Risk-stratified assessment of 180-day mortality in severe alcoholic hepatitis

Abstract

BACKGROUND

The clinical heterogeneity of severe alcoholic hepatitis (SAH) poses challenges in precision prognosis.

AIM

To identify distinct clinical phenotypes of SAH and develop a simplified, risk-stratified model for predicting 180-day mortality.

METHODS

A total of 1536 patients with SAH were categorized into training and testing cohorts. After comparing nine machine learning (ML) models, we employed the SHapley Additive exPlanations (SHAP) analysis to identify the optimal model and key predictors. Based on these predictors [creatinine (CRE) and total bilirubin (TBIL) levels, prothrombin time, and cholinesterase activity], we performed latent class analysis (LCA) to determine underlying phenotypes. We developed a two-step risk stratification system: Primary and secondary stratifications based on CRE and TBIL level, respectively.

RESULTS

Overall, 1536 patients were enrolled. Among the nine ML models, XGBoost demonstrated the best performance (test set area under the curve of 0.848). SHAP analysis identified CRE as the most important predictor. The subsequent LCA revealed two distinct phenotypes primarily differentiated by CRE level, each associated with markedly different 180-day survival rates: 78.5% and 47.9%, respectively. Within both CRE subtypes, TBIL level was the most significant prognostic factor. The final two-step CRE-TBIL risk assessment stratified patients into four tiers, demonstrating significant mortality differences ranging from 5% to over 50% in both the training and validation cohorts.

CONCLUSION

By integrating ML interpretability with phenotypic analysis, we developed a CRE-TBIL-based 180-day mortality risk assessment card to enable clinicians to rapidly evaluate risk and guide treatment decisions.

Key Words: Maddrey discriminant function; Model for end-stage liver disease; Machine learning; Prognosis; Predictive model

Core Tip: This study identified distinct clinical phenotypes of severe alcoholic hepatitis (SAH) using interpretable machine learning, enhancing prognosis precision. By analyzing 1536 patients, we found that creatinine and total bilirubin levels are key predictors of 180-day mortality. A novel two-step risk stratification model revealed significant mortality variations, classifying patients into four tiers with survival rates ranging from 5% to over 50%. This tool facilitates rapid risk assessment and informed treatment decisions, marking a significant advancement in managing SAH.

INTRODUCTION

Severe alcoholic hepatitis (SAH) represents the most critical form of alcoholic liver disease (ALD), a progressive condition characterized by hepatic failure. Typically, this condition is defined by Maddrey discriminant function (MDF) score ≥ 32 or model for end-stage liver disease (MELD) score > 20[1], and is associated with high short-term mortality[2]. Traditional scoring systems such as MDF[3], MELD[4], and Lille scoring systems[5] are widely used in clinical practice. These scoring systems employ linear regression algorithms to convert laboratory parameters into quantitative risk scores, providing preliminary guidance for clinical decision-making.

In recent years, artificial intelligence (AI) and machine learning (ML) technologies have considerably improved the accuracy of prognostic prediction by integrating high-dimensional clinical data and modeling nonlinear relationships[6]. A 2024 global multicenter study used an integrated model constructed with random forests (RF), gradient boosters, and XGBoost (the ALCoholic Hepatitis Artificial Intelligence Integration Score), which significantly outperformed conventional models in predicting 90-day mortality with an area under the curve (AUC) of 0.799 (P < 0.001). However, its predictive ability for long-term prognosis (e.g., 180 days) was not evaluated[7]. Therefore, developing an ML-based 180-day prognostic model to more accurately predict the long-term mortality risk in patients with SAH is of clinical importance.

In this study, we hypothesized that by integrating high-performance ML models with advanced SHapley Additive exPlanations (SHAP) interpretability techniques, we can develop and validate a novel, data-driven hierarchical risk model to predict 180-day mortality in patients with SAH. This method enables clinicians to rapidly identify patients at risk of mortality and adjust treatment strategies in time.

MATERIALS AND METHODS

Sex as a biological variable

Our study involved both men and women (Table 1).

Table 1 Baseline characteristics of the study population.

Variable	Overall (n = 1536)	Survived (n = 962)	Died (n = 574)	P value
MDF	51.00 (40.00-68.00)	45.50 (37.00-57.00)	64.00 (51.00-82.75)	< 0.001
MELD	16.00 (13.00-20.00)	14.00 (12.00-17.00)	19.00 (16.00-24.00)	< 0.001
Gender (male)	1495 (97.3)	934 (97.1)	561 (97.7)	0.515
Age (year)	48.00 (42.00-54.00)	48.00 (41.00-54.00)	48.00 (43.00-55.00)	0.086
ALB (g/L)	26.00 (22.00-29.00)	27.00 (23.00-30.00)	24.00 (20.00-28.00)	< 0.001
GLO (g/L)	29.00 (23.00-35.00)	30.00 (24.00-36.00)	27.00 (21.00-33.00)	< 0.001
ALT (U/L)	31.00 (21.00-47.00)	29.00 (20.00-42.00)	35.00 (23.00-54.00)	< 0.001
AST (U/L)	69.00 (46.00-111.00)	64.00 (44.00-102.75)	81.00 (52.00-124.00)	< 0.001
ALP (U/L)	136.00 (103.00-182.00)	139.00 (106.00-186.00)	130.00 (95.00-177.00)	0.008
GGT (U/L)	70.00 (32.00-166.00)	73.00 (33.00-165.00)	66.50 (30.00-167.50)	0.371
TBA (μmol/L)	131.00 (70.83-198.25)	127.75 (72.00-199.75)	136.00 (65.25-197.00)	0.952
PTS (second)	20.00 (18.00-23.00)	19.00 (18.00-21.75)	22.00 (20.00-26.00)	< 0.001
PTA (%)	37.00 (31.10-42.10)	38.90 (34.00-42.80)	33.50 (27.02-39.80)	< 0.001
TBIL (μmol/L)	165.40 (97.10-302.15)	129.05 (80.53-235.73)	241.35 (149.90-372.05)	< 0.001
INR	1.74 (1.60-1.99)	1.68 (1.58-1.88)	1.90 (1.68-2.23)	< 0.001
CRE (μmol/L)	83.00 (68.00-112.25)	76.00 (64.00-95.00)	107.00 (79.25-164.75)	< 0.001
TG (mmol/L)	0.90 (0.60-1.37)	0.95 (0.64-1.37)	0.82 (0.55-1.39)	0.001
CHE (U/L)	2067.00 (1504.75-2707.25)	2236.00 (1672.25-2842.75)	1762.00 (1312.00-2469.50)	< 0.001
CK (U/L)	67.00 (39.00-118.00)	69.00 (40.00-118.00)	65.00 (39.00-119.00)	0.979
FE (μmol/L)	19.80 (13.50-26.10)	20.00 (13.12-27.00)	19.50 (14.10-24.78)	0.388
AFP (ng/mL)	4.54 (2.77-8.00)	4.40 (2.81-7.84)	4.80 (2.73-8.88)	0.305
HGB (g/L)	97.00 (79.00-113.00)	99.00 (83.00-115.00)	92.70 (74.00-108.00)	< 0.001
PLT (× 109/L)	71.00 (46.00-113.00)	72.50 (48.00-117.75)	68.20 (43.20-106.00)	0.017
SBP (yes)	473 (30.8)	238 (24.7)	235 (40.9)	< 0.001
Pulmonary infection (yes)	261 (17.0)	107 (11.1)	154 (26.8)	< 0.001
HRS (yes)	108 (7.0)	31 (3.2)	77 (13.4)	< 0.001
HE (yes)	373 (24.3)	139 (14.4)	234 (40.8)	< 0.001
Gastrointestinal bleeding (yes)	174 (11.3)	76 (7.9)	98 (17.1)	< 0.001
Glucocorticoid (yes)	96 (6.2)	41 (4.2)	55 (9.6)	< 0.001

Continuous variables are expressed as median (interquartile range) and categorical variables as n (%). MDF: Maddrey discriminant function; MELD: Model for end-stage liver disease; ALB: Albumin; GLO: Globulin; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; ALP: Alkaline phosphatase; GGT: Gamma-glutamyl transferase; TBA: Total bile acid; PTS: Prothrombin time seconds; PTA: Prothrombin activity; TBIL: Total bilirubin; INR: International normalized ratio; CRE: Creatinine; TG: Triglyceride; CHE: Cholinesterase; CK: Creatine kinase; FE: Iron; AFP: Alpha-fetoprotein; HGB: Hemoglobin; PLT: Platelet Count; SBP: Spontaneous bacterial peritonitis; HRS: Hepatorenal syndrome; HE: Hepatic encephalopathy.

Study design and patient population

This retrospective study involved patients with SAH admitted at the Fifth Medical Center of Chinese PLA General Hospital between January 1, 2013 and December 31, 2022. The inclusion criteria followed the American Gastroenterological Association’s clinical diagnostic criteria for alcoholic hepatitis[8], specifically: (1) Serum total bilirubin (TBIL) level > 3 mg/dL; (2) Aspartate aminotransferase (AST) level > 50 U/L, with an AST/alanine aminotransferase (ALT) ratio of > 1.5, and both ALT and AST levels < 400 U/L; (3) Jaundice onset within 8 weeks; and (4) Persistent alcohol consumption for ≥ 6 months, with daily alcohol intake > 40 g for female patients and > 60 g for male patients, and an abstinence period of < 60 days before jaundice onset. SAH was defined as MDF score > 32 or MELD score > 20. The exclusion criteria were as follows: (1) Viral hepatitis, including hepatitis A, B, C, and E; (2) Autoimmune liver disease; (3) Drug-induced liver injury; (4) Congenital metabolic liver disease or liver disease due to other etiologies; and (5) Severe comorbidities, including cardiovascular, cerebrovascular, respiratory, renal, psychiatric disorders, and malignancies. All patients enrolled in this study were Chinese nationals.

The study protocol was approved by the Ethics Committee of the Fifth Medical Center of Chinese PLA General Hospital (No. KY-ZDZX-2025-11-212-1).

Data collection and variables

We retrospectively collected baseline data within 24 hours of admission via the electronic medical record system. These data included demographic characteristics (age and sex), clinical manifestations (complications such as hepatic encephalopathy and hepatorenal syndrome), laboratory test results (complete blood count, hepatic and renal function assay results, and coagulation parameters), and status of glucocorticoid therapy (yes or no). Based on these data, we calculated the traditional MELD score and MDF score as benchmarks for performance comparison.

Outcome definition and data processing

The primary outcome of this study was mortality within 180 days of admission. For missing data, we employed multiple imputation independently for the training and testing datasets. Subsequently, all patients were stratified by outcome and randomly assigned at a 7:3 ratio to form the training and independent testing cohorts.

Statistical analysis and model development

All statistical analyses were performed using R software (version 4.3.2). Continuous variables are expressed as median (interquartile range) and categorical variables as n (%). Continuous variables were compared using the Mann-Whitney U test and categorical variables were compared using the χ² test.

ML model development, selection, and validation

Within the training set, we developed and compared nine ML models to predict 180-day mortality. Model performance was evaluated using 10-fold cross-validation, with the model yielding the highest mean AUC selected as the optimal model. To assess the clinical value of the model, we further compared its receiver operating characteristic curves against traditional MELD and MDF scores in an independent test set, evaluating AUC differences for significance via DeLong’s test.

Model interpretation and identification of key predictors

To gain insights into the internal decision-making mechanisms of the optimal model (XGBoost), we employed the SHAP method. The SHAP analysis provides precise quantitative values for each feature’s contribution in every prediction[9]. By generating SHAP feature importance maps and beeswarm plots, we identified the top predictors with the greatest overall impact on the risk of 180-day mortality.

Clinical phenotype exploration

To explore the clinical heterogeneity of SAH, we performed latent class analysis (LCA). The top predictors identified using the SHAP analysis [creatinine (CRE) level, TBIL level, prothrombin time, and cholinesterase activity] were incorporated into the LCA model and the optimal number of patient subgroups (phenotypes) was determined by comparing Bayesian Information Criteria values. The clinical significance of these phenotypes was assessed using Kaplan-Meier survival analysis.

Development of a two-step risk stratification tool

We developed a novel two-step stratification tool with the following construction logic. Primary stratification: Serum CRE level, identified as a core variable based on both SHAP analysis and LCA results, serves as the primary stratification indicator. The optimal cut-off point of serum CRE level within the training set was determined by maximizing the Youden index, where Youden index = sensitivity + specificity - 1. This method provides the optimal cut-off point that maximizes the overall correct classification rate between survivors and non-survivors[10]. By applying this cut-off, we categorized patients into high-CRE (≥ 62 μmol/L) and low-CRE (< 62 μmol/L) subgroups. Secondary stratification: To enable more refined risk assessment within subgroups of patients with differing renal function states, we performed independent SHAP subgroup analyses for the high- and low-CRE subgroups. Consistent results indicated that TBIL level was the most significant secondary prognostic determinant in both subgroups. Consequently, TBIL level was selected as the secondary stratification variable, with its optimal cut-off likewise determined within each CRE subgroup of the training set using the Youden index. This yielded distinct TBIL thresholds: ≥ 127 μmol/L for the high-CRE subgroup and ≥ 144 μmol/L for the low-CRE subgroup. Critically, all threshold determinations were confined to the training set to prevent any information leakage from the test set[11]. This process ultimately yielded a 2 × 2 risk assessment tool comprising four risk levels.

Performance evaluation and validation

We first demonstrated the robustness of this stratification strategy in distinguishing mortality across the training and testing cohorts using histograms. Subsequently, the final performance of the tool was comprehensively validated on an independent testing cohort. Kaplan-Meier survival analysis and the Log-rank test were employed to assess differences in survival curves across the four risk groups. Finally, decision curve analysis (DCA) quantified the clinical net benefit of this stratification tool at different risk thresholds to evaluate its clinical utility[12]. All statistical tests were two-sided, and results with P < 0.05 were considered statistically significant.

RESULTS

Baseline characteristics of patients

Based on the inclusion and exclusion criteria, 1536 patients diagnosed with SAH were finally included in this study. The detailed screening process is shown in Figure 1. As presented in Table 1, comparison of baseline characteristics between survivors and non-survivors showed that patients who died exhibited significantly higher MELD scores, serum CRE levels, and TBIL levels at the time of admission (all P < 0.001). In contrast, the serum albumin level and prothrombin time were significantly lower in patients who died (both P < 0.001). These findings provide preliminary insights into key clinical risk factors associated with poor prognosis. All patients were randomized into training (n = 1076) and independent test sets (n = 460). As shown in Table 2, the training and test sets were not statistically different with respect to all baseline characteristics (all P > 0.05), confirming the homogeneity of the groups.

Figure 1 Flowchart of patient selection and model development. A total of 1536 patients with severe alcoholic hepatitis were included based on the predefined inclusion and exclusion criteria. The cohort was randomly split into a training set (n = 1076) for model development and internal validation, and an independent test set (n = 460) for the final evaluation of model performance. The number of survivors and non-survivors in each set is also detailed. SAH: Severe alcoholic hepatitis.

Table 2 Comparison of baseline characteristics between the training and test sets.

Variable	Training set (n = 1076)	Test set (n = 460)	P value
MDF	51.00 (40.00-67.25)	51.00 (40.00-69.00)	0.736
MELD	16.00 (13.00-20.00)	15.00 (13.00-19.00)	0.357
Gender (male)	1048 (97.4)	447 (97.2)	0.863
Age (years)	48.00 (42.00-54.00)	49.00 (43.00-55.00)	0.093
ALB (g/L)	26.00 (22.00-29.00)	26.00 (22.00-29.00)	0.313
GLO (g/L)	29.00 (23.00-35.00)	29.00 (23.00-36.00)	0.464
ALT (U/L)	30.00 (21.00-45.25)	31.00 (21.00-49.00)	0.288
AST (U/L)	69.00 (46.00-110.00)	69.00 (47.00-113.25)	0.664
ALP (U/L)	137.00 (103.00-185.00)	135.00 (104.00-176.25)	0.726
GGT (U/L)	69.00 (31.00-168.00)	74.50 (35.00-159.50)	0.254
TBA (μmol/L)	131.00 (69.00-197.00)	130.50 (74.07-201.75)	0.417
PTS (second)	20.00 (18.00-23.00)	20.00 (18.00-23.00)	0.484
PTA (%)	36.95 (31.08-42.30)	37.10 (31.27-42.00)	0.631
TBIL (μmol/L)	165.05 (97.10-305.95)	166.10 (97.08-293.00)	0.863
INR	1.73 (1.60-1.99)	1.76 (1.60-2.00)	0.404
CRE (μmol/L)	83.00 (67.00-116.00)	83.00 (68.00-107.00)	0.434
TG (mmol/L)	0.88 (0.60-1.37)	0.94 (0.63-1.37)	0.349
CHE (U/L)	2054.00 (1492.75-2703.00)	2080.50 (1524.50-2734.25)	0.343
CK (U/L)	69.00 (40.00-119.00)	65.00 (38.00-115.25)	0.298
FE (μmol/L)	19.55 (13.20-26.00)	20.25 (14.57-26.70)	0.133
AFP (ng/mL)	4.54 (2.75-8.02)	4.53 (2.81-8.00)	0.871
HGB (g/L)	96.00 (79.00-113.00)	99.00 (80.00-113.00)	0.503
PLT (× 109/L)	71.20 (46.00-113.00)	71.00 (47.00-112.25)	0.875
SBP (yes)	323 (30.0)	150 (32.6)	0.334
Pulmonary infection (yes)	195 (18.1)	66 (14.3)	0.075
HRS (yes)	82 (7.6)	26 (5.7)	0.191
HE (yes)	263 (24.4)	110 (23.9)	0.846
Gastrointestinal bleeding (yes)	126 (11.7)	48 (10.4)	0.538
Glucocorticoid (yes)	73 (6.7)	24 (5.0)	0.247

Continuous variables are expressed as median (interquartile range) and categorical variables as n (%). MDF: Maddrey discriminant function; MELD: Model for end-stage liver disease; ALB: Albumin; GLO: Globulin; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; ALP: Alkaline phosphatase; GGT: Gamma-glutamyl transferase; TBA: Total bile acid; PTS: Prothrombin time seconds; PTA: Prothrombin activity; TBIL: Total bilirubin; INR: International normalized ratio; CRE: Creatinine; TG: Triglyceride; CHE: Cholinesterase; CK: Creatine kinase; FE: Iron; AFP: Alpha-fetoprotein; HGB: Hemoglobin; PLT: Platelet Count; SBP: Spontaneous bacterial peritonitis; HRS: Hepatorenal syndrome; HE: Hepatic encephalopathy.

Model performance and key predictor identification

Among the nine ML models compared, the XGBoost model demonstrated the optimal overall prediction performance in both the training (Figure 2A) and test sets (Figure 2B). When compared with the traditional scoring system, the AUC of the XGBoost model was higher than that of the MELD and MDF scores for both the training (Figure 2C) and test sets (test set AUC: XGBoost 0.848 vs MELD 0.821 vs MDF 0.787, DeLong’s test P value were 0.0586 and 0.00058, respectively; Figure 2D).

Figure 2 Machine learning model performance and comparison. A: Receiver operating characteristic (ROC) curves of nine machine learning models in the training set (internal validation); B: ROC curves of the nine models in the independent test set; C: ROC curve comparison between the best-performing model (XGBoost) and traditional scores (model for end-stage liver disease and Maddrey discriminant function) in the training set; D: ROC curve comparison between the XGBoost model and traditional scores in the independent test set. AUC: Area under the curve; ML: Machine learning; MELD: Model for end-stage liver disease; MDF: Maddrey discriminant function.

SHAP analysis was performed on the trained XGBoost model to reveal the decision basis. Global feature construction and validation of the two-step risk card and SHAP swarm plots (Figure 3) consistently identified serum CRE level as the strongest driver for predicting 180-day mortality risk. The subsequent LCA similarly identified two clinical phenotypes whose core difference was in the level of CRE and were associated with significantly different survival outcomes (Supplementary Figure 1 and Supplementary Table 1), thereby validating the plausibility of CRE level as a primary stratification metric from an unsupervised learning perspective.

Figure 3 Interpretation of the XGBoost model using Shapley Additive exPlanations analysis on the training set. A: Shapley Additive exPlanations (SHAP) summary plot illustrating the global importance and impact of the top 20 predictors. Each point represents a patient for a given feature, with the color indicating the feature's value (red for high, blue for low); B: SHAP beeswarm plot providing a detailed distribution of each feature's impact on the model output. CRE: Creatinine; TBIL: Total bilirubin; HE: Hepatic encephalopathy; PTS: Prothrombin time seconds; CHE: Cholinesterase; PTA: Prothrombin activity; GGT: Gamma-glutamyl transferase; AFP: Alpha-fetoprotein; FE: Iron; ALB: Albumin; INR: International normalized ratio; AST: Aspartate aminotransferase; HGB: Hemoglobin; ALP: Alkaline phosphatase; SHAP: Shapley Additive exPlanations.

Construction and validation of the two-step risk card

Based on the central role of CRE level, we first identified its optimal cut-off point (62 μmol/L) in the training set and categorized patients into high-CRE (≥ 62 μmol/L) and low-CRE (< 62 μmol/L) subgroups. This one-level stratification strategy was effective in differentiating the risk of death in both the training and test sets (Figure 4A).

Figure 4 Development of the two-step hierarchical stratification strategy. A: Mortality rates for the primary stratification based on serum creatinine (CRE) level in the training and test sets; B: Shapley Additive exPlanations feature importance plots for the high-CRE and low-CRE subgroups, both identifying total bilirubin (TBIL) level as the top predictor; C: Mortality rates for the secondary stratification by TBIL level within the high-CRE subgroup; D: Mortality rates for the secondary stratification by TBIL level within the low-CRE subgroup. CRE: Creatinine; TBIL: Total bilirubin.

Subsequently, to enable finer risk assessment within subgroups of different renal functional status, we performed separate SHAP subgroup analyses for the high- and low-CRE subgroups. The results showed that TBIL level was the most important predictor of mortality in both subgroups, despite differences in baseline risk and some secondary predictors between the subgroups (Figure 4B). Therefore, we performed secondary stratification within each CRE subgroup based on TBIL level, using optimal thresholds determined within each subgroup (≥ 127 μmol/L for the high-CRE subgroup and ≥ 144 μmol/L for the low-CRE subgroup). Secondary stratification further enhanced risk differentiation, with mortality in the high-TBIL subgroup being significantly higher than that in the low-TBIL subgroup, in both the high- (Figure 4C) and low-CRE subgroups (Figure 4D). This effect remained consistent across the training and test sets.

Final risk assessment

The two-step strategy was ultimately consolidated into a concise and intuitive four-group classification. We performed the final validation of this assessment card in an independent test set. The Kaplan-Meier survival analysis showed that the tool successfully classified patients into four risk groups with very different prognoses (global Log-rank P < 0.001; Figure 5A). The difference in 180-day mortality among the four groups was significant, rising from 5.1% in the lowest risk group to 51.6% in the ultra-high-risk group. DCA further confirmed the clinical utility of this risk assessment card, which consistently outperformed the traditional strategy in terms of net benefit over a wide range of clinical decision thresholds (Figure 5B). Detailed performance metrics of the optimal model vs traditional scoring systems are shown in Table 3.

Figure 5 Validation of the four-group risk stratification. A: Kaplan-Meier survival curves for the four distinct risk groups. The P value was calculated using the log-rank test; B: Decision curve analysis demonstrating the clinical net benefit of the four-group stratification model across a range of threshold probabilities. CRE: Creatinine; TBIL: Total bilirubin.

Table 3 Performance of the best-machine learning model and traditional scores in the training and test sets.

Dataset	Model	AUC	Accuracy	Sensitivity	Specificity	Kappa
Test set	Best-ML (XGBoost)	0.848 (0.813-0.883)	0.785	0.719	0.824	0.541
Training set	Best-ML (XGBoost)	0.837 (0.795-0.880)	0.752	0.733	0.764	0.491
Test set	MELD	0.821 (0.784-0.859)	0.715	0.374	0.917	0.323
Training set	MELD	0.804 (0.757-0.851)	0.693	0.397	0.895	0.314
Test set	MDF	0.787 (0.745-0.829)	0.709	0.380	0.903	0.312
Training set	MDF	0.713 (0.658-0.769)	0.643	0.305	0.874	0.195

AUC: Area under the curve; ML: Machine learning; MDF: Maddrey discriminant function; MELD: Model for end-stage liver disease.

DISCUSSION

Patients with MDF scores ≥ 32 exhibit rapid disease progression, treatment resistance, and poor prognosis, with a 30-day natural mortality rate as high as 20%-50%[13]. In China, viral hepatitis remains the predominant type of liver disease; however, owing to widespread vaccination and improved living standards, the incidence of ALD has risen steadily over the past decade[14]. Consequently, the diagnosis and treatment of ALD have garnered increasing attention. Although Western countries exhibit high ALD prevalence and extensive research has been conducted in this field, clinical characteristics may differ significantly owing to factors such as ethnicity and drinking habit[15]. Currently, China lacks comprehensive, large-scale prognostic studies in patients with SAH. In recent years, the advancements of AI and ML technologies have resulted in the development of novel computational approaches to predict prognosis by integrating high-dimensional clinical data and modeling nonlinear relationships. Integrating advanced computational methods with clinical practice to provide clinicians with simple, intuitive prognostic tools holds substantial practical value.

The European Association for the Study of the Liver guidelines for ALD lists five predictive models for assessing the short-term prognosis of patients with alcoholic hepatitis: MDF, MELD, Glasgow Alcoholic Hepatitis Score, Age-Bilirubin-International Normalized Ratio-Creatinine score, and Lille score[1]. A 2024 global multicenter study employed RF, gradient-boosted machines, and XGBoost to construct an ensemble model (AI-integrated score for alcoholic hepatitis), achieving an AUC of 0.799 for predicting 90-day mortality—significantly outperforming traditional models (P < 0.001)[7]. Consistent with the findings of these studies, our research confirmed the advantages of ML. Our data indicate that a substantial proportion of fatalities occur beyond 90 days. Therefore, we selected 180 days as the evaluation endpoint to extend SAH management beyond the acute hospitalization phase to identify high-risk outpatients. Among the nine algorithms compared, the XGBoost model demonstrated the best performance, achieving an AUC of 0.848 on an independent test set—outperforming the clinically prevalent MELD and MDF scores. These findings suggest that advanced algorithms such as XGBoost can more effectively capture nonlinear relationships among factors within the complex pathophysiological network of SAH. Consequently, they may provide earlier warnings for the long-term management of high-risk patients, thereby strengthening the management of out-of-hospital patients.

However, purely predictive model performance improvement does not directly translate into optimization of clinical practice. More importantly, clinicians need to understand the decision logic behind the model and integrate data-driven findings with clinical experience. Therefore, the core innovation of this study lies in extracting interpretable clinical wisdom from the model. Through SHAP analysis, we identified serum CRE and TBIL levels as the core predictors driving 180-day survival in patients with SAH. These two predictors overlap with the components of the MELD score (TBIL level, CRE level, and international normalized ratio), confirming that our model captures the key to disease prognosis. More importantly, our analysis revealed the hierarchical relationship of these factors in different clinical contexts.

To validate the plausibility of CRE level as a primary stratification metric, we applied an unsupervised method of LCA[16,17]. The LCA naturally classified patients into two clinical phenotypes with significantly different survival prognoses without using any outcome information. The survival curves of these two groups of patients showed significant differences (group I, with a 180-day survival rate of 79%, and group II, with a survival rate of only 48%). The central distinguishing feature between these phenotypes was the CRE level. This finding strongly demonstrates that CRE level is not only a mere prognostic predictor, but also a key axis that delineates the core pathophysiologic status of patients with SAH, particularly the state of renal function. Clinically, this finding suggests that renal impairment may be the primary event determining the overall disease trajectory in patients with SAH.

Having established CRE level as the cornerstone of the primary stratification, we further explored the high- and low-CRE subgroups. The SHAP subgroup analysis revealed a clinically important phenomenon: Regardless of the presence or absence of significant renal impairment, TBIL level, which represents the severity of hepatic failure, was consistently the most critical factor influencing the prognosis of the patients. This finding prompted the construction of the final CRE-TBIL two-step stratification system. The system clearly categorizes patients into four risk classes, with those in the “high CRE-high bilirubin” group having a 180-day mortality rate of > 50%. These patients are likely to require the most aggressive interventions, close monitoring, and even early evaluation for liver transplantation. Unlike complex mathematical models, this intuitive risk identification card allows clinicians to quickly make a semi-quantitative assessment of a patient’s prognosis using two routine laboratory markers, which can directly guide treatment decisions.

Currently, the causes of severe hepatitis in China are hepatitis B virus infection and alcoholic liver damage[18]. However, patients with SAH are different from those with severe hepatitis B. Patients with severe hepatitis B mainly exhibit single-organ failure, namely liver failure, with jaundice and decreased coagulation mechanism as the main clinical manifestations, which are also the core prognostic indicators[19]. In contrast, patients with SAH mainly exhibit multi-organ failure, with renal dysfunction being the most important prognostic indicator[20]. However, advanced SAH is characterized by multiple organ failure, of which renal dysfunction is pronounced; our study further confirmed that renal impairment is the core prognostic factor. We found that the incidence of infection (including spontaneous peritonitis and pulmonary infection) was higher in the high-CRE group than in the low-CRE group (45% vs 18%) and that infections were likely to lead to the deterioration of renal function. Therefore, we hypothesized that “infection-kidney damage-death” is an important mechanism affecting the survival of patients with SAH for 180 days. In the case of SAH, preventing infection appears to be one of the most effective strategies to improve patient outcomes, although it is not the only determinant of prognosis.

This study has some limitations. First, the single-center retrospective design and the absence of an independent external validation cohort may limit the generalizability of our findings and pose a risk of overfitting. Second, owing to the lack of complete post-treatment dynamic data, we were unable to evaluate the Lille score in patients receiving hormone therapy or analyze in depth the effects of different treatment regimens on various risk subgroups. Future studies should increase the number of participants, and validate our risk assessment card in a multi-center prospective cohort. Studies should also explore whether the individualized treatment strategy based on this stratification can truly improve the clinical outcomes of patients.

CONCLUSION

In conclusion, by integrating interpretable ML with clinical phenotyping, we constructed a model with superior predictive performance and refined and validated a simple two-step stratification system based on CRE and TBIL levels. This system provides clinicians with a prognostic assessment tool that is fast, intuitive, and based on solid data evidence, which is expected to guide individualized and precise treatment of SAH.

ACKNOWLEDGEMENTS

We thank Long-Xin Guo and Sheng-Kai Zhu for their support in data analysis.