5-year recurrence prediction after hepatocellular carcinoma resection: deep learning vs. Cox regression models

Hon-Yi Shi; King-The Lee; Chong-Chi Chiu; Jhi-Joung Wang; Ding-Ping Sun; Hao-Hsien Lee

5-year recurrence prediction after hepatocellular carcinoma resection: deep learning vs. Cox regression models

Am J Cancer Res. 2022 Jun 15;12(6):2876-2890. eCollection 2022.

Authors

Hon-Yi Shi^{1

2

3

4}, King-The Lee^{1

5}, Chong-Chi Chiu^{6

7}, Jhi-Joung Wang^{8

9}, Ding-Ping Sun¹⁰, Hao-Hsien Lee¹¹

Affiliations

¹ Department of Healthcare Administration and Medical Informatics, Kaohsiung Medical University Kaohsiung 80708, Taiwan.
² Department of Business Management, National Sun Yat-sen University Kaohsiung 80420, Taiwan.
³ Department of Medical Research, Kaohsiung Medical University Hospital Kaohsiung 80708, Taiwan.
⁴ Department of Medical Research, China Medical University Hospital, China Medical University Taichung 40402, Taiwan.
⁵ Hepatobiliary-Pancreatic Surgery, Park One International Hospital Kaohsiung 81357, Taiwan.
⁶ Department of General Surgery, E-Da Cancer Hospital Kaohsiung 82445, Taiwan.
⁷ School of Medicine, College of Medicine, I-Shou University Kaohsiung 82445, Taiwan.
⁸ Department of Medical Research, Chi Mei Medical Center Yongkang, Tainan 71004, Taiwan.
⁹ Allied AI Biomed Center, Southern Taiwan University of Science and Technology Tainan 71005, Taiwan.
¹⁰ Department of Surgery, Chi Mei Medical Center Yongkang, Tainan 71004, Taiwan.
¹¹ Department of Surgery, Chi Mei Medical Center Liouying, Tainan 73658, Taiwan.

PMID: 35812048
PMCID: PMC9251698

Abstract

Deep learning algorithms have yet to be used for predicting clinical prognosis after cancer surgery. Therefore, this study compared performance indices and permutation importance of potential confounders in three models for predicting 5-year recurrence after hepatocellular carcinoma (HCC) resection: a deep-learning deep neural network (DNN) model, a recurrent neural network (RNN) model, and a Cox proportional hazard (CPH) regression model. Data for 725 patients who had received HCC resection at three medical centers in southern Taiwan between April, 2011, and December, 2015, were randomly divided into three datasets: a training dataset containing data for 507 subjects was used for model development, a testing dataset containing data for 109 subjects was used for internal validation, and a validating dataset containing data for 109 subjects was used for external validation. Feature importance analysis was also performed to identify potential predictors of recurrence after HCC resection. Univariate Cox proportional hazards regression analyses were performed to identify potential significant predictors of 5-year recurrence after HCC resection, which were included in the forecasting models (P < 0.05). All performance indices for the DNN model were significantly higher than those for the RNN model and the conventional CPH model (P < 0.001). The most important potential predictor of 5-year recurrence after HCC resection was surgeon volume followed by, in order of importance, hospital volume, preoperative Beck Depression Scale score, preoperative Beck Anxiety Scale score, co-residence with family, tumor stage, and tumor size. The feature importance analysis performed to investigate interpretability in this study elucidated the potential use of deep learning models for predicting recurrence after HCC resection and for identifying predictors of recurrence. Further experiments using the proposed DNN model would clarify its potential uses for developing, promoting, and improving health policies for treating HCC patients after surgery.

Keywords: Hepatocellular carcinoma; deep learning; feature importance; recurrence; resection.