# A machine learning approach to identify optimal candidates for transarterial chemoembolization in unresectable HBV-related hepatocellular carcinoma complicated by first-branch portal vein tumor thrombus: a multicenter study

## TL;DR

This study uses machine learning to identify patients with a specific type of liver cancer who may benefit from a treatment called TACE, improving clinical decision-making.

## Contribution

A machine learning model is developed to select optimal TACE candidates for HCC with PVTT1, showing comparable survival outcomes to patients without PVTT.

## Key findings

- The RSF model identified tumor size, tumor number, AST, INR, and age as top predictors for TACE-treated PVTT1 patients.
- Low-risk PVTT1 patients had survival outcomes similar to those without PVTT, even after PSM.
- TACE provided a survival advantage over sorafenib for PVTT1 patients.

## Abstract

Portal vein tumor thrombus (PVTT) is a critical factor influencing prognosis and treatment allocation for Hepatitis B virus (HBV)-related hepatocellular carcinoma (HCC). However, accurately selecting patients with unresectable HCC and first-order branch PVTT (PVTT1) who would benefit from transarterial chemoembolization (TACE) remains a significant clinical challenge. This study aimed to leverage machine learning to address this issue.

We conducted a large-scale, retrospective multicenter study utilizing data from 15 tertiary hospitals in China (2012-2021). A Random Survival Forest (RSF) model was constructed to identify key prognostic variables and stratify risk among TACE-treated PVTT1 patients. Model performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC). To mitigate selection bias and confounding factors for survival comparisons, Propensity Score Matching (PSM) was performed.

Of the 3,948 patients enrolled, 763 constituted the TACE-PVTT1 group. The RSF model exhibited robust predictive accuracy for this group, identifying tumor size, tumor number, AST, INR, and age as the top five clinical predictors. Patients in the bottom risk-score tertile were classified as low-risk. Notably, the overall survival (OS) of this low-risk TACE-PVTT1 group was not significantly different from that of the 3,073 TACE-PVTT0 patients (P = 0.19), a finding that was maintained after PSM (P = 0.54). A multivariate Cox analysis confirmed that in this low-risk context, PVTT1 status was not a significant prognostic factor (P = 0.08575). Additionally, TACE conferred a significant survival advantage over sorafenib in patients with PVTT1.

The integrated application of RSF and PSM can effectively identify low-risk candidates for TACE among patients with unresectable HCC and PVTT1. Our findings provide strong evidence that for this carefully selected patient subgroup, TACE offers survival outcomes comparable to those for patients without PVTT, highlighting the clinical utility of machine learning in guiding treatment decisions for this challenging disease.