Predictive models for overall mortality risk in the elderly population in Ningbo, China: a comparative analysis

Objective To compare the effects of the multivariate long short term memory neural network (LSTM) model versus the autoregressive integrated moving average model-X (ARIMAX) model in predicting total deaths among the elderly population in Ningbo, China.

Methods The total number of deaths among the elderly population, meteorological data, and air pollutant data were collected in Ningbo from January 1, 2014 to December 31, 2018. The multivariate LSTM and ARIMAX models were established with the weekly data from January 1, 2014 to February 28, 2018 as the training set and the weekly data from March 1 to December 31, 2018 as the test set for prediction. The predictive effects of the two models were compared based on predictive curves and error assessment indices such as root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), goodness of fit (R²).

Results In this study, a total of 146 606 deaths were collected in Ningbo from January 1, 2014 to December 31, 2018, with a median of 76 deaths per week. The weekly mean maximum temperature, relative humidity, and precipitation were 23.37℃, 80.53%, and 0.86 mm, respectively, and the weekly mean mass concentrations of SO₂, NO₂, CO, O₃-8 h, PM₁₀, and PM_2.5 were 11.71 μg/m³, 37.43 μg/m³, 0.79 mg/m³, 95.43 μg/m³, 58.43 μg/m³, and 35.93 μg/m³, respectively. Based on the Spearman rank correlation analysis of meteorological factors and air pollutants, five factors were finally included in analysis, i.e., maximum temperature, relative humidity, precipitation, O₃-8 h, and PM_2.5. The multivariable LSTM and ARIMAX (3, 1, 2) models were used for prediction, and the multivariate LSTM model had an RMSE of 4.90, an MAE of 3.77, an MAPE of 4.77, and an R² of 0.82, while the ARIMAX (3, 1, 2) model had an RMSE of 8.68, an MAE of 5.80, an MAPE of 7.53, and an R² of 0.97. The ARIMAX had a higher degree of curve fitting than the LSTM model.

Conclusion Multivariate ARIMAX has better predictive ability than multivariate LSTM for the weekly mean number of deaths among the elderly population in Ningbo from March 1 to December 31, 2018.