Deep learning for hepatocellular carcinoma recurrence before and after liver transplantation: a multicenter cohort study

Hepatocellular carcinoma (HCC) recurrence after liver transplantation (LT) is a major contributor to mortality. We developed a recurrence prediction system for HCC patients before and after LT. Data from patients with HCC who underwent LT were retrospectively collected from three specialist centres in China. Pre- and post-operative variables were selected using support vector machine, random forest, and logistic regression (LR). Then, pre- and post-operative models were developed using three machine learning methods: LR, stacking, and two survival-based approaches. Models were evaluated using seven assessment indices, and patients were classified as either high- or low-risk based on recurrence risk. 466 patients were included and followed for a median of 51.0 months (95% CI 47.8-54.2). The pre-DeepSurv model (pre-DSM) had a C-index of 0.790 ± 0.003 during training, 0.775 ± 0.037 during testing, and 0.765 ± 0.001 and 0.819 ± 0.002 during external validation. After incorporating clinicopathologic variables, the post-DeepSurv model (post-DSM) had a 0.835 ± 0.008 C-index during training, 0.812 ± 0.082 during testing, and 0.839 ± 0.001 and 0.831 ± 0.002 during external validation. The post-DSM outperformed the Milan criteria by more accurately identifying patients at high risk of recurrence. Tumour recurrence predictions also improved significantly with DeepSurv. Both pre- and post-DSMs have the potential to guide personalised surveillance strategies for LT patients with HCC.