparative analyses within the external cohort (Table S16), the nomogram (AUC 0.818) discriminated better than SOFA alone (0.770) and the base model without log-SIRI (0.754), and the reclassification improvement of adding log-SIRI remained significant (IDI 0.062, P = 0.005; continuous NRI 0.767, P = 0.0003). The paired DeLong tests did not reach significance (nomogram vs SOFA P = 0.289; vs base P = 0.184), consistent with the limited number of external events. Decision curve analysis showed net benefit above the default “treat-all” and “treat-none” strategies across threshold probabilities of 0.1 to 0.5 (Fig. 4B). External calibration was assessed by applying the locked MIMIC-IV coefficients to the eICU-CRD cohort (Figure S3, Table S12). Calibration-in-the-large was 0.001 (95% CI − 0.506 to 0.508), with mean predicted risk equal to the observed event rate (both 16.5%); the Brier score was 0.110 and the Hosmer–Lemeshow test was non-significant (P = 0.331). The calibration slope was 1.62 (95% CI 0.866 to 2.376), with the confidence interval including 1.

Fig. 4.

External validation of the nomogram in the eICU-CRD cohort (n = 121; 20 events). A ROC curve. B Decision curve analysis

Discussion

In MIMIC-IV and eICU-CRD, we built a five-variable nomogram for 28-day ICU mortality in critically ill patients with cirrhosis and atrial fibrillation. The variables were log-SIRI, age, sex, SOFA, and MELD-Na. In MIMIC-IV, the model showed good discrimination and calibration, with little optimism after bootstrap correction. Compared with the base clinical model, the gain in AUC from adding log-SIRI was modest in magnitude (+0.030) but was statistically significant on the DeLong test (P = 0.026), the likelihood ratio test (P = 0.002), and the continuous NRI (P = 0.037); calibration also improved substantially, so the value of log-SIRI is best understood as complementary to SOFA and MELD-Na rather than dominant. In the smaller external validation cohort (eICU-CRD, n = 121; 20 events), the point estimate of the AUC was 0.818 with a wide 95% confidence interval (0.722–0.914), consistent with the limited number of external events. Reclassification and decision curve analyses in MIMIC-IV favored the nomogram over SOFA, MELD-Na, and CHA₂DS₂-VASc, although the AUC advantage over SOFA alone did not reach statistical significance. As expected, CHA₂DS₂-VASc, a stroke risk score, performed poorly for ICU mortality prediction and served only as a secondary reference. Most prior work on inflammation biomarkers in cirrhosis has been done in chronic outpatients or short-term inpatient cohorts. Critically ill patients with co-existing AF have not been specifically studied. Our work contributes two things: a head-to-head comparison of five hematologic inflammation indices in this setting, and a preliminary nomogram that combines log-SIRI with standard severity scores. The nomogram is intended as a research and exploratory tool. Prospective external calibration, impact assessment, and validation across additional healthcare systems and AF subtypes will be required before it can be considered for clinical adoption. The nomogram is designed to be applied at 24 hours after ICU admission, once the first 24-hour clinical data (including the complete blood count required for log-SIRI, SOFA, and MELD-Na) are available. It is not intended for use at the moment of admission before these data have been collected.

The link between SIRI and mortality is biologically plausible in cirrhosis because advanced liver disease combines innate immune activation, monocyte/macrophage dysfunction, and lymphopenia [23]. Decompensated cirrhosis involves cirrhosis-associated immune dysfunction: a coexistence of immunodeficiency and persistent systemic inflammation, driven by pathogen- and damage-associated molecular patterns translocating from the gut [24]. Each of the three cell types in SIRI maps onto a different part of this response. Neutrophilia reflects innate activation, usually triggered by infection or sterile tissue injury. Monocytes sit at the bridge between innate and adaptive immunity and rise with reticuloendothelial stress [25,26]. Lymphopenia signals adaptive immune exhaustion, which is itself a known mortality predictor in cirrhosis [27]. Because SIRI multiplies these three cell counts together, the score rises sharply when all three abnormalities coexist, a pattern typical of patients progressing toward acute-on-chronic liver failure [28]. AF may add to this picture: systemic inflammation has been linked both to arrhythmogenesis and to hemodynamic decompensation [29]. Log-SIRI remained statistically associated with 28-day ICU mortality after adjustment for SOFA and MELD-Na. Whether this residual association reflects distinct inflammatory biology not captured by the two severity scores, or shared variation with unmeasured factors, cannot be determined from the present analysis and would require mechanistic biomarker work and prospective validation to establish. Age, sex, and MELD-Na did not reach statistical significance in the multivariable model, and the model should not be interpreted as demonstrating independent prognostic value for each of the five variables. They were retained as prespecified structural predictors rather than as data-driven independent predictors, following standard practice for prediction models. MELD-Na in particular was retained because it carries INR, bilirubin, and sodium information not covered by SOFA and serves as the standard liver-specific severity anchor in cirrhotic populations; removing it would leave the model without a liver-specific severity component and reduce its transportability to hepatology settings where MELD-Na is the accepted reference measure.

The choice of log-SIRI over SII deserves explicit discussion. SII had a marginally higher AUC and lower AIC than log-SIRI in the head-to-head comparison, and its near-unity per-unit odds ratio could in principle be rescaled per SD, per IQR, or per 100 units to yield an interpretable coefficient. We therefore based the choice of log-SIRI on three grounds rather than statistical performance alone. The most important is biological suitability in cirrhosis. SII depends on platelet count, which is systematically depressed by portal hypertension and hypersplenism in cirrhotic patients; platelet variation in this population reflects hepatic hemodynamics as much as inflammation. SIRI’s three-lineage formulation (neutrophils, monocytes, lymphocytes) replaces platelets with monocytes and may align more closely with the reticuloendothelial and immune-lineage dysfunction proposed in cirrhosis-associated immune dysfunction. Second, the discrimination difference between log-SIRI and SII was a difference of 0.007 in AUC, which is well within sampling variability at this sample size, so log-SIRI was not selected as inferior on discrimination. Third, log-SIRI shares a similar dynamic range with the other continuous predictors in the model, producing a visually balanced nomogram. We acknowledge that this choice reflects a considered trade-off rather than a purely statistical decision, and that a rescaled SII would be a defensible alternative in populations where hypersplenism is not a dominant issue. Restricted cubic spline analysis showed an approximately linear relationship between log-SIRI and mortality, and log-SIRI was therefore entered as a linear term.

In the eICU-CRD external validation, the point estimate of the AUC (0.818) was numerically higher than in the development cohort (0.751). This apparent improvement should be read with strong caution. The eICU-CRD cohort had only 121 patients and 20 outcome events, and the AUC confidence interval was correspondingly wide (0.722–0.914). The intervals for the development and validation AUCs overlap substantially, so the numerical difference most likely reflects sampling variability rather than a true improvement in generalization. Differences in case mix between the single-center MIMIC-IV cohort and the multicenter eICU-CRD cohort may also contribute. Decision curve analysis showed that the nomogram retained clinical utility in the external cohort, but given the limited external sample size, external estimates should be regarded as preliminary and stronger conclusions about generalizability will require validation in additional cohorts.

The two cohorts also differed substantially in treatment intensity (vasoactive support 62.8% vs 14.9%; mechanical ventilation 82.6% vs 32.2%), reflecting the higher acuity of the single-center development cohort. However, the prognostic value of log-SIRI did not vary by treatment: interactions between log-SIRI and vasoactive support (P = 0.889) and mechanical ventilation (P = 0.327) were both non-significant, and stratified estimates were consistent with the overall effect. Because treatment variables reflect post-baseline responses rather than admission-time predictors, they were not included in the nomogram.

The improvement was supported by concordant changes across multiple metrics. We interpret IDI and continuous NRI as supportive rather than definitive, given that these indices can overstate improvement when prespecified risk categories are not used; the threshold-based sensitivity, specificity, and predictive values (Table S11) and the decision curve analysis provide more directly interpretable evidence of incremental clinical value.

Our study has some strengths. The development and reporting followed the TRIPOD guideline, and candidate predictors were prespecified rather than chosen by stepwise variable selection on the data. With 89 events in the development cohort and only five variables in the final model, the events-per-variable ratio was 17.8, well above the conventional minimum of 10 for logistic regression. We performed internal validation using 1000-resample bootstrap to obtain optimism-corrected estimates rather than a simple split-sample approach. External validation was carried out in a separate multicenter database, which is more informative about generalizability than internal cross-validation alone.

Our study has several limitations. First, ICU discharge is a competing event for in-ICU death. Two sensitivity analyses addressed this concern: a 28-day all-cause mortality outcome (170 events; Table S8) and exclusion of the 6 patients who died within the first 24 hours (Table S9). Both preserved the log-SIRI effect (OR 1.47 and 1.39 vs 1.48 in the main analysis) and its incremental value over the base clinical model, indicating that the primary results are not driven by informative ICU discharge or by predictor-outcome overlap in the earliest deaths. Second, the outcome is restricted to in-ICU death rather than 28-day all-cause mortality; the model therefore predicts short-term ICU risk. Third, cases were identified by ICD-9/10 codes without chart review, which introduces potential misclassification and clinical heterogeneity. We could not systematically distinguish infection, gastrointestinal bleeding, ACLF stage, cirrhosis etiology, or AF subtype, each of which may relate differently to systemic inflammation and mortality; the model’s behavior across these subgroups will need to be examined in prospective cohorts. Relatedly, cirrhosis etiology, the primary reason for ICU admission, and the incidence of hepatic encephalopathy and ascites could not be reliably ascertained from administrative codes and were therefore not reported; characterizing the model across these clinically important strata is a priority for future prospective studies. Fourth, both databases are retrospective and built from routine electronic health records, which limits granularity. Fifth, the external validation cohort included only 121 patients and 20 events, so external estimates carried wide confidence intervals and precluded meaningful subgroup analyses. Although external calibration-in-the-large was excellent (0.001; mean predicted risk equal to observed), the calibration slope was estimated imprecisely (1.62, 95% CI 0.866 to 2.376) owing to the small number of events, so larger prospective validation is needed before individual-level risk prediction. In particular, paired AUC comparisons in the external cohort were underpowered (20 events), so external discrimination differences between models should be interpreted alongside the point estimates and reclassification metrics rather than DeLong P values alone. Sixth, missing data were handled by KNN imputation with the outcome and inflammation markers excluded from the distance metric, and KNN imputation in eICU-CRD was performed within eICU-CRD rather than by matching to MIMIC-IV neighbors; residual bias from incomplete data cannot be ruled out, and prospective validation with standardized data collection is a needed next step. Imputation was performed once before bootstrap resampling rather than nested within each iteration, which may yield mildly optimistic performance estimates, although the observed bootstrap optimism was small (0.017). Seventh, the nomogram uses a single baseline log-SIRI value; whether longitudinal trajectories add prognostic information remains open. Eighth, we did not evaluate clinical implementation, so it is not yet clear whether using the nomogram at the bedside changes patient-relevant outcomes. Ninth, APACHE II was not available in a validated form in our MIMIC-IV extract, so ICU-score comparison was limited to OASIS, SAPS II, and SOFA, all of which the nomogram matched or exceeded. Finally, both cohorts are based on North American ICUs, and the model’s behavior in other healthcare systems and ethnic groups will need to be tested before broad use.

Conclusion

In critically ill patients with cirrhosis and atrial fibrillation, log-SIRI was associated with 28-day ICU mortality after adjustment for age, sex, SOFA, and MELD-Na. A five-variable nomogram built around log-SIRI showed acceptable discrimination and calibration in MIMIC-IV and suggestive external performance in the smaller eICU-CRD cohort. Compared with the base clinical model, the gain in AUC was modest, but reclassification, likelihood ratio testing, and decision curve analysis were consistent with incremental prognostic value. These findings should be regarded as exploratory rather than confirmatory. Prospective external calibration, evaluation across additional healthcare systems and AF subtypes, and implementation studies are needed before the model can be considered for individual clinical use.

Author contributions

Xiangle He and Jiajing Han designed the study, performed the data extraction and statistical analyses, and prepared the materials. Xiangle He drafted the manuscript, which was critically revised by Jiajing Han. All authors reviewed and approved the final manuscript.

Funding

Not applicable.

Data availability

The data used in this study are publicly available from PhysioNet: MIMIC-IV v3.1 (https://physionet.org/content/mimiciv/3.1/) and eICU-CRD v2.0 (https://physionet.org/content/eicu-crd/2.0/). Analysis code is available from the corresponding author upon reasonable request.

Declarations

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1186/s12876-026-05291-6.