🚨 TRANSPLANT INFORMATICS 🚨

Introduction:

Globally the trend of Liver Transplantation(LT) rates are increasing and the global transplant demand is expected to rise by nearly 25% by 2040. It is mostly influenced by the increased end stage liver disease prevalence, yet the post-LT survival rate, in one year and within five years, is 90% and 70% respectively which is due to persistent Biliary Complications(BC) as well as increased post operative mortality. The incidence of BC is  5-20% after transplantation and is the major cause of death. The liver transplantation is dependent upon the surgeons, which exponentially increases the uncertainty in outcomes.  

Presently in order to address these issues and increase the post operative accuracy, Artificial Intelligence (AI) as well as Machine Learning (ML) models are inculcated in surgical setup. Many studies reflect that ML models outperforms traditional models in predicting LT complications. This study used survival analysis to build and benchmark ML models for predicting post-LT mortality and biliary complications.

Methods:

A prospective cohort of pediatric and adult LT recipients at Shiraz Transplant Center (March 2018 to March 2023, median follow-up 19 months) was studied, excluding multi-organ transplants and incomplete follow-up (n=96). Forty key donor/recipient variables—including demographics, comorbidities, perioperative data, immunosuppression and early labs (excluding MELD/PELD components) were analyzed.

Two primary outcomes were studied: 

• BC: Identified through clinical evaluation, labs, and imaging, requiring intervention. 

• Mortality: Documented from patients medical records. 

Data Preprocessing: Missing values (<10%) were filled using mean or median. Continuous data were scaled, and categorical data were one-hot encoded. 

Feature Selection: Four methods (Cox-P, Cox-c, RSF, LASSO) identified the top 20 features. Models using all features served as a baseline.

Model Development, Evaluation & Interpretation:

• Seven survival ML models (LASSO, Ridge, RSF, E-NET, GBS, C-GBS, FS-SVM) and Cox regression were trained with 5-fold cross-validation, random over-sampling & hyperparameter tuning.

• Models were ranked by average C-index and stability; top performers visualized via heatmaps.

• Feature importance was assessed for best models: Ridge (biliary complications) via standardized coefficients; RSF (mortality) via permutation importance. Scores were normalized for comparison.

• Stability, indicating sensitivity to training data changes, was measured by standard deviation of C-index; higher stability implies more reliable models.

Results:

• Dataset: 1799 LT patients (21% pediatric), mostly deceased-donor grafts; biliary stricture was the most common complication.
  BC Prediction: Best: Ridge + RSF selection (C-index 0.699). RSF selection outperformed others; Cox PH and C-GBS also performed strongly.

• Mortality Prediction: Best: RSF + RSF selection (C-index 0.784). RSF selection ranked highest; Cox PH, Ridge, GBS, C-GBS competitive.
  Top Predictors: BC: graft type, recipient IBD, BMI, PVT history, prior LT. Mortality: post-transplant AST, creatinine, age, ALT, tacrolimus use.

• Model Stability: BC best model SD: 0.035 (range 0.027–0.050). Mortality best model SD: 0.013 (range 0.010–0.044). 

Key Predictors of Biliary Complications (BC)

• Liver Graft Type: Split grafts increase BC risk due to bile leakage.

• Inflammatory Bowel Disease (IBD): Associated with BC via Primary Sclerosing Cholangitis, worsened by immune and microbiome factors.

• Recipient BMI: Higher BMI raises BC risk by ~3% per unit, linked to fatty liver, poor blood flow, inflammation, and surgical challenges.

• Previous Portal Vein Thrombosis (PVT): Recurrence impairs biliary blood flow, causing ischemia and damage.

• Previous Liver Transplant: Raises BC risk from repeated ischemia, complex surgery, and stronger immunosuppression.

Key Predictors of Mortality:

• Top Factors: Immediate post-transplant AST, creatinine, recipient age, ALT, and tacrolimus use.

• Liver Enzymes: Elevated AST/ALT signal graft injury or rejection, increasing mortality risk.

• Creatinine: Post-LT renal impairment strongly predicts mortality.

• Tacrolimus: Effective immunosuppression lowers rejection but requires balance to limit side effects.

• Age: Older recipients face higher mortality due to comorbidities.

Other Factors influencing Liver Transplantations:

• Longer cold ischemia time increases graft loss risk (~3.4% per hour).

• High graft steatosis worsens ischemic damage and enzyme elevation.

• Close monitoring of post-transplant labs is vital to reduce mortality.

Strengths

• First study to use survival ML models for predicting biliary complications (BC) after LT.

• Identified key predictors for BC and mortality to support clinical decision-making.

• Applicable to routine follow-up visits for early risk detection and intervention.

Limitations

• Missing key variables (e.g., T-tube use, graft steatosis, Rh mismatch).

• Single centered study so no external dataset for model validation.

• Retrospective design may involve unmeasured confounders.

• Focused only on BC within 5 years; excluded long-term outcomes. 

• No long-term lab trends (e.g., tacrolimus levels).

• Combined adult and pediatric data; separate models may provide deeper insights. 

• Insufficient data to analyze BC subtypes individually.

Conclusion:

Survival ML models demonstrated high prognostic accuracy for post-LT biliary complications and mortality, supporting their utility in risk stratification and individualized management. Integrating such models into liver transplantation, a critical domain of advanced surgical care can optimize outcomes through data-driven clinical decision-making.