Abstract:

Objective: To evaluate the predictive value of the Autoregressive Integrated Moving Average (ARIMA) model and the Holt-Winters exponential smoothing model for pulmonary tuberculosis incidence in the Xinjiang Uygur Autonomous Region (abbreviated as “Xinjiang”), aiming to provide references for formulating pulmonary tuberculosis prevention and control strategies in the region. Methods: Monthly reported cases of pulmonary tuberculosis in Xinjiang from January 2014 to June 2024 were collected from the “Tuberculosis Management Information System”, a subsystem of the “China Disease Prevention and Control Information System”. Data from January 2014 to June 2023 were used as the training set to construct ARIMA and Holt-Winters exponential smoothing models. These two models were then employed to predict the reported incidence of pulmonary tuberculosis from July 2023 to June 2024. Finally, the predictive accuracy of both ARIMA and Holt-Winters models was evaluated by comparing the predicted values with the actual values. Results: Joinpoint regression analysis showed that the reported incidence of pulmonary tuberculosis in Xinjiang exhibited a general trend of first rising and then declining, with fluctuations at low levels in recent years from 2014 to 2024. It steadily increased from 176.00/100000 (39851/22643000) in 2014 to a peak of 304.94/100000 (74549/24446712) in 2018, followed by a rapid decline starting in 2019 (169.05/100000 (42043/24869966)), and further decreased to 77.68/100000 (20182/25980011) in 2024, demonstrating an overall downward trend (AAPC (95%CI)=-9.85% (-17.45% to -1.55%), P=0.026). Monthly seasonal indices for reported tuberculosis cases ranged from 81.51% (February) to 127.01% (June), indicating the time series fluctuated periodically over time with distinct seasonality. Following stationarity testing and systematic model identification—including examination of ACF and PACF plots, automated ARIMA selection via the auto.arima() function in R, and comparison of Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) information criteria—the multiplicative seasonal ARIMA(0,1,0)(1,0,0)₁₂ model was selected as the optimal model (AIC=1857.65; BIC=1863.10; MAPE=19.78%). In parallel, Holt-Winters modeling identified the multiplicative variant as superior to the additive form (MAPE=20.77%). For the predicted tuberculosis incidence from July 2023 to June 2024, the ARIMA(0,1,0)(1,0,0)₁₂ model demonstrated lower prediction errors, with RMSE=309.13, MAE=248.58, and MAPE=13.01%, compared to the Holt-Winters additive model (RMSE=786.48, MAE=732.00, MAPE=45.17%), indicating a better fitting effect. Discussion: The ARIMA model has a significantly better fitting effect for the monthly incidence of tuberculosis in Xinjiang compared to the Holt-Winters exponential smoothing model, and can serve as a more effective predictive tool.

Key words: Tuberculosis, pulmonary, Models, Forecasting, Comparative study