Temporal trends and comparative forecasting of echinococcosis in endemic Xinjiang (2010–2024): changes during and after the COVID-19 pandemic

Abstract

To analyze temporal trends of reported echinococcosis cases in Xinjiang from 2010 to 2024 and evaluate the long-horizon out-of-sample forecasting performance of different time series models during the COVID-19 period and subsequent years. Data were sourced from the China Infectious Disease Surveillance and Reporting System. Annual trends were analyzed using Joinpoint regression to calculate Annual Percentage Change (APC) and Average Annual Percentage Change (AAPC). Monthly data were summarized descriptively to characterize monthly reporting patterns. Five forecasting approaches—Naive, Seasonal Naive, SARIMA, Prophet, and LSTM—were trained using 2010–2019 data and evaluated on 2020–2024 data under a common fixed-origin framework, without updating or refitting using observed test-period values. Performance was assessed using root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and symmetric mean absolute percentage error (sMAPE). Expanding-window rolling-origin one-month-ahead forecasting was additionally conducted as a sensitivity analysis. A total of 20,302 cases were reported, peaking in 2017 and reaching a low in 2022. In the one-joinpoint specification, reported cases increased during 2010–2017 (APC = 8.91%, 95% CI: 2.93% to 22.25%) and decreased during 2017–2024 (APC = − 12.56%, 95% CI: −23.20% to − 7.49%); however, the overall AAPC was − 2.42% (95% CI: −5.80% to 0.83%). Segment-specific trend inference was sensitive to model specification. Monthly reporting patterns also changed over time. In the primary fixed-origin analysis, LSTM (seed = 42) yielded the lowest errors (RMSE = 47.08, MAE = 39.24, MAPE = 105.91%, and sMAPE = 46.07%). In the rolling-origin sensitivity analysis, SARIMA yielded the lowest RMSE, MAE, and sMAPE. Reported echinococcosis cases in Xinjiang showed a rise-then-fall pattern under the one-joinpoint specification, although trend inference was sensitive to model specification. Monthly reporting patterns also changed over time. LSTM (seed = 42) yielded the lowest point-forecast errors in the primary fixed-origin analysis, whereas the rolling-origin sensitivity analysis showed a different model ranking, with SARIMA yielding the lowest RMSE, MAE, and sMAPE, while LSTM showed comparable performance. These findings indicate that comparative model performance depended on the forecasting framework.