Machine learning-based risk factors for acute-on-chronic liver failure in alcohol-associated liver disease

Abstract

Background

In patients with alcohol-associated liver disease (ALD), heavy and prolonged alcohol consumption can trigger acute-on-chronic liver failure (ACLF), a condition associated with high early mortality and significant clinical challenges. Early identification of patients at risk is critical for improving outcomes.

Aims

We conducted a retrospective observational study of patients diagnosed with ALD between January 2000 and December 2024 to develop a predictive model for ACLF.

Methods

Key clinical indicators were selected using LASSO-regularized logistic regression (LR). The final LR model was visualized as a nomogram and compared with four additional machine learning algorithms. Model performance was evaluated using ten-fold cross-validation and area under the curve (AUC), while feature importance was assessed with Shapley Additive exPlanations values.

Results

Among 210 patients with ALD, LASSO identified four independent predictors of ACLF: total bilirubin (TBIL), folate, vitamin B12 (VitB12) and the difference from the normal value of prothrombin time (ΔPT). The LR model achieved an AUC of 0.970, indicating excellent predictive accuracy.

Conclusion

We developed a robust clinical prediction model combining LR and machine learning approaches. TBIL, folate, VitB12 and ΔPT are key prognostic markers that may enable early risk stratification and timely intervention, potentially reducing ACLF incidence in ALD patients.

Introduction

Alcohol-associated liver disease (ALD) is a metabolic disorder caused by chronic alcohol intake, with clinical manifestations ranging from reversible steatosis to advanced cirrhosis and hepatocellular carcinoma [1]. Over the past three decades, alcohol consumption among Chinese individuals has increased markedly, paralleled by rising numbers of hospitalized ALD patients and a steady increase in the annual hospitalization rate. Currently, ALD represents the second leading cause of end-stage liver disease in China, following viral hepatitis [2].

Acute-on-chronic liver failure (ACLF) frequently complicates the course of ALD. ACLF is characterized by acute hepatic decompensation in the setting of chronic liver disease and carries a 28-day mortality rate exceeding 50% [3]. Several scoring systems have been developed to predict outcomes in patients with alcohol-associated hepatitis (AH), including the Maddrey discriminant function (MDF), the Model for End-Stage Liver Disease (MELD) and the age-bilirubin-international normalized ratio-creatinine (ABIC) score. However, some of these models, such as MDF, demonstrated poor prognostic performance in the STOPAH trial [4]. Although these tools are widely used and easy to calculate, they were primarily developed in Western populations and may not be directly applicable to Asian patients. This underscores the urgent need for prognostic models that incorporate clinical and biomarker characteristics specific to Asian populations.

ALD represents a major and rapidly growing cause of advanced liver disease in Asia and has emerged as one of the leading aetiologies of ACLF. ALD-related ACLF is characterized by rapid clinical deterioration and high short-term mortality, underscoring the urgent need for early risk stratification in this vulnerable population. Despite its clinical importance, tools specifically designed to identify patients with ALD who are at highest risk of progressing to ACLF remain limited. Machine-learning approaches, which are capable of capturing complex nonlinear relationships among clinical variables, offer substantial potential to improve early identification of patients at high risk for ACLF [5,6]. In this study, we aimed to develop and externally validate a dynamic, machine-learning–based prediction model for ACLF in Chinese patients with ALD. By integrating routinely available clinical parameters with metabolic biomarkers, our goal was to facilitate individualized risk assessment, support timely clinical decision-making and ultimately improve outcomes in patients with ALD.

Methods

Study design

This retrospective cohort study included 518 patients hospitalized for alcoholic liver disease (ALD) at Peking University First Hospital between January 2000 and December 2024. The inclusion criteria were: (1) primary diagnosis of ALD; (2) age ≥18 years; and (3) availability of complete medical records, laboratory test results and other necessary medical documentation. Exclusion criteria were: (1) age <18 years; (2) pregnancy or lactation; (3) liver injury from other aetiologies (e.g. viral hepatitis, drug-induced liver injury, autoimmune hepatitis); (4) recurrent hospitalization or non-primary diagnosis of ALD; and (5) incomplete records or missing laboratory/clinical data. The patient enrolment process is shown in Figure 1.

Figure 1.

Flowchart of patient selection. A total of 210 patients with alcohol-associated liver disease (ALD) were included, of whom 89 developed acute-on-chronic liver failure (ACLF) and 121 did not. Note: 95%CI: 95% confidence interval.

Participants

Between January 2000 and December 2024, patients meeting the inclusion and exclusion criteria were screened. A total of 210 patients were ultimately enrolled. Of these, 121 patients (57.6%) did not develop ACLF during hospitalization, while 89 patients (42.4%) developed ACLF.

Outcome definition

The primary outcome was the development of ALD-related ACLF (ALD-ACLF) during hospitalization. The diagnosis of ACLF was based on the 2024 Chinese Medical Association Guidelines for the Diagnosis and Treatment of Liver Failure (2024 edition) [7]. Patients with chronic ALD who developed acute hepatic deterioration, defined by elevated bilirubin levels and coagulopathy, with or without extrahepatic organ failure, were classified as having ALD-ACLF.

Data collection

Clinical data were retrieved from the institutional electronic health record system. Baseline demographic and anthropometric variables included age, sex, height, weight and body mass index (BMI). Alcohol consumption history was collected, including presence or absence of alcohol use and length of alcohol consumption (LOAC). The daily amount of alcohol consumption (AOAC, g/day) was calculated using the formula: AOAC (g) =AOAC (mL)×alcohol content (%)×0.8. Baseline laboratory parameters included white blood cell count (WBC), neutrophil count (NE), lymphocyte count (LY), haemoglobin (HGB), platelet (PLT), mean corpuscular volume (MCV), aspartate aminotransferase (AST), alanine aminotransferase (ALT), alkaline phosphatase (ALP), gamma-glutamyl transferase (GGT), total bilirubin (TBIL), serum creatinine (CREA), albumin (ALB), folate, vitamin B12 (VitB12), prothrombin time (PT), the difference between PT and normal values (ΔPT), fibrinogen (FIB) and international normalized ratio (INR). Clinical comorbidities at baseline, including ascites and hepatic encephalopathy (HE), were also recorded.

Model construction

To overcome the racial and regional limitations of existing predictive models, we constructed a dynamic prediction system based on multimodal machine learning. Unlike traditional logistic regression models that assume linear associations, machine learning approaches can capture nonlinear feature interactions and detect early signals of organ dysfunction [5,6]. Given the modest sample size, we systematically evaluated five algorithms: logistic regression (LR), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), support vector machine (SVM) and k-nearest neighbours (KNN).

Each algorithm has unique strengths. LR estimates event probabilities through linear predictor combinations transformed by the logit function. XGBoost constructs decision trees with built-in regularization to mitigate overfitting, while LightGBM applies histogram-based learning and leaf-wise tree growth for faster computation and lower memory usage. SVM identifies optimal hyperplanes for binary classification and is particularly suited for small datasets and nonlinear decision boundaries. KNN classifies new cases according to the labels of their closest neighbours in multidimensional space.

We trained the LR model using a 70/30 training-validation split with ten-fold cross-validation. LASSO regression was employed for variable selection, and the penalty parameter (λ) was optimized via cross-validation using the ‘glmnet’ package in R. A multivariate LR model was then constructed with significant predictors, and the final formula was visualized as a nomogram. In parallel, four additional machine learning models (XGBoost, LightGBM, SVM and KNN) were developed. To address class imbalance between ACLF and non-ACLF patients, the Borderline-SMOTE technique was applied to oversample the minority class. Concurrently, we performed a comparative analysis between the newly developed model and established scoring systems, including MDF, MELD, ABIC and Child–Pugh Scores.

An ensemble framework was developed by integrating LR with the four machine learning algorithms. Model interpretability was enhanced using Shapley Additive exPlanation (SHAP) values, which quantified the contribution of each variable to model predictions [8,9]. Importantly, folate and VitB12 levels, along with coagulation parameters, were incorporated into the models as novel metabolic predictors. This integrative framework provided not only improved predictive performance but also biological insight into ALD-ACLF risk.

Model evaluation

Model performance was comprehensively evaluated across multiple dimensions. Discrimination was quantified by the area under the curve (AUC), reflecting the model’s ability to distinguish between patients who developed ACLF and those who did not. Calibration was assessed using calibration plots and the Hosmer–Lemeshow goodness-of-fit test to compare predicted versus observed outcomes. Clinical utility was examined with decision curve analysis (DCA), which estimated the net benefit across a range of clinically relevant thresholds. In addition, overall performance metrics, including accuracy, sensitivity, specificity and the F1 score, were calculated to provide a balanced assessment of predictive ability. Finally, SHAP analysis was used to interpret the relative contribution of each predictor, thereby improving transparency and clinical applicability of the models.

Statistical analysis

We examined the association between baseline clinical and biochemical parameters in ALD patients and the development of ACLF. Descriptive statistics were used to summarize baseline characteristics. Continuous variables were first assessed for normality using the Kolmogorov–Smirnov test. Normally distributed data are expressed as mean ± standard deviation (SD), while skewed data are presented as median with interquartile range (M [Q₁–Q₃]). Categorical variables are reported as frequency and percentage (n [%]). Multicollinearity among covariates was evaluated using the variance inflation factor (VIF), with values <10 considered acceptable. For between-group comparisons, independent-sample t tests were applied to normally distributed variables, Mann–Whitney U tests to non-normally distributed variables and chi-square or Fisher’s exact tests to categorical data. A two-sided p value <.05 was considered statistically significant. Analyses were performed using SPSS version 25.0 (IBM Corp., Armonk, NY) and R software.

Results

Baseline characteristics of patients

We summarized the baseline clinical characteristics of 210 patients with ALD, stratified into non-ACLF and ACLF groups. The cohort consisted of 202 men (96.2%) and 8 women (3.8%) (Supplementary Table 1). Comparative analyses revealed significant differences in age, presence of ascites, HE and length of alcohol consumption (LOAC) between the two groups (p < .05). Laboratory parameters were consistently higher in the ACLF group compared with the non-ACLF group. Specifically, serum AST, GGT, ALP, WBC, NE/LY, NE%, PLT, MCV, HGB, Folate, VitB12, ALB, TBIL, CREA, INR and PT levels were all significantly elevated in patients with ACLF (p < .05).

Feature selection

To address multicollinearity, VIF analysis was performed, and variables with VIF >10 were excluded. These included PT (VIF = 614.833), INR (VIF = 186.788), MCV (VIF = 44.702), age (VIF = 40.872), NE (VIF = 33.078), BMI (VIF = 27.196) and sex (VIF = 18.082). Subsequently, we examined potential risk factors such as ascites, AST/ALT ratio, ALP, GGT, TBIL, WBC, HGB, CREA, Folate, VitB12, LOAC and ΔPT. LASSO regression was applied to identify the most relevant predictors of ACLF. Twelve candidate variables were initially included, and the optimal λ value was determined using 10-fold cross-validation (Supplementary Figures 1A and 1B). The final variables retained in the model were TBIL, Folate, VitB12 and ΔPT. Both univariate and multivariate regression analyses confirmed that these four factors were independent risk predictors for ACLF in ALD patients (Figure 2, Supplementary Table 2).

Figure 2.

Forest plot showing the results of univariate and multivariate regression analyses of risk factors for acute-on-chronic liver failure (ACLF) in patients with alcohol-associated liver disease (ALD). Total bilirubin (TBIL), folate, vitamin B12 and ΔPT were identified as independent predictors of ACLF, highlighting their potential value in risk assessment.

Logit(P) = 0.006 × TBIL − 0.037 × Folate + 0.003 × ViB12 + 0.262 × ΔPT.

Development and validation of the nomogram

Using these independent predictors, we constructed a nomogram model to estimate the risk of ACLF in ALD patients (Figure 3). Binary logistic regression was applied, yielding the following predictive formula: logit (P) = 0.006 × TBIL − 0.037 × Folate + 0.003 × VitB12 + 0.262 × ΔPT. Then obtain the predicted probability of the target event for that patient using the P = e^logit(P)/[1 + e^logit(P)].

Figure 3.

Nomogram for predicting the risk of acute-on-chronic liver failure (ACLF) in patients with alcohol-associated liver disease (ALD). The model was developed using a training dataset and incorporates total bilirubin (TBIL), folate, vitamin B12 and ΔPT as predictive parameters.

The dataset was randomly split into training (n = 147) and validation (n = 63) cohorts. The nomogram was internally validated using bootstrap resampling with 1000 iterations. Calibration plots demonstrated good concordance between predicted and observed probabilities, with minimal deviation from the ideal line (Figure 4(A,B)). Discriminatory performance was excellent: the AUC was 0.968 (95% CI: 0.939–0.997) in the training set and 0.970 (95% CI: 0.918–1.000) in the validation set (Figure 5(A,B)). Clinical utility was further confirmed using DCA, which showed a favourable net benefit across a wide range of threshold probabilities (Figure 6). Finally, robustness was confirmed by 10-fold cross-validation, producing AUC values of 0.985 in the training cohort and 0.981 in the validation cohort (Supplementary Figures 2A and 2B).

Figure 4.

Calibration curves for the nomogram predicting progression of ALD to ACLF. Panel A: training set; Panel B: validation set. The apparent and bias-corrected curves closely follow the ideal line, indicating strong agreement between predicted and observed probabilities.

Figure 5.

Receiver operating characteristic (ROC) curves assessing the predictive performance of the nomogram for acute-on-chronic liver failure (ACLF) in patients with alcohol-associated liver disease (ALD). Panel A: training set (AUC = 0.968); Panel B: validation set (AUC = 0.970). High and comparable AUCs indicate consistent predictive accuracy across datasets.

Figure 6.

Decision curve analysis (DCA) showing the net benefit of predicting progression from alcohol-associated liver disease (ALD) to acute-on-chronic liver failure (ACLF) across a range of threshold probabilities. The x-axis represents threshold probability (%), and the y-axis represents mean net benefit. The curve demonstrates the high clinical utility of the logistic regression model compared with treating all or no patients.

Overall, the nomogram demonstrated high accuracy, robustness and clinical applicability, enabling precise risk stratification of ACLF in patients with ALD and supporting optimized management and treatment strategies.

Development of predictive models using machine learning

To improve early diagnosis and intervention for ALD-ACLF, we constructed five machine learning models using the four key clinical indicators identified (TBIL, Folate, VitB12 and ΔPT): LR, XGBoost, LightGBM, SVM and KNN. In the training cohort, XGBoost (AUC = 0.980) and KNN (AUC = 0.989) demonstrated the highest performance, outperforming LR in terms of accuracy, specificity and other evaluation metrics (Supplementary Figures 3A and Supplementary Table 3). However, when applied to the validation cohort, LR achieved the best overall performance with an AUC of 0.958 and accuracy of 0.919 (Supplementary Figures 3B and Supplementary Table 4). Calibration plots and decision curve analysis further supported LR’s stability and clinical utility (Supplementary Figures 4A and 4B). In contrast, KNN showed evidence of overfitting, performing well in training but less consistently in validation. Taken together, LR provided the most robust and reliable model for predicting ACLF risk in this dataset.

In this study, we compared the newly developed model with established scoring systems, including MDF, MELD, ABIC and Child–Pugh, using receiver operating characteristic (ROC) analysis, calibration curves and decision curve analysis. The new model demonstrated superior discriminative performance (AUC = 0.960), with higher sensitivity and specificity than the MDF (AUC = 0.953), MELD (AUC = 0.930), ABIC (AUC = 0.776) and Child–Pugh (AUC = 0.947) models (Supplementary Figure 5). In addition, calibration and DCA indicated that the new model exhibits high stability and greater clinical utility, providing effective support for clinical decision-making and highlighting its advantages over traditional scoring systems (Supplementary Figures 6 and 7).

Assessment of variable importance

To enhance model interpretability, SHAP values were applied to quantify the contribution of individual variables to model predictions. The beeswarm plot (Supplementary Figure 8A) illustrates depicts the distribution of SHAP values for each feature, with each dot representing an individual prediction. ΔPT exhibited the widest range of SHAP values, indicating the strongest influence on ACLF risk. Higher ΔPT values (shown in red) were strongly associated with increased ACLF risk, as were elevated vitamin B12 and TBIL levels. In contrast, lower folate levels were negatively associated with ACLF outcomes. The SHAP summary bar plot (Supplementary Figure 8B), which ranks variables by mean absolute SHAP values, further confirmed ΔPT and vitamin B12 as the most influential predictors, followed by TBIL and folate. Collectively, these findings provide mechanistic insight into model behaviour and underscore the prominent roles of coagulation dysfunction and metabolic derangements in predicting ACLF.

Discussion

ACLF, characterized by multiorgan dysfunction and rapidly escalating mortality, represents an acute-on-chronic decompensation event in the natural history of chronic liver disease [10]. Aetiologies such as viral hepatitis, ALD and autoimmune liver disease can all serve as the underlying substrate. ALD, a globally prevalent cause of chronic liver injury, can progress to ACLF due to sustained ethanol exposure, leading to irreversible hepatocyte damage. Recent studies highlight the contributions of systemic inflammation, immune dysregulation, gut microbiota disruption and mitochondrial dysfunction to the transition from ALD to ACLF [11]. Observational data further demonstrate an increasing proportion of ACLF cases induced by severe AH among ICU admissions, underscoring the association with adverse clinical outcomes [12].

Beyond ethanol toxicity, oxidative stress and genetic factors, methionine-cycle disruption has emerged as a pivotal mechanism in ALD pathogenesis [13]. ALD is increasingly conceptualized as a protein-calorie malnutrition syndrome exacerbated by deficits in folate, Vit B6, and particularly Vit B12 [14]. In our study, SHAP analysis confirmed VitB12 as one of the most influential predictors of ACLF risk, consistent with mechanistic links to NF-κB/TNF-α overactivation and hepatic decompensation [15]. Vitamin B12 in liver disease serves a dual role: it may function as a deficient micronutrient while also acting as a biomarker of hepatocellular injury, systemic inflammation and metabolic dysregulation. Specifically, hepatocellular damage impairs hepatic vitamin B12 storage capacity, resulting in increased release of vitamin B12 into the circulation, predominantly in the form of holotranscobalamin, and consequently elevated serum vitamin B12 levels [15]. Furthermore, multiple clinical studies have demonstrated that serum vitamin B12 concentrations are significantly associated with disease severity and prognostic outcomes in patients with ACLF [16]. These findings reinforce the concept of vitamin metabolic reprogramming in liver disease and underscore the need for refined detection methods to assess nutritional status more accurately in ALD. Folate is another critical water-soluble vitamin, facilitating one-carbon transfer for purine and pyrimidine synthesis [17]. Folate deficiency triggers proinflammatory cytokine release, disrupts hepatic lipid homeostasis and promotes steatosis and fibrosis [17]. Low folate levels are epidemiologically associated with MASLD, viral hepatitis and other liver disorders [17]. In ALD, alcohol intake reduces dietary folate, impairs absorption due to mucosal injury and decreases hepatic storage and tetrahydrofolate production [17]. While prior studies have highlighted folate deficiency in liver disease, its role in ALD-to-ACLF progression has been underexplored. Our study incorporated folate as a predictive factor, filling an important gap and emphasizing the role of metabolic derangements in ALD.

Coagulation function plays a central role in liver disease pathophysiology. The liver regulates haemostasis through the synthesis of clotting and anticoagulant factors, fibrinolysis and clearance of activated enzymes [18]. In chronic liver disease, inflammatory and coagulation pathways are closely interconnected, resulting in reduced factor synthesis, impaired clearance, altered platelet function and enhanced fibrinolysis, which together create a fragile balance between bleeding and thrombosis [19,20]. PT, a simple yet informative index, reflects liver synthetic capacity and inflammation [21]. Nomograms incorporating ALP, GGT and PT have been shown to predict inflammation severity in chronic liver disease [22]. Moreover, coagulation indices independently predict ACLF risk, reflecting impaired factor synthesis [23]. In this study, our integrative analysis identified ΔPT as a predictive marker of ALD-to-ACLF progression, highlighting the clinical utility of coagulation parameters in disease stratification.

TBIL remains one of the most reliable markers of hepatic metabolic and excretory capacity and is integral to the Child-Pugh classification. Elevated TBIL correlates with liver injury severity and short-term mortality [24]. Prior studies demonstrate TBIL as an independent predictor of 90-day transplant-free survival in cirrhosis [24], and dynamic changes in TBIL are highly relevant to ACLF prognosis [25]. In patients with HBV-related ACLF, TBIL has been incorporated into prognostic models, highlighting its central role [26]. Our findings extend this evidence by confirming TBIL as a critical predictor in ALD-to-ACLF progression.

With the digital transformation of healthcare, machine learning offers new opportunities for risk stratification. We compared five algorithms (LR, XGBoost, LightGBM, SVM, KNN) for predicting ACLF under sample-size constraints. LR provided the most stable performance, combining transparency, interpretability and computational efficiency. Its calibration demonstrated strong agreement between predicted and observed probabilities, supporting bedside applicability and potential integration into transplant prioritization. Although KNN achieved high specificity (96.7%) and positive predictive value (94.6%), it showed signs of overfitting. In contrast, LR achieved the best balance with sensitivity of 88.8%, accuracy of 91.9% and AUC of 0.958, reducing the likelihood of missed diagnoses. Decision curve and calibration analyses further validated LR’s clinical utility. XGBoost offered complementary strengths by capturing complex biomarker interactions, with an F1 score of 87.7%. While SVM and LightGBM performed comparably, they lacked clear advantages. Robustness across models was confirmed through cross-validation, with LR demonstrating the highest consistency. This comparative framework suggests LR as optimal for rapid bedside triage, XGBoost as a tool for specialized centres and KNN as a potential approach in resource-limited settings.

During the validation phase, model performance was comprehensively evaluated using ROC analysis, calibration curves and decision curve analysis. The results demonstrate that the new model exhibits high sensitivity, specificity and stability, providing effective support for clinical decision-making and highlighting its advantages over traditional scoring systems.

Despite the strengths of this study, several limitations should be acknowledged. First, the retrospective design and reliance on electronic health records introduce the potential for incomplete or inaccurate data, resulting in information bias. Second, the single-centre nature of the cohort enhances internal consistency but limits the generalizability of the findings. Third, although LASSO regression was applied to reduce overfitting, certain relevant biomarkers may not have been captured or included in the model. Fourth, this study did not systematically evaluate the relationship between inflammatory markers and vitamin B12 levels, nor did it fully account for comorbid conditions that may influence vitamin B12 concentrations; these issues warrant further investigation in future studies.

In summary, we developed and validated a novel risk-stratification model for ALD-ACLF using LASSO combined with logistic regression and compared its performance with XGBoost, LightGBM, SVM and KNN. Key predictors included TBIL, Folate, VitB12 and ΔPT. A nomogram was constructed for intuitive bedside use, highlighting the prognostic importance of metabolic and coagulation pathways in ALD. SHAP analysis clarified variable importance and interactions, supporting personalized risk assessment. While alternative algorithms demonstrated complementary strengths, LR emerged as the most clinically practical model. This framework offers an early prediction tool to support preventive strategies and optimize patient management. Future studies should pursue large-scale, multicentre external validation, precisely quantify the interval between acute deterioration and ACLF onset and incorporate novel biomarkers and inflammatory indices to further enhance clinical applicability and patient outcomes.

Funding Statement

This research received no funding from any agency in the public, commercial, or non-profit sectors.

Ethical approval

This retrospective study of de-identified patient records was approved by the Institutional Review Board of Peking University First Hospital (IRB No. 2024-693-001) and was conducted in accordance with the ethical principles outlined in the Declaration of Helsinki (6th revision, 2008). Given the retrospective design and minimal risk to participants, the requirement for informed consent was waived by the IRB.

Consent form

The IRB of Peking University First Hospital waived the requirement for informed consent.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The datasets used and/or analysed during the current study are available from the corresponding author upon reasonable request.