组合模型对恶性肿瘤死亡率拟合度评价及预测方法的研究

Effect of combination model on fitting cancer mortality and prediction

目的 用6种单项预测方法对金昌队列13年恶性肿瘤死亡率进行拟合并预测。方法 采用动态数列、线性回归、指数平滑、自回归移动平均（ARIMA）模型、灰色模型、Joinpoint回归6种方法，利用金昌队列2001-2013年数据进行恶性肿瘤死亡率拟合及方法比较。采用组合模型进行模型优化，基于算术平均法、方差倒数法、均方误差倒数法、简单加权平均法计算组合模型权重系数。结果 对恶性肿瘤死亡率以Joinpoint线性回归拟合精度最高（87.64%），线性回归法、动态数列、GM（1，1）模型、指数平滑法、ARIMA（1，0，0）拟合精度分别为87.32%、86.99%、86.25%、85.72%、81.98%。基于灰色模型与线性回归的组合模型预测精度（〉99%）高于基于ARIMA（1，0，0）与灰色模型的组合模型，其中算术平均法和均方误差倒数法权重系数组合模型（灰色模型与线性回归）拟合效果最好。结论 组合模型预测恶性肿瘤死亡率优于单项预测法，预测精度〉95%。

Objective To reduce the cancer burden in the Jinchang cohort and provide evidence for developing cancer prevention strategies and performing effectiveness evaluation in the Jinchang cohort. We are fitting thirteen years of cancer mortality data from the Jinchang cohort by using six kinds of predicting methods to compare relative fitness and to select good predicting methods for the prediction of cancer mortality trends. Methods The mortality data of cancer in Jinchnag cohort from 2001-2013 were fitted using six kinds of predicting methods：dynamic series, linear regression, exponential smoothing, autoregressive integrated moving average （ARIMA） model, grey model （GM）, and Joinpoint regression. Weight coefficients of combination models were calculated by four methods：the arithmetic average method, the variance inverse method, the mean square error inverse method, and the simple weighted average method. Results The cancer mortality was fitted and compared by using six kinds of forecasting methods; the fitting precision of the Joinpoint linear regression had the highest accuracy （87.64%）, followed by linear regression （87.32%）, the dynamic series （86.99%）, GM （1, 1） （86.25%）, exponential smoothing （85.72%） and ARIMA （1, 0, 0） （81.98%）, respectively. Prediction accuracy of the combination model derived from GM （1, 1） and linear regression （〉99%） was higher than that of the combination model derived from ARIMA （1, 0, 0） and GM （1, 1）. The combination model derived from the GM （1, 1） and linear regression, with weight coefficients based on the arithmetic average method and the mean square error inverse method, had the best prediction effect of the four weight calculation methods. Conclusion Prediction accuracy of the combination model, with accuracy 〉95%, was higher than that of the single prediction methods.