Individual Mortality Risk Predictive System of ACLF Based on RSF Model

0
0
0

Individual Mortality Risk Predictive System of Patients with Acute-on-Chronic Liver Failure Based on a Random Survival Forest Model

Introduction
Acute-on-chronic liver failure (ACLF), characterized by rapid deterioration of liver function in patients with chronic liver diseases, carries a high short-term mortality rate of 60%–70% within three months. Current prognostic models, such as the Model for End-Stage Liver Disease (MELD), International Normalized Ratio and Creatinine Score (ABIC), and Integrated MELD (iMELD), offer group-level mortality predictions but fail to provide individualized risk assessments for clinical decision-making. The Random Survival Forest (RSF) algorithm, a nonparametric machine learning method, addresses these limitations by handling nonlinear relationships, variable interactions, and missing data effectively. This study aimed to develop an online individual mortality risk predictive tool for ACLF patients using RSF, enabling dynamic mortality risk curves and personalized survival predictions.

Study Design and Methods
Patient Cohorts and Data Collection
A retrospective analysis enrolled 391 ACLF patients from three Chinese hospitals (Shunde Hospital, Jiangmen Central Hospital, and The First People’s Hospital of Foshan). After excluding patients with incomplete data, comorbidities (e.g., liver cancer, autoimmune diseases), or insufficient follow-up, 276 patients formed the model cohort. A validation cohort (n=276) was generated using bootstrap resampling to assess model generalizability.

Diagnostic Criteria and Variables
ACLF diagnosis followed Asian Pacific Association for the Study of the Liver guidelines, with complications like hepatic encephalopathy (HE), acute kidney injury (AKI), and hepatorenal syndrome (HRS) defined using established criteria. Clinical variables included demographics, laboratory parameters (e.g., serum sodium, INR, RDW), and disease-specific scores (MELD, ABIC, iMELD).

Model Development and Validation

  1. Cox Proportional Hazards Model: Identified independent predictors via stepwise multivariate regression.
  2. RSF Model: Constructed using variables from the Cox analysis (HE, age, serum sodium, AKI, RDW, INR). The RSF algorithm’s ensemble survival trees estimated survival probabilities, with variable importance ranked through permutation.
  3. Performance Metrics:
    • Time-dependent ROC curves: Calculated for 3-, 6-, and 12-month mortality.
    • Brier scores: Evaluated calibration accuracy (lower scores indicate better performance).
    • Decision curve analysis: Compared clinical utility across models.

Online Predictive Tool
A web-based tool (hosted at https://zhangzhiqiao13.shinyapps.io/Individual_mortality_risk_predictive_tool_for_liver_failure/) was developed to generate individualized mortality curves, predicted survival percentages, and 95% confidence intervals at user-defined time points.

Key Findings
Variable Importance and Risk Factors
The RSF algorithm ranked AKI, HRS, HE, age, RDW, and INR as top predictors (Figure 1). Multivariate Cox regression confirmed these as independent mortality risk factors (Table 2):

  • HE: HR=2.408 (95% CI:1.624–3.571, P<0.001).
  • Age: HR=1.035 per year (95% CI:1.021–1.050, P<0.001).
  • AKI: HR=3.289 (95% CI:2.170–4.985, P<0.001).
  • INR: HR=1.897 (95% CI:1.530–2.350, P<0.001).
  • Serum sodium: Protective effect (HR=0.978 per mmol/L, 95% CI:0.959–0.998, P=0.027).
  • RDW: Protective effect (HR=0.974 per fL, 95% CI:0.963–0.986, P<0.001).

Model Performance

  1. RSF vs. Cox Model:

    • AUROC (Model Cohort): RSF outperformed Cox for 3-month (0.916 vs. 0.872), 6-month (0.916 vs. 0.866), and 12-month (0.905 vs. 0.848) predictions.
    • AUROC (Validation Cohort): RSF maintained superiority (0.912, 0.910, and 0.880 for 3-, 6-, and 12-month predictions, respectively).
    • Brier Scores: RSF demonstrated better calibration (3-month: 0.119 vs. 0.138; 12-month: 0.128 vs. 0.156).

  2. Comparison to Traditional Scores:
    RSF surpassed MELD, ABIC, and iMELD across all metrics. For example, MELD’s 3-month AUROC was 0.683, significantly lower than RSF’s 0.916.

Survival Stratification
Patients stratified into high- and low-risk groups by RSF showed significant survival differences (P0.5) had median survival of 2.3 months vs. 34.6 months in the low-risk group.

Online Tool Outputs
The tool provides two survival curves (RSF and Cox predictions) alongside numerical predictions (e.g., 12-month survival: 58% [95% CI:52–64%] for a 45-year-old patient with HE, AKI, serum sodium=135 mmol/L, INR=2.0, RDW=45 fL).

Discussion
Advantages of the RSF Model

  1. Individualized Predictions: Unlike traditional scores, the RSF model generates dynamic survival curves, enabling tailored clinical decisions.
  2. Handling Complex Interactions: RSF accounts for nonlinear relationships (e.g., age and serum sodium) and variable interdependencies (e.g., AKI and HE).
  3. Robust Validation: Bootstrap resampling ensured model stability, with consistent performance in the validation cohort.

Clinical Implications

  • Early Intervention: High-risk patients (identified via the online tool) may benefit from prioritized liver transplantation or intensive care.
  • Resource Allocation: Hospitals can optimize bed allocation and monitoring based on individualized risk.

Limitations

  1. Retrospective Design: Potential selection bias despite rigorous exclusion criteria.
  2. Lack of External Validation: Future studies should test the model in diverse populations.
  3. Missing Variables: Thyroid function and imaging-based liver volumetry were excluded due to incomplete data.

Future Directions

  • Prospective Multicenter Studies: Validate the tool across global cohorts.
  • Integration with Omics Data: Enhance predictions using genomic or metabolomic biomarkers.

Conclusion
This study developed an RSF-based online tool that provides individualized mortality risk curves and time-specific survival probabilities for ACLF patients. By outperforming traditional prognostic models, the tool enhances clinical decision-making, offering a paradigm shift from population-level to personalized risk assessment.

doi.org/10.1097/CM9.0000000000001539

Was this helpful?

0 / 0