基于深度学习算法开发和验证的肝细胞癌预后预测模型：一项大样本队列和外部验证研究

目的 基于深度学习算法开发和验证可评估肝细胞癌（hepatocellular carcinoma，HCC）患者预后的预测模型，并评估其价值。方法 选择2011年1月—2015年12月美国国立癌症研究所的监测、流行病学和最终结果（Surveillance，Epidemiology and Results，SEER）数据库中经病理确诊的HCC患者为训练队列用于构建模型，选择同期在本院经病理确诊的HCC患者为外部验证队列用于评估模型。主要预测结局为1、3、5年总生存率。分别使用深度学习算法DeepSurv、随机生存森林（RFS）、Cox比例风险回归开发模型，使用C-index评价模型的区分度，一致性曲线评估模型的校准度，log-rank检验评估危险分层能力。结果 训练队列9 730例患者和外部验证队列405例患者被纳入分析。在训练队列中，DeepSurv算法1、3、5年的C-index分别为0.85 （95%CI：0.80~0.90）、0.82 （95%CI：0.77~0.89）、0.80 （95%CI：0.73~0.87），在外部验证队列中分别为0.83 （95%CI：0.78~0.87）、0.79 （95%CI：0.74~0.83）、0.72 （95%CI：0.67~0.77）。无论在训练队列还是外部验证队列中，DeepSurv算法的C-index和校准度均优于RFS、Cox回归和TNM分期（均P<0.05）；log-rank检验显示，DeepSurv算法具有良好的危险分层能力（P<0.001）。结论 基于深度学习算法DeepSurv开发的预测模型可有效预测HCC患者死亡率，且优于常规的算法和预后评价指标。 

Deep learning- based development and validation of a predictive model for evaluating the prognosis of hepatocellular carcinoma: A study of a large-sample cohort and external validation

Objective To develop and validate a predictive model for evaluating the prognosis of patients with hepatocellular carcinoma (HCC) based on deep-learning algorithms, and to evaluate its value. Methods The SEER (Surveillance, Epidemiology and Results) database of the pathologically diagnosed HCC patients in the National Cancer Institute (USA) from January 2011 to December 2015 were selected as the training cohort to construct predictive models, and the HCC patients who were also pathologically diagnosed in the People's Hospital of Guigang City during the same period were selected as the external verification cohort to evaluate the model. The main predictions were 1-, 3-, and 5-year overall survival rates. The deep-learning algorithm DeepSurv, random survival forest (RFS), Cox proportional hazard regression were used to develop the models. C-index was used to evaluate the discrimination, the calibration curve was used to evaluate the calibration, and the log-rank test was used to evaluate the ability of risk stratification. Results A total of 9, 730 patients in the training cohort and 405 patients in the external verification cohort were finally included in the study. In the training cohort, the C-index of DeepSurv algorithm in 1-, 3-, and 5-year were 0.85 (95%CI: 0.80-0.90), 0.82 (95%CI: 0.77-0.89) and 0.80 (95%CI: 0.73-0.87), respectively; in the external validation cohort, they were 0.83 (95%CI: 0.78-0.87), 0.79 (95%CI: 0.74-0.83), 0.72 (95%CI: 0.67-0.77), respectively. The C-index and calibration of DeepSurv algorithm were better than those of RFS, Cox regression and TNM staging in both the training and external validation cohort (all P<0.05). The log-rank test showed that the DeepSurv algorithm had a good capability of risk stratification (P<0.001). Conclusions The predictive model developed based on the deep learning algorithms DeepSurv can effectively predict the mortality of HCC patients, and is superior to conventional algorithms and prognostic evaluation indicators.