Open AccessA Hybrid Machine Learning Model Based on Semantic Information Can Optimize Treatment Decision for Naïve Single 3–5-cm HCC Patients
Author(s) -
Ding Wenzhen,
Wang Zhen,
Liu Fang-Yi,
Cheng Zhi-Gang,
Yu Xiaoling,
Han Zhiyu,
Zhong Hui,
Yu Jie,
Liang Ping
Publication year - 2022
Publication title -
liver cancer
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.916
H-Index - 34
eISSN - 1664-5553
pISSN - 2235-1795
DOI - 10.1159/000522123
Subject(s) - research article
Background: Tumor recurrence is an abomination for hepatocellular carcinoma (HCC) patients receiving local treatment. Purpose: The aim of the study was to build a hybrid machine learning model to recommend optimized first treatment (laparoscopic hepatectomy [LH] or microwave ablation [MWA]) for naïve single 3–5-cm HCC patients based on early recurrence (ER, ≤2 years) probability. Methods: This retrospective study collected 20 semantic variables of 582 patients (LH: 300, MWA: 282) from 13 hospitals with at least 24 months follow-up. Both groups were divided into training, validation, and test set, respectively. Five algorithms (logistics regression, random forest, neural network, stochastic gradient boosting, and eXtreme Gradient Boosting [XGB]) were used for model building. A model with highest area under the receiver operating characteristic curve (AUC) in a validation set of LH and MWA was selected to connect as a hybrid model which made decision based on ER probability. Model testing was performed in a comprehensive set comprising LH and MWA test sets. Results: Four variables in each group were selected to build LH and MWA models, respectively. LH-XGB model (AUC = 0.744) and MWA-stochastic gradient method (AUC = 0.750) model were selected for model building. In the comprehensive set, a treatment confusion matrix was established based on recommended and actual treatment. The predicted ER probabilities were comparable with the actual ER rates for various types of patients in matrix ( p > 0.05). ER rate of patients whose actual treatment consistent with recommendation was lower than that of inconsistent patients (LH: 21.2% vs. 46.2%, p = 0.042; MWA: 26.3% vs. 54.1%, p = 0.048). By recommending optimal treatment, the hybrid model can significantly reduce ER probability from 38.2% to 25.6% for overall patients ( p < 0.001). Conclusions: The hybrid model can accurately predict ER probability of different treatments and thereby provide reliable evidence to make optimal treatment decision for patients with single 3–5-cm HCC.