Abstract: Application value of gradient boosting machine model for predicting prognosis after resection of intrahepatic cholangiocarcinoma        Fan Ye, JI Gu-wei, XU Zheng-gang, et al. Hepatobiliary Center, the First Affiliated Hospital of Nanjing Medical University, Nanjing210029, China
Corresponding author: WANG Ke, E-mail: lancetwk@163.com
Abstract    Objective    To explore the efficacy of gradient boosting machine (GBM) model in predicting the prognosis after resection of intrahepatic cholangiocarcinoma (ICC). Methods    A total of 649 ICC patients treated with surgical resection between 2004 and 2015 from the Surveillance, Epidemiology, and End Results (SEER) database were retrospectively analyzed. All the patients were randomly divided into 433 case in the training dataset and 216 cases in the test dataset in a 2∶1 ratio. The prediction model was built by GBM algorithm. The prediction accuracy of the model was measured by the concordance index (C-index) and the model fit was assessed by the calibration curve. The prediction performance of GBM model and ICC staging system was compared. All models were validated in the independent test dataset. Results    Median cancer-specific survival was 39.0 months in the training dataset while that in the validation dataset was 40.0 months, there was no significant difference between the two groups (χ2=0.131, P=0.718). The GBM model consisted of 6 commonly used clinicopathological characteristics: age, tumor size, number of regional lymph node metastasis, tumor number, histological grade and vascular invasion. In both training and test datasets, the GBM model predicted cancer-specific survival with respective C-index of 0.741 (95%CI 0.709-0.773) and 0.719 (95%CI 0.671-0.766), respectively. The prediction accuracy of GBM model was superior to that of TNM staging system with statistically significant difference (P<0.05). The calibration curves demonstrated good agreement between GBM model-predicted probabilities and observed outcomes. All patients could be stratified into low-risk, intermediate-risk and high-risk group based on GBM model, statistically significant differences among three risk groups were observed in both training and test datasets (P<0.001). GBM model could better identify ICC patients with favorable prognosis than TNM staging system. Conclusion    The proposed GBM model assembled with 6 commonly used clinicopathological characteristics in this study, which can significantly improve the prognosis prediction of ICC. By exploring the interactions between clinicopathological characteristics, machine learning makes accurate prediction via ensemble learning that may have important clinical value in the era of big data.

Key words: intrahepatic cholangiocarcinoma, machine learning, prediction model, surgical resection, prognosis

摘要： 目的    探讨梯度提升机（GBM）模型在肝内胆管癌（ICC）手术预后预测中的效能。方法    回顾性分析美国国立癌症研究所监测、流行病学和最终结果（SEER）数据库2004—2015年间接受手术治疗的649例ICC病人数据，按2∶1随机分为训练集433例和测试集216例。采用GBM算法构建预后预测模型。通过一致性指数（C-index）衡量模型预测的准确性，校准曲线反映模型的拟合情况。比较GBM模型和ICC分期系统的预测效能。所有模型均在独立的测试集内进行验证。结果    训练集内病人中位肿瘤特异性生存时间为39.0个月，测试集内病人中位肿瘤特异性生存时间为40.0个月，两组病人生存情况比较，差异无统计学意义（χ2=0.131，P=0.718）。GBM模型由6个常用临床病理特征构成：年龄、肿瘤大小、区域淋巴结转移数目、肿瘤数目、肿瘤分化程度和血管侵犯。在训练集和测试集中，GBM模型对术后肿瘤特异性生存预测的C-index值分别为0.741（95%CI 0.709-0.773）和0.719（95%CI 0.671-0.766），GBM模型的预测效能优于TNM分期系统，差异均具有统计学意义（P<0.05）。校准曲线示GBM模型预测概率与实际观察值具有较好的一致性。GBM模型可将ICC病人分为低危、中危和高危组，3组病人手术预后情况比较，在训练集和测试集内的差异均有统计学意义（P<0.001）。GBM模型比TNM分期能够更好地识别出具有良好手术预后的ICC病人。结论    构建的GBM模型集合了6个常用临床病理特征，可显著改善对ICC手术预后的评估。机器学习通过深入挖掘临床病理特征间的相互作用并以集成学习的方式实现精准预测，在大数据时代具有重要的临床价值。

关键词: 肝内胆管癌, 机器学习, 预测模型, 手术切除, 预后