Nam Joon Yeul, Lee Jeong-Hoon, Bae Junho, Chang Young, Cho Yuri, Sinn Dong Hyun, Kim Bo Hyun, Kim Seoung Hoon, Yi Nam-Joon, Lee Kwang-Woong, Kim Jong Man, Park Joong-Won, Kim Yoon Jun, Yoon Jung-Hwan, Joh Jae-Won, Suh Kyung-Suk
Department of Internal Medicine and Liver Research Institute, Seoul National University College of Medicine, Seoul 03080, Korea.
DEEPNOID Inc., Seoul 08376, Korea.
Cancers (Basel). 2020 Sep 29;12(10):2791. doi: 10.3390/cancers12102791.
Several models have been developed using conventional regression approaches to extend the criteria for liver transplantation (LT) in hepatocellular carcinoma (HCC) beyond the Milan criteria. We aimed to develop a novel model to predict tumor recurrence after LT by adopting artificial intelligence (MoRAL-AI). This study included 563 patients who underwent LT for HCC at three large LT centers in Korea. Derivation ( = 349) and validation ( = 214) cohorts were independently established. The primary outcome was time-to-recurrence after LT. A MoRAL-AI was derived from the derivation cohort with a residual block-based deep neural network. The median follow-up duration was 74.7 months (interquartile-range, 18.5-107.4); 204 patients (36.2%) had HCC beyond the Milan criteria. The optimal model consisted of seven layers including two residual blocks. In the validation cohort, the MoRAL-AI showed significantly better discrimination function (c-index = 0.75) than the Milan (c-index = 0.64), MoRAL (c-index = 0.69), University of California San Francisco (c-index = 0.62), up-to-seven (c-index = 0.50), and Kyoto (c-index = 0.50) criteria (all 0.001). The largest weighted parameter in the MoRAL-AI was tumor diameter, followed by alpha-fetoprotein, age, and protein induced by vitamin K absence-II. The MoRAL-AI had better predictability of tumor recurrence after LT than conventional models. The MoRAL-AI can also evolve with further data.
已经开发了几种使用传统回归方法的模型，以将肝细胞癌（HCC）肝移植（LT）的标准扩展到米兰标准之外。我们旨在通过采用人工智能（MoRAL-AI）开发一种新型模型来预测LT后的肿瘤复发。本研究纳入了韩国三个大型LT中心接受HCC-LT的563例患者。独立建立了推导队列（n = 349）和验证队列（n = 214）。主要结局是LT后的复发时间。MoRAL-AI由基于残差块的深度神经网络从推导队列中得出。中位随访时间为74.7个月（四分位间距，18.5-107.4）；204例患者（36.2%）的HCC超出米兰标准。最佳模型由包括两个残差块的七层组成。在验证队列中，MoRAL-AI显示出比米兰标准（c指数=0.64）、MoRAL（c指数=0.69）、加利福尼亚大学旧金山分校标准（c指数=0.62）、七项扩展标准（c指数=0.50）和京都标准（c指数=0.50）显著更好的区分功能（c指数=0.75）（均P<0.001）。MoRAL-AI中权重最大的参数是肿瘤直径，其次是甲胎蛋白、年龄和维生素K缺乏诱导蛋白-II。与传统模型相比，MoRAL-AI对LT后肿瘤复发具有更好的预测性。MoRAL-AI也可以随着更多数据而不断发展。
