FRI-484 machine learning in predicting 3-month mortality for liver transplant candidates with HCC: a paradigm shift
Résumé
Background and aims: Predicting waitlist mortality in liver transplant (LT) candidates with hepatocellular carcinoma (HCC) is both critical and challenging. Traditional scores, like MELD scores, assess liver disease but ignore tumor risks, while HCC-specific models such as AFP model focus on tumors but overlook liver failure. This creates an unmet clinical need for an approach that evaluates both risks to optimize prioritization for transplantation. We address this gap using machine learning (ML) to predict 3-month mortality by combining liver dysfunction and tumor progression features. Method: We used the OPTN/UNOS data registry to compare ML models with traditional scores to predict 3-month mortality in LT candidates with HCC. Inclusion criteria were adult patients listed for LT who (i) either died or were delisted due to clinical deterioration within 3 months, or (ii) survived beyond 3 months. Patients listed for multi-organ transplants were excluded. The cohort included 25 clinical, laboratory, and disease-related variables, with an additional 6 dynamic variables derived from changes in laboratory values, leading to 31 variables. Logistic Regression and Random Forest (RF) models were assessed using both feature sets. To address class imbalance, a down-sampling strategy was applied, and performance was assessed using 3-fold cross-validation. Results: 11,641 patients were included, of whom 448 died or were removed from the waitlist within 3 months (mean: age 60, raw MELD 22) and 11,193 surviving on the waiting list beyond 3 months (mean: age 60, raw MELD 14). The AFP model performed the worst (AUC: 0.601). ALBI outperformed the Child-Pugh score (AUC: 0.710 vs. 0.699), while MELD scores demonstrated superior performance: MELD 3.0 achieved the highest AUC (0.752) compared to MELD (0.736) and MELD-Na (0.745). However, all traditional scores showed imbalance between sensitivity and specificity. ML models using 25 variables outperformed traditional scores, with RF achieving an AUC of 0.800. Incorporating 31 dynamic variables further improved performance, with RF achieving an AUC of 0.825 and balanced sensitivity (76.06%) and specificity (76.21%). Gini importance analysis identified new key predictors beyond MELD components, including largest tumor size, alpha-fetoprotein dynamics, BMI, and ascites severity. Conclusion: RF outperformed traditional scores in predicting 3-month waitlist mortality for LT candidates with HCC, by integrating both liver dysfunction and tumor progression risks. This approach, pending validation, could enhance organ allocation.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |