基于滞后效应的多变量LSTM模型对儿童呼吸道疾病就诊人数的预测

  目的  构建儿童常见呼吸道疾病日就诊人数的预测模型, 分析未来就诊人数的变化趋势, 为儿童常见呼吸道疾病的科学防控提供数据支撑。  方法  利用某医院2017年1月1日-2019年12月31日每日就诊病例及同期气象和大气污染物数据, 采用分布滞后非线性模型(distribution lag nonlinear models, DLNM)分别分析日均气温及污染物浓度对春、秋季学期日就诊人数的影响及滞后效应, 在此基础上构造多变量长短期记忆(long short-term memory, LSTM)模型对春、秋季学期日就诊人数进行预测。  结果  选取春、秋季学期日均气温的中位数进行研究, 发现日均气温对秋季学期日就诊人数的影响滞后7 d其后持续约10 d, 而对春季学期表现为即时效应且持续约4 d。结合滞后效应构造的多变量LSTM模型对春、秋季学期日就诊人数均能较好地预测, 测试集上的平均绝对百分比误差(mean absolute percentage error, MAPE)分别为4.59%和4.77%。  结论  考虑滞后效应的多变量LSTM模型能够较准确地对儿童常见呼吸道疾病日就诊人数进行预测, 为疾病的预防和控制提供科学依据。

Prediction for the outpatient amount of childhood common respiratory diseases based on multivariate LSTM model with lag effect

  Objective  To construct the prediction model of the daily outpatient amount of childhood common respiratory diseases and analyze the trend of the outpatient amount in the future, which will provide data support for the scientific prevention and control of common respiratory diseases in children.  Methods  Based on the daily outpatient cases of a hospital and meteorological and air pollutant data from January 1, 2017 to December 31 2019, the distributed lag nonlinear model (DLNM) was used to analyze the influence and lag effect of average daily temperature and pollutant concentration on the daily outpatient amount in spring and autumn semesters.A multivariate long and short-term memory (LSTM) model was constructed to predict daily outpatient amounts in the spring and autumn semesters.  Results  The median average daily temperature in the spring and autumn semesters was selected for research, and we found that the impact of average daily temperature on the daily outpatient amount in the autumn semester lagged 7 days and lasted for about 10 days, while the effect on the spring semester was immediate and lasted for about 4 days.The multivariate LSTM model combined with the lag effect can predict daily outpatient amount in spring and autumn semesters well, the mean absolute percentage error (MAPE) on the test set was 4.59% and 4.77%, respectively.  Conclusion  The multivariate LSTM model combined with the lag effect can accurately predict the daily outpatient amount, which provides a scientific basis for the prevention and control of diseases.