基于Stacking集成学习的医用直线加速器主要故障联锁预测模型研究  

Research on prediction model for major fault interlock of medical linear accelerators based on stacking ensemble learning

作　　者：李亮 何威震 沙冠辰 解昕 陈勇 章龙珍 LI Liang;HE Weizhen;SHA Guanchen;XIE Xin;CHEN Yong;ZHANG Longzhen(Department of Radiotherapy,the Affiliated Hospital of Xuzhou Medical University,Xuzhou,Jiangsu Province 221006;School of Biomedical Engineering,Department of medicine,Shenzhen University,Shendong,Guangzhou Province 518055;School of Precision Instrument and Opto-Electronics Engineering,Tianjin University,Tianjin 300110)

机构地区：[1]徐州医科大学附属医院肿瘤放射治疗科,江苏徐州221006 [2]深圳大学医学部生物医学工程学院,广东深圳518055 [3]天津大学精密仪器与光电子工程学院,天津300110

出　　处：《北京生物医学工程》2025年第1期68-73,共6页Beijing Biomedical Engineering

基　　金：国家自然科学基金面上项目(81972845);徐州市引进临床医学专家团队项目(2019TD003)资助。

摘　　要：目的研究基于Stacking集成学习模型应用于医用直线加速器主要故障联锁预测的可行性。方法回顾性收集119个月瓦里安23EX直线加速器频次最多的4种故障联锁(代码:MLC、HWFA、GFIL与UDRS),并将加速器使用时间(月)、月治疗人数、月射野数与月MU数考虑为故障联锁的影响因素。利用Stacking集成学习方法构造医用直线加速器主要故障联锁预测模型,通过比较故障联锁频次曲线与真实故障联锁频次曲线的相似程度、均方根误差、平均绝对值误差和决定系数,对各基模型和集成学习模型进行预测精度和预测性能评估。结果相较于各基模型,集成学习模型的各故障联锁频次曲线与真实故障联锁频次曲线更为相似,集成学习模型的均方根误差、平均绝对值误差和决定系数在MLC联锁故障预测中分别为0.41、0.33和83.2%;在HWFA联锁故障预测中分别为0.19、0.17和74.2%;在GFIL联锁故障预测中分别为0.19、0.16和67.9%;在UDRS联锁故障预测中分别为0.20、0.17和71.5%。各指标结果均优于单一的基模型。结论基于Stacking集成学习模型能够较为准确地对直线加速器主要故障联锁趋势进行预测,对于加速器的预防性维护和故障维修管理具有一定的应用价值。Objective To study the feasibility of applying the stacking ensemble learning model to the prediction of major fault interlocking of medical linear accelerators.Methods The four fault interlocks(codes:MLC,HWFA,GWIL and UDRS)with the highest frequency of Varian 23EX linear accelerator at 119 months were retrospectively collected,and the accelerator use time(months),monthly number of treatment,monthly number of shooting fields and monthly MU were considered as the influencing factors of fault interlocking.The Stacking ensemble learning method is used to construct the prediction model of the main fault interlocking of medical linear accelerators,and the prediction accuracy and prediction performance of each base model and the ensemble learning model are evaluated by comparing the similarity,root mean square error,mean absolute value error and coefficient of determination between the fault interlocking frequency curve and the real fault interlocking frequency curve.Results Compared with the base models,the fault interlocking frequency curves of the ensemble learning model are more similar to the real fault interlocking frequency curves,and the root mean square error,mean absolute value error and coefficient of determination of the ensemble learning model are 0.41,0.33 and 83.2%in MLC interlock fault prediction,respectively.In the prediction of HWFA interlock faults,they were 0.19,0.17 and 74.2%,respectively.In the GFIL interlock fault prediction,they were 0.19,0.16 and 67.9%,respectively.In the UDRS interlock fault prediction,they are 0.20,0.17 and 71.5%,respectively.The results of each indicator were better than the single base model.Conclusions Based on the Stacking ensemble learning model,the main fault interlocking trend of linear accelerator can be predicted more accurately,which has certain application value for preventive maintenance and fault repair management of accelerator.

关 键 词：直线加速器 故障联锁预测 集成学习 长短期记忆网络 

分 类 号：R318.04[医药卫生—生物医学工程]