Open session Open full presentation page
Oral Presentation

Body-composition abnormalities and post-transplant prognosis in hepatocellular carcinoma liver transplantation: a multi-task clinical surrogate model

Presenter
Jinyan Chen (CN)

zhejiang university ยท Hangzhou

Profile

Authors

Jinyan Chen1, Xiao Xu2, Di Lu3

Affiliations

1 Hangzhou Medical College
2 The First Affiliated Hospital of Nanjing Medical University
3 Zhejiang university

Abstract

Background: Sarcopenia, myosteatosis, visceral obesity and high subcutaneous adipose tissue density (SATD) are common before liver transplantation (LT) for hepatocellular carcinoma (HCC), but their combined prognostic impact and simple clinical surrogates are unclear.
Methods: We analysed 409 adults who underwent LT for HCC (2015–2020) at two centres with pre-transplant L3 CT within 3 months. Skeletal muscle index, muscle radiodensity, visceral adipose tissue (VAT) area and SAT attenuation were measured at L3. Cohort-derived cut-offs (maximally selected rank statistics) defined four body-composition abnormalities and a cumulative burden (≤1, 2, ≥3 abnormalities). Recurrence-free survival (RFS) and overall survival (OS) were evaluated by Kaplan–Meier and multivariable Cox models. Using 42 routine clinical variables, we trained four multi-task models (70:30 train–test split): shared-feature random forests, a classifier chain, a multi-task neural network and a stacking multi-task ensemble.
Results: Each abnormality was associated with shorter RFS and OS. In multivariable models adjusted for beyond-Milan status, AFP >400 ng/mL and microvascular invasion, all four remained independent predictors of RFS (HR 1.69 for sarcopenia, 1.58 for myosteatosis, 2.82 for visceral obesity and 2.31 for high SATD; all P ≤0.021). Five-year RFS declined from 68.2% (≤1 abnormality) to 41.6% (2) and 17.9% (≥3); 5-year OS from 77.3% to 68.1% and 19.6% (all log-rank P ≤0.0001). On the test set, the classifier chain performed best, with AUCs of 0.789 (high SATD), 0.878 (myosteatosis), 0.882 (sarcopenia) and 0.865 (visceral obesity), and provided the greatest net clinical benefit.

Figure 1
Performance evaluation and clinical utility of multi-task learning models for body composition abnormality prediction.

Figure 2
Feature importance and chain dependency analysis of the Classifier Chain model.

Conclusions: Four CT-defined body-composition abnormalities independently and additively worsen post-transplant outcomes in HCC–LT. A multi-task classifier-chain model can approximate these abnormalities from routine clinical data with good accuracy, offering a practical route to embed body-composition risk into pre-transplant assessment when full CT analysis is not available.