35: Q&A

Hepatocellular carcinoma (HCC) is the most rapidly growing cause of cancer death in the US. Despite strict criteria for transplantation, some will develop liver (LR) and extrahepatic recurrences (ER). The goal of this study is to use machine learning to predict recurrence patterns after liver transplantation for HCC.


Methods: Consecutive HCC patients transplanted at 20 US centers from 2002-2013 were studied for clinicopathologic features associated with recurrence. A classification model based on the Random Forest algorithm (an ensemble machine learning method using decision trees) was constructed. This model uses a 10-fold cross-validation strategy trained on 10% of the cohort and validated on the remaining 90%. To predict recurrence patterns, the Label Powerset method was used to predict recurrence site (only LR vs only ER vs ER + LR vs none) using multilabel classification.


Results: Of 4980 patients, 594 (11.9%) recurred with a median follow up of 47 months. Only LR occurred in 137 (23.1%), only ER in 372 (62.6%), and 63 (10.6%) had ER + LR. Those that recurred more commonly had higher peak and pre-transplant alpha fetoprotein (AFP) levels (58 and 24 ng/mL vs 16 and 8 ng/mL, p< 0.001), more tumors ≥ 5 cm (24.7% vs 8.1%, p< 0.001), micro and macrovascular invasion (37.9% and 16.2% vs 16.4% and 2.8%, p< 0.001), and higher neutrophil-to-lymphocyte ratio (NLR 3.3 vs 2.9, p=0.019). Based on feature importance analysis performed at each cross-validation iteration, these five features were the most important recurrence predictors. Notably, tumor size and peak AFP were always the most important features. On averaging the area under the curve (AUC) for the ten strategies, a mean AUC=0.78 was obtained (Figure 1). Using these features, multilabel classification predicted the recurrence site with mean AUC=0.74.


Conclusions: Recurrence after liver transplantation for HCC was uncommon but usually occurred as an ER. Maximum AFP and largest tumor size were the most important predictors and may help guide surveillance for patients at risk for recurrence by location.