What is the clinical evidence from this study?

Study design: Observational. Population: Hand, foot and mouth disease (HFMD) (n=917285). Intervention: Multivariate BP neural network model vs. ARIMA models and univariate BP model. Primary outcome: Mean absolute percentage error (MAPE) on testing set.

October 7, 2019Open Access

Forecasting incidence of hand, foot and mouth disease using BP neural networks in Jiangsu province, China

Key Result

The multivariate BP neural network model incorporating meteorological factors accurately predicted the incidence of hand, foot and mouth disease with a mean absolute percentage error of 18.57%.

Study Design

Type

Observational (n=917,285)

Structured PICO

Does a multivariate BP neural network model with meteorological factors improve the prediction accuracy of HFMD incidence compared to ARIMA models?

Population

917,285 cases of hand, foot and mouth disease (HFMD) detected during 2009-2016 in Jiangsu province, China.

Intervention

Multivariate back propagation (BP) neural network model incorporating HFMD incidences lagged from 1 to 4 months, mean temperature, rainfall, and their one order lagged terms.

Comparator

Univariate BP model, univariate ARIMA model, and multivariate ARIMAX model.

Outcome

Prediction accuracy of HFMD incidence, evaluated by Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE).

A multivariate BP neural network model incorporating meteorological factors accurately predicts HFMD incidence and outperforms traditional ARIMA models.

Main Result

Absolute Event Rate: 18.57% vs 36.67%

Limitations

Only meteorological variables were considered, ignoring other potential predictors such as social environmental or etiological factors.
Reported data may underestimate HFMD incidence due to mild cases using home therapies or misdiagnosed atypical cases.
Generalizability to other regions with different epidemic characteristics and climates may be limited.
The BP model cannot explain the specific association between risk factors and the disease.

Abstract

Abstract Background Hand, foot and mouth disease (HFMD) is a rising public health problem and has attracted considerable attention worldwide. The purpose of this study was to develop an optimal model with meteorological factors to predict the epidemic of HFMD. Methods Two types of methods, back propagation neural networks (BP) and auto-regressive integrated moving average (ARIMA), were employed to develop forecasting models, based on the monthly HFMD incidences and meteorological factors during 2009–2016 in Jiangsu province, China. Root mean square error (RMSE) and mean absolute percentage error (MAPE) were employed to select model and evaluate the performance of the models. Results Four models were constructed. The multivariate BP model was constructed using the HFMD incidences lagged from 1 to 4 months, mean temperature, rainfall and their one order lagged terms as inputs. The other BP model was fitted just using the lagged HFMD incidences as inputs. The univariate ARIMA model was specified as ARIMA (1,0,1)(1,1,0) 12 (AIC = 1132.12, BIC = 1440.43). And the multivariate ARIMAX with one order lagged temperature as external predictor was fitted based on this ARIMA model (AIC = 1132.37, BIC = 1142.76). The multivariate BP model performed the best in both model fitting stage and prospective forecasting stage, with a MAPE no more than 20%. The performance of the multivariate ARIMAX model was similar to that of the univariate ARIMA model. Both performed much worse than the two BP models, with a high MAPE near to 40%. Conclusion The multivariate BP model effectively integrated the autocorrelation of the HFMD incidence series. Meanwhile, it also comprehensively combined the climatic variables and their hysteresis effects. The introduction of the climate terms significantly improved the prediction accuracy of the BP model. This model could be an ideal method to predict the epidemic level of HFMD, which is of great importance for the public health authorities.

Bookmark

View Full Paper