A comparative modeling study evaluating monthly reported pulmonary tuberculosis incidence across Kashgar Prefecture in Xinjiang, China, revealed that model performance varied substantially depending on the selected forecast horizon. At an extended nine-month horizon, a TimeMixer-type neural network architecture achieved the lowest predictive errors among tested approaches, recording a mean absolute error of 6.580 per 100,000 population and an 11.4 percent error reduction relative to naive persistence. Conversely, for shorter-range forecasts spanning three and six months, naive persistence benchmarks consistently outperformed more complex statistical and deep learning formulations, posting mean absolute errors of 2.193 and 3.217 per 100,000, respectively.
The research addressed critical questions surrounding the temporal patterns and predictability of tuberculosis notifications at the local administrative scale. Because tuberculosis notification data reflect a confluence of true disease occurrence, clinical progression, diagnostic case detection, and administrative reporting procedures, the monthly incidence time series demonstrated notable temporal fluctuations across the region. Historical registration data indicated that the mean monthly reported incidence reached 32.05 cases per 100,000 population between 2016 and 2020, representing a higher incidence rate than observed in other periods of the investigation, while overall seasonal strength evaluated through seasonal-trend decomposition using LOESS remained low at 0.090.
To evaluate predictive utility, investigators constructed monthly reported incidence time series using registration records from twelve counties and cities across Kashgar Prefecture spanning from January 2011 through September 2025. The analytical framework trained candidate models on data from January 2011 to December 2024, preserving the nine-month interval from January through September 2025 as an independent temporal holdout test set. Evaluated architectures included advanced neural network designs—specifically iTransformer-type models, TimeMixer-type models, and a bidirectional recurrent neural network—alongside traditional forecasting benchmarks comprising naive persistence, seasonal-naive, and seasonal autoregressive integrated moving average models.
Model evaluation was executed across multiple verification protocols to probe consistency across forecast origins, assess sensitivity to random parameter initialization, and compute forecast errors across three distinct horizons. In short-term assessments, complex neural networks struggled to improve upon basic historical persistence. For the three-month forecast horizon, naive persistence maintained the lowest mean absolute error at 2.193 per 100,000, whereas the TimeMixer architecture yielded a higher mean absolute error of 4.804 per 100,000. At the intermediate six-month horizon, naive persistence likewise retained superior accuracy with a mean absolute error of 3.217 per 100,000, indicating that recent reported incidence served as an effective benchmark over brief observation windows.
At the nine-month horizon, the TimeMixer model demonstrated the strongest overall forecasting accuracy, recording the lowest root mean squared error of 15.960 per 100,000 alongside its mean absolute error of 6.580 per 100,000, while achieving a coefficient of determination of 0.039. Comparative baseline persistence at this horizon registered a higher mean absolute error of 7.430 per 100,000. Additional diagnostic evaluations included screening for data anomalies, which identified three county-month observations in 2025 based on an outlier-screening rule established from training data. When researchers excluded these flagged outliers in a sensitivity analysis, forecast errors decreased across several models, yet the relative ranking of model accuracy at the nine-month horizon remained entirely unchanged.
Despite the modest accuracy gains demonstrated by TimeMixer over longer horizons, the investigation revealed considerable uncertainty surrounding model predictions. At the nine-month mark, empirical coverage of calculated prediction intervals reached 97.2 percent, but the mean interval widths spanned from 49.87 to 58.94 per 100,000, substantially exceeding the mean observed reported incidence documented over that same timeframe. Consequently, the study concluded that while multimodel forecasts can supply valuable supplementary information for routine epidemiological surveillance and data verification, public health resource allocation should remain grounded in broader epidemiological assessments and localized service needs rather than automated projections alone, with model transferability requiring further evaluation in alternate settings and subsequent periods.
Subject of Research: Comparative evaluation of neural network and statistical models for county-level tuberculosis forecasting
Article Title: Multimodel forecasting of county-level reported pulmonary tuberculosis incidence in Kashgar Prefecture, China: a comparative evaluation of neural network and statistical models across forecast horizons
Article References: Ku, T., Yimamu, M., Ma, J., Chen, Z., Guo, G., Zhang, L., Li, L., & Zheng, Y. (2026). Multimodel forecasting of county-level reported pulmonary tuberculosis incidence in Kashgar Prefecture, China: a comparative evaluation of neural network and statistical models across forecast horizons. BMC Infectious Diseases. https://doi.org/10.1186/s12879-026-14832-z
Image Credits: AI Generated
DOI: 10.1186/s12879-026-14832-z
Keywords: tuberculosis, pulmonary tuberculosis, Kashgar Prefecture, Xinjiang, TimeMixer, iTransformer, bidirectional recurrent neural network, naive persistence, seasonal-naive, SARIMA, forecast horizons, prediction intervals, mean absolute error, root mean squared error, county-level incidence, disease surveillance, temporal holdout, outlier sensitivity
News Source: Ophelia Keating. (November 21, 2026). TimeMixer Model Reduced Nine-Month Tuberculosis Forecasting Errors in Kashgar County Evaluation. Scienmag.



