基于机器学习算法的自体外周血造血干细胞采集预测模型构建与应用

Prediction model construction and application of machine learning algorithms for outcome prediction in autologous peripheral blood hematopoietic stem cell collection

目的筛选自体外周血造血干细胞(peripheral blood stem cell,PBSC)采集的危险因素并建立个体风险预测模型,以提高临床中自体PBSC采集的成功率。方法通过大数据平台收集2013年2月至2021年5月在海军军医大学第一附属医院血液科行自体PBSC采集术的恶性血液病患者757例,对患者进行单因素显著性统计学分析和多因素Logistic回归分析对PBSC采集危险因素进行筛选。采用Python 3.8.8版本、PyCharm 2021.1.3集成开发环境构建Logistic回归模型和前馈神经网络、最小二乘支持向量机、自动机器学习3种机器学习模型,并采用多种模型评价指标对其评价。结果共收集患者PBSC采集前的指标24项,单因素和Logistic回归分析筛选出11项PBSC采集危险因素。所构建的Logistic回归模型、前馈神经网络、最小二乘支持向量机和自动机器学习模型对自体PBSC采集风险预测的准确度分别为0.822、0.873、0.875和0.973。结论本研究所建自动机器学习模型能够准确预测自体PBSC采集结果,对提高临床自体PBSC采集成功率具有重要参考价值。

Objective To screen risk factors for autologous peripheral blood hematopoietic stem cell(PBSC)collection and develop an individual risk prediction model to improve the success rate of autologous PBSC collection in the clinic.Methods A total of 757 patients with hematologic malignancies who underwent PBSC collection in the Department of Hematology,First Affiliated Hospital of Naval University from Feb 2013 to May 2021 were collected through the big data platform,and the patients were screened for risk factors of PBSC collection by univariate statistical analysis and multivariate Logistic regression.Logistic regression models and three machine learning models,BP neural network(BPNN),least squares support vector machine(LSSVM),and automated machine learning(Auto-ML),were constructed using Python version 3.8.8,Pycharm 2021.1.3 integrated development environment,and the models were evaluated using several model evaluation metrics.Results A total of 24 items of the index before PBSC collection from patients were collected,and 11 items with risk factors for PBSC collection were screened by univariate and logistic regression analysis.The accuracies of the constructed logistic,BPNN,LSSVM and Auto-ML models for risk prediction of autologous PBSC collection were 0.822,0.873,0.875 and 0.973,respectively.Conclusion The established Auto-ML model can accurately predict the outcome of autologous PBSC collection and will be valuable for improving the success rate of autologous PBSC collection in the clinic.