Abstract:

 Based on the data of stroke emergency visits from four hospitals in Beijing during 2008 to 2012, as well as the observed daily meteorological and environmental factors, meteorological and environmental predictors were selected. All the data were divided into two groups, that is, calibration set (2008-2011) and validation set (2012) . Causing-stroke weather forecasting model was constructed using four methods including Stepwise Regression Model (SRM), BP Neural Network Model, Decision Tree Model (DTM) and Generalized Additive Model (GAM). The best model for forecasting the number of stroke patients was determined by comparing the goodness of fit and forecasting accuracy of different models. The results show that the selected predictors vary are varied along with different seasons and models.  and The time series factor is the most important factorindispensable. GAM produces , SRM, and DTM methods give the best performance in at forecasting number of the high-level-stoke., SRM gives the best result for the prediction of the medium-level-stroke.  andDTM provides the best estimate for the low-level-stroke patients, respectively.. The sequence of averaged forecasting accuracy over the four seasons from of the three different models is as follows: GAM＞BP Neural Network Model＞SRM＞DTM. GAM has the highest forecasting accuracy for the number of averaged and high–level-stroke patients . For, the cerebral hemorrhagic stroke (CHS), the forecasting accuracy is 68.3% if the forecasted grade is exactly correct. For the bias grade  ≤±1difference between forecasted grade and actual grade no more than 1, the forecasting accuracy is 98%. The result indicates  reveals that GAM can basically meet the demand for medical meteorological forecasting of CHS.

Key words: Medical meteorology, Stroke, Generalized Additive Model (GAM), BP neural network, Decision tree method