Hayashi Koki, Ono Yoshihiro, Takamatsu Manabu, Oba Atsushi, Ito Hiromichi, Sato Takafumi, Inoue Yosuke, Saiura Akio, Takahashi Yu
Division of Hepatobiliary and Pancreatic Surgery, Cancer Institute Hospital of the Japanese Foundation for Cancer Research, Koto-ku, Tokyo, Japan.
Division of Pathology, Department of Pathology, Cancer Institute, Cancer Institute Hospital of the Japanese Foundation for Cancer Research, Koto-ku, Tokyo, Japan.
Ann Surg Oncol. 2022 Mar 1. doi: 10.1245/s10434-022-11471-x.
Patients with pancreatic cancer (PC) have poor prognosis and a high incidence of recurrence. Since further treatment is applicable for specific recurrent events, it is important to predict recurrence patterns after surgery. This study aimed to identify and predict early and late recurrence patterns of PC using a histology-based machine learning model.
Patients who underwent upfront curative surgery for PC between 2001 and 2014 were included. The timing of recurrence and prognosis of each first recurrence site were examined. A histology-based supervised machine learning method, which combined convolutional neural networks and random forest, was used to predict the recurrence and respective sites of metastasis. Accuracy was evaluated using area under the receiver operating characteristic curve (AUC).
In total, 524 patients were included. Recurrence in the liver accounted for 47.8% of all recurrence events in the first year after surgery. Meanwhile, recurrence in the lung occurred later and could become apparent more than 5 years post-surgery, with indications for further surgery. In terms of substantial distant organ metastases, liver and lung metastases were identified as representative early and late recurrence events. The predictive AUCs of the machine learning model for training and test data were 1.000 and 0.861, respectively, and for predicting nonrecurrence were 1.000 for both.
We identified the liver and lung as early and late recurrence sites, which could be distinguished with high probability using a machine learning model. Prediction of recurrence sites using this model may be useful for further treatment of patients with PC.
胰腺癌（PC）患者预后较差且复发率高。由于针对特定复发事件可进行进一步治疗，因此预测手术后的复发模式很重要。本研究旨在使用基于组织学的机器学习模型识别和预测PC的早期和晚期复发模式。
纳入2001年至2014年间接受PC初次根治性手术的患者。检查了每次首次复发部位的复发时间和预后情况。使用一种基于组织学的监督机器学习方法，该方法结合了卷积神经网络和随机森林，来预测复发及转移的各个部位。使用受试者工作特征曲线下面积（AUC）评估准确性。
总共纳入524例患者。术后第一年肝脏复发占所有复发事件的47.8%。同时，肺部复发出现较晚，可能在术后5年以上才明显，有进一步手术的指征。就实质性远处器官转移而言，肝转移和肺转移被确定为具有代表性的早期和晚期复发事件。机器学习模型对训练数据和测试数据的预测AUC分别为1.000和0.861，对预测无复发的AUC两者均为1.000。
我们确定肝脏和肺部为早期和晚期复发部位，使用机器学习模型可高概率区分这两者。使用该模型预测复发部位可能有助于PC患者的进一步治疗。
