Seasonality and Trend Forecasting of Tuberculosis Prevalence Data in Eastern Cape, South Africa, Using a Hybrid Model

Abstract

Background: Tuberculosis (TB) is a deadly infectious disease caused by Mycobacteria tuberculosis. Tuberculosis as a chronic and highly infectious disease is prevalent in almost every part of the globe. More than 95% of TB mortality occurs in low/middle income countries. In 2014, approximately 10 million people were diagnosed with active TB and two million died from the disease. In this study, our aim is to compare the predictive powers of the seasonal autoregressive integrated moving average (SARIMA) and neural network auto-regression (SARIMA-NNAR) models of TB incidence and analyse its seasonality in South Africa.

Methods: TB incidence cases data from January 2010 to December 2015 were extracted from the Eastern Cape Health facility report of the electronic Tuberculosis Register (ERT.Net). A SARIMA model and a combined model of SARIMA model and a neural network auto-regression (SARIMA-NNAR) model were used in analysing and predicting the TB data from 2010 to 2015. Simulation performance parameters of mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), mean percent error (MPE), mean absolute scaled error (MASE) and mean absolute percentage error (MAPE) were applied to assess the better performance of prediction between the models.

Results: Though practically, both models could predict TB incidence, the combined model displayed better performance. For the combined model, the Akaike information criterion (AIC), second-order AIC (AICc) and Bayesian information criterion (BIC) are 288.56, 308.31 and 299.09 respectively, which were lower than the SARIMA model with corresponding values of 329.02, 327.20 and 341.99, respectively. The seasonality trend of TB incidence was forecast to have a slightly increased seasonal TB incidence trend from the SARIMA-NNAR model compared to the single model.

Conclusions: The combined model indicated a better TB incidence forecasting with a lower AICc. The model also indicates the need for resolute intervention to reduce infectious disease transmission with co-infection with HIV and other concomitant diseases, and also at festival peak periods.

Keywords: autocorrelation; co-infection; neutral-network; non-seasonality; prediction.

Figure 1

Map showing Eastern Cape Province, South Africa. Map data ©2016 AfriGIS (Pty) Ltd., Google.

Figure 2

Monthly reported cases of TB prevalence data from 2010 to 2015.

Figure 3

Additive decomposition of monthly time series cases of TB prevalence data.

Figure 4

Seasonally adjusted values showing the effects on the monthly reported TB case prevalence.

Figure 5

Time plot, ACF and PACF plot for differenced seasonality adjusted monthly TB cases prevalence.

Figure 6

Standardized residuals from the SARIMA model applied to TB prevalence.

Figure 7

Forecast from SARIMA model applied to TB case prevalence.

Figure 8

Forecast from SARIMA-NNAR model applied to TB case prevalence.

Figure 9

Forecast from ARIMA model with non-zero mean applied to TB case prevalence.