摘要
急诊患者到达预测是医生排班的基础,是解决急诊拥堵的关键。现有预测多为单一预测算法,针对每天、每月进行,缺乏更短时间预测。构建基于堆叠法(Stacking)集成学习模型的预测方法,分别以小时、天、周为时间单位,对患者到达进行预测,探究不同时间单位的预测效果。在随机森林(RF)、梯度提升决策树(GBDT)、极端梯度提升树(XGB)、Light GBM(LGB)、支持向量回归(SVR)、K近邻学习(KNN)中选择预测效果较佳的方法作为Stacking集成的初级学习器,以线性回归作为次级学习器,进行集成预测。在上海某三甲综合医院的急诊数据集上,考虑气温、降雨、空气质量、节假日等变量预测,试验表明多项指标上,Stacking集成方法优于单模型。预测时间长度越长,预测效果越好。
Emergency patient arrival forecast is the foundation of staff shift arrangement and is important for the overflow control of the hospital.Most research based on emergency patient arrival forecast are sole predicting methods that focus on daily and monthly problems,lacking the forecast of a shorter period.Therefore,a predicting model based on ensemble learning that covers hourly,daily and weekly period is established.Meanwhile,the relationship between different predicting time periods and the prediction error is studied.Basic models like Random Forest(RF),Gradient Boosting Decision Tree(GBDT),eXtreme Gradient Boosting(XGB),Light GBM(LGB),Support Vector Regression(SVR),K-Nearest Neighbor(KNN) are used as the first layer learners,while linear regression is used as the meta learner.The whole model is based on Stacking policy,one of the common ensemble methods.The data is from the hospital in Shanghai,along with data of weather and air quality.The results show that ensemble model is better when compared with single models.Meanwhile,the longer the predicting period,the better the predicting result.
作者
李瑶琦
周鑫
高卫益
柏志安
耿娜
LI Yao-qi;ZHOU Xin;GAO Wei-yi;BAI Zhi-an;GENG Na(Department of Industrial Engineering and Management,Shanghai Jiao Tong University,Shanghai 200240,China;Ruijin Hospital affiliated to School of Medicine,Shanghai Jiao Tong University,Shanghai 200025,China;Sino-US Global Logistics Institute,Shanghai Jiao Tong University,Shanghai 200030,China)
出处
《工业工程与管理》
CSSCI
北大核心
2019年第6期180-187,194,共9页
Industrial Engineering and Management
基金
国家自然科学基金资助项目(71471113,71432006)
关键词
急诊
患者到达
预测
集成学习
STACKING
emergency department
ED patient arrivals
forecasting
ensemble learning
Stacking
