基于机器学习算法的胃肠癌患者术后短期并发症的预测研究

Prediction of Short-Term Postoperative Complications in Patients with Gastrointestinal Cancer Based on Machine Learning Algorithm

摘要: 目的：基于最优机器学习算法确定与胃肠癌术后短期并发症高权重危险因素，开发并验证预测并发症的模型。方法：本研究共纳入胃癌、结直肠癌手术的335例患者，按时间顺序将其分为训练集(268例)和验证集(67例)，所有患者均严格监测围术期临床数据和术后短期并发症的发生，利用三种机器学习算法比较确定最优算法，包括随机森林(RF)、支持向量机(SVM)、人工神经网络(ANN)，基于最优算法确定与术后7天、30天内发生的并发症相关的高权重危险因素。进而利用确定的危险因素建立基于人工智能的术后7天内并发症、30天内并发症的列线图预测模型。结果：基于最优算法——随机森林成功构建预测术后7天内并发症的模型Nomogram-A’和30天的模型Nomogram-B’。Nomogram-A’在训练集和验证集中的AUC值为0.888和0.805，Nomogram-B’在训练集和验证集中的AUC值为0.937和0.895。表明两模型的预测准确性良好。结论：本研究基于机器学习算法成功构建了预测术后7天内和30天并发症的列线图模型Nomogram-A’和Nomogram-B’，模型有良好的预测准确性。

Abstract: Objective: To identify high-weighted risk factors for short-term postoperative complications of gas-trointestinal cancer based on optimal machine learning algorithm, and to develop and validate models for predicting complications. Methods: A total of 335 patients undergoing surgery for gastric cancer and colorectal cancer were included in this study. They were divided into a training set (268 cases) and a verification set (67 cases) according to the time sequence. The perioperative clinical data and the occurrence of short-term postoperative complications of all patients were strictly monitored, and the optimal algorithm was compared and determined using three machine learning algorithms, including random forest (RF), support vector machine (SVM), and artificial neural net-work (ANN). High-weight risk factors related to complications occurring within 7 and 30 days after surgery were determined based on the optimal algorithm. Then the established risk factors were used to establish an artificial intelligence-based nomogram prediction model for complications within 7 days and 30 days after surgery. Results: Models Nomogram-A’ and Nomogram-B’ for pre-dicting postoperative complications within 7 days and 30 days, respectively, were successfully con-structed based on the optimal algorithm, random forest. The AUC values for Nomogram-A’ in the training and validation sets were 0.888 and 0.805, and the AUC values for Nomogram-B’ in the training and validation sets were 0.937 and 0.895. The results show that the two models have good prediction accuracy. Conclusions: In this study, we successfully constructed nomogram models Nomogram-A’ and Nomogram-B’ based on machine learning algorithm to predict complications within 7 and 30 days after surgery, with good prediction accuracy.