基于机器学习算法预测腹膜后脂肪肉瘤患者总体生存期的临床研究

Clinical study of predicting overall survival of patients with retroperitoneal liposarcoma based on machine learning algorithms

目的:比较不同类型的机器学习算法对腹膜后脂肪肉瘤(RP-LPS)患者总体生存期的预测效果,选择最佳的算法对临床诊疗进行指导。方法:通过美国SEER*Stat软件搜集2000至2019年确诊为腹膜后脂肪肉瘤的病例(2 147例)作为训练集进行研究,我们将南通大学附属医院2014年至2019年确诊的腹膜后脂肪肉瘤的病例(55例)作为外部验证集,使用的机器学习算法包括支持向量机、自适应提升法、决策树、随机森林和神经网络。分别比较这几种算法在训练集和外部验证集中的预测效果。结果:通过比较各个机器学习算法的预测性能指标,包括准确率、敏感度、AUC、F1分数等,得出自适应提升法的预测效果最佳,在训练集中的准确率为69.1%,敏感度为76.5%,AUC为0.70,在外部验证集的准确率为74.5%,敏感度为72.0%,AUC为0.74。再与传统的TNM模型进行比较,也表现出更佳的预测性能。结论:机器学习算法提供了比传统预测模型更加准确和个性化的腹膜后脂肪肉瘤患者的预后信息,可用于辅助医生判断患者预后情况及治疗效果,制定个性化的诊疗方案。

Objective:To compare the prediction effects of different types of machine learning algorithms on the overall survival of retroperitoneal liposarcoma(RP-LPS)patients,and choose the best algorithm to guide clinical diagnosis and treatment.Methods:The cases(2147 cases)diagnosed with retroperitoneal liposarcoma from 2000 to 2019 were collected by SEER*Stat software in the United States as a training set for research.We selected patients(55 cases)with retroperitoneal liposarcoma diagnosed in Nantong University Affiliated Hospital from 2014 to 2019 as an external validation set,using machine learning algorithms including support vector machine,adaptive boosting,decision tree,random forest,and neural network.The prediction performance of these algorithms in the training set and the validation set was compared respectively.Results:By comparing the prediction performance indicators of each machine learning algorithm,including accuracy,sensitivity,AUC,F1 score,etc.,it was concluded that the prediction effect of the adaptive improvement algorithm was the best,in the training set,the accuracy rate was 69.1%,the sensitivity was 76.5%and AUC was 0.70.In the external validation set,the accuracy was 74.5%,the sensitivity was 72.0%and AUC was 0.74.Compared with the traditional TNM model,it also shows better prediction performance.Conclusion:The machine learning algorithm provides more accurate and personalized prognostic information about retroperitoneal liposarcoma than traditional prediction models,which can be used to assist doctors in judging the prognosis and treatment effect of patients,and formulate a personalized diagnosis and treatment plans.