Page 169 - Read Online

Page 62 Endo et al. Art Int Surg 2024;4:59-67 https://dx.doi.org/10.20517/ais.2024.09

Table 1. Summary of representative studies using machine learning related to hepatobiliary cancers
Ref. Year Patient Clinical application Machine learning approach
Famularo et al. [37] 2022 Hepatocellular carcinoma Prediction of survival after recurrence Least Absolute Shrinkage and
Selection Operator
[38]
Moazzam et al. 2024 Hepatocellular carcinoma Prediction of survival after recurrence Optimal Survival Tree
[42]
Iseke et al. 2023 Hepatocellular carcinoma Prediction of recurrence incorporating MRI data Conventional neural network
[43]
Saillard et al. 2020 Hepatocellular carcinoma Prediction of survival incorporating digitazed Conventional neural network
histological whole-slide data
[49]
Jiang et al. 2021 Hepatocellular carcinoma Prediction of microvascular invasion Radiomics and conventional neural
network
[51]
Alaimo et al. 2023 Intrahepatic Prediction of early recurrence Random forest
cholangiocarcinoma
[53]
Cotter et al. 2022 Gallbladder cancer Prediction of overall survival Classification and Regression Tree
Tsilimigras et al. [54] 2020 Intrahepatic Classification based on machine learning Hierarchical machine-learning
cholangiocarcinoma
[55]
Chen et al. 2023 Intrahepatic Prediction of very early recurrence Radiomics and K-means clustering
cholangiocarcinoma
[60]
Alaimo et al. 2023 Intrahepatic Prediction of survival and recurrence relative to Optimal policy tree
cholangiocarcinoma margin width
institutional database to formulate a prediction model for SAR. In this study, the SARScore was proposed as
a prediction model based on clinicopathological determinants such as cirrhosis, number of primary tumors,
macrovascular invasion, R1 resection margin, alpha-fetoprotein (AFP) > 400 ng/mL on diagnosis of
recurrent disease, extrahepatic recurrence, radiologic size and number of recurrent lesions, radiologic
recurrent bilobar disease, and recurrence within 24 months after hepatectomy. The clinical applicability of
[12]
SARScore was assessed using Optimal Survival Tree (OST) analysis . This ML-based tool demonstrated
that patients with high SARScores experienced the worst survival outcomes (5-year AUC; training: 0.79 vs.
testing: 0.71). In turn, the combination of SARScore and OST analysis can provide risk stratification and
therapeutic guidance in the treatment of individuals with recurrent HCC.

AI has also been applied to develop prognostic risk scores following hepatectomy [40,41] . Traditionally,
prognostic risk stratification has relied on “structured” or pre-specified data including patient
characteristics, and tumor size/number. In recent years, AI has evolved to integrate structured information
with more detailed yet previously untapped data. For example, Iseke et al. employed ML to predict
recurrence using pretreatment laboratory, clinical, and magnetic resonance imaging (MRI) data among
patients with early-stage HCC initially eligible for liver transplantation . This study demonstrated that
[42]
combining MRI radiomics with clinical parameters yielded the most accurate prediction of post-treatment
recurrence. These data suggested that ML-based models can forecast recurrence before therapy allocation
among patients eligible for liver transplantation with early-stage HCC. The incorporation of MRI data into
the model significantly enhanced predictive performance compared with reliance on clinical parameters
alone. Saillard et al. adopted a deep learning technique, utilizing whole-slide digitized histological slides (i.e.,
whole-slide imaging; WSI) to construct models to predict the survival of HCC patients undergoing surgical
resection . Notably, the derived scores demonstrated high C-indexes of 0.78 and 0.75, respectively,
[43]
outperforming other models using structured clinicopathologic features. AI enabled the exploration of
previously untapped data and the identification of distinctive features that may have been overlooked. One
significant challenge, however, lies in the external application of imaging-based AI models to other cohorts,
as the lack of a user-friendly interface hinders the broad adoption of this technique . Addressing this
[44]
challenge with the creation of easy-to-use, online applications will be crucial in the future.