Page 84 - Read Online

Ma et al. Hepatoma Res. 2026;12:43 Page 15 of 22

objective response to the first TACE. Limitations include survival calculated from response assessment date
rather than treatment date, introducing timing variability; accuracy dependent on mRECIST with inter-
observer variation; and no incorporation of subsequent TACE sessions.

Frontier and future models: towards comprehensive and individualized prediction
Currently, to achieve accurate and individualized prognosis prediction, the model will be developed towards
higher-dimensional data integration, smarter time series analysis, and more forward-looking decision
simulation, aiming to build a smarter and more forward-looking prognosis prediction system.

HCC death risk prediction model and HCC progression risk prediction model
The HCC death risk (HDR) prediction model and HCC progression risk (HPR) prediction model were
constructed by Dang et al., who collected plasma samples from 30 HCC patients before and 4-6 weeks after
TACE and performed whole-genome sequencing and capture-based mitochondrial DNA (mtDNA)
sequencing . The core input variables of the two models were the same: the three characteristics of cell-free
[55]
mitochondrial DNA (cf-mtDNA): mtDNA copy number, number of mtDNA mutations, and HCC
Prognosis Prediction (HPP) score (a prognostic prediction score that integrates mtDNA fragment omics
characteristics). Patients were divided into two groups according to the median risk score of the patient
cohort: high-risk group (risk score > median) and low-risk group (risk score ≤ median). The results showed
that the progression free survival (PFS) and OS of patients in the high-risk group before and after TACE
treatment were significantly shorter (mPFS before treatment: HR = 0.37, P <​ 0.01; mOS before treatment: HR
= 0.30, P <​ 0.01; mPFS after treatment: HR = 0.28, P <​ 0.01; mOS after treatment: HR = 0.28, P <​ 0.01), that is,
the risk of disease progression and death in the high-risk group increased significantly. The study also
monitored the change trend of the three characteristics before and after treatment to make a dynamic
prediction. If any two of the three characteristics were elevated after treatment, the patient was classified as a
high-risk group; otherwise, it was a low-risk group. The OS and PFS of the high-risk group were significantly
worse than those of the low-risk group. This study proved that cf-mtDNA multi-feature analysis is a
powerful tool with excellent performance, consistent with the gold standard, and superior to existing liquid
biopsy methods (CNV load) in predicting the efficacy and prognosis of TACE in HCC patients. However,
this study also has limitations: (1) small sample size; (2) lack of external validation and multicenter data; (3)
unclear mechanism of cf-mtDNA release, requiring further investigation into its biological basis.

CatBoost model
The CatBoost model is a prediction model based on machine learning developed and validated by Wei et al.
based on preoperative clinical and CT image characteristics . It was used to predict the risk of advanced-
[56]
stage progression (ASP) in patients with intermediate HCC after TACE and to evaluate whether it could
guide the choice of postoperative systemic treatment. A total of 34 preoperative clinical and CT imaging
variables were included in this model, and 11 key variables were finally selected for modeling. This study
compared six machine learning algorithms (CatBoost, XGBoost, GBDT, LGBM, RF, and LR). CatBoost
performed best on the three datasets, and the C-index and time-dependent area under the curve (AUC) of
the CatBoost model were significantly better than those of all existing staging systems (all P <​ 0.01). Using
the predicted probability of the model output, an optimal risk threshold (66.27 points) was determined by X-
tile software, and the patients were divided into the “high-risk group” and “low-risk group”. The results
showed that systemic therapy after TACE in the high-risk group significantly improved PFS and OS (P <​
0.01), whereas the low-risk group showed no significant benefit, indicating that the model can effectively
identify patients who may benefit from postoperative adjuvant therapy. However, this model has limitations:
(1) The multicenter retrospective design introduces heterogeneity and a large time span, while the low ASP
incidence (6.7%-8.4%) leads to class imbalance; (2) The study population was highly specific (mostly Chinese
patients with HBV-related HCC), and pathological or genetic factors were not included; (3) Exclusion of
patients who received systemic therapy before or after TACE limits model generalizability.