https://orcid.org/0009-0003-8830-5203
Figures
Abstract
Background
While time-series models have been applied to forecast brucellosis incidence in China, systematic comparisons of multiple models remain relatively limited. This study aimed to elucidate the epidemic characteristics of human brucellosis and to provide a comparative assessment of several time-series prediction models, in order to identify a suitable predictive framework for future incidence forecasting.
Methods
Monthly and annual incidence rates (per 100,000 population) of brucellosis in China from January 2011 to December 2020 were used as raw data. Seven time-series models were developed and compared using R software (version 4.3.1): Seasonal Autoregressive Integrated Moving Average (SARIMA), Holt-Winters additive model, Holt-Winters multiplicative model, Neural Network Autoregressive (NNAR) model, Exponential Smoothing State Space (ETS) model, TBATS model, and Prophet model. A rolling-window cross-validation was applied to assess model stability. Model performance was evaluated using root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and mean absolute scaled error (MASE).
Results
Among the seven models evaluated, the Holt-Winters multiplicative model demonstrated the most stable and superior predictive performance on the test set (MAE = 0.034, RMSE = 0.040, MAPE = 14.881%, MASE = 0.891), which serves as strong evidence for its best generalization capability among the compared models.
Conclusions
Given its stable and superior performance in the test set, the Holt-Winters multiplicative model is recommended for short-term brucellosis forecasting in China. It captures the characteristic spring-summer peak, and its integration into surveillance systems could enhance early warning and targeted interventions.
Citation: Jiang Y, Zhao J, Long J, Deng P, Qin S, Zhang Y (2026) Modeling the seasonal epidemic of human brucellosis in China: A comparative time series analysis. PLoS One 21(3): e0344908. https://doi.org/10.1371/journal.pone.0344908
Editor: Maryam Afshari, Hamadan University of Medical Sciences, IRAN, ISLAMIC REPUBLIC OF
Received: September 9, 2025; Accepted: February 26, 2026; Published: March 25, 2026
Copyright: © 2026 Jiang et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the manuscript and its Supporting information files.
Funding: This study was supported by grants from the Scientific Research Project of the 2024 National Public Health Talent Training Support Project, National Key Research and Development Program Project (No. 2024YFC2311500), and the “Kunlun Talents - Plateau Famous Doctors” Project (Qing Health Office [2021] No.104).
Competing interests: The authors have declared that no competing interests exist.
1. Introduction
Brucellosis is a zoonotic infectious disease caused by Brucella. It mainly affects domestic animals such as cattle, sheep, and pigs, and can also be transmitted to humans. It is also a Class B infectious disease specified in the Law of the People’s Republic of China on the Prevention and Control of Infectious Diseases [1–3]. Globally, the disease remains prevalent in regions including the Middle East, the Mediterranean, and parts of the Americas, with an estimated 16,000–20,000 human cases annually [4]. Osteoarticular involvement is the most common complication, usually manifesting as spondylitis, sacroiliitis, and peripheral arthritis, with great social harm [5,6]. The risk of brucellosis is increasing in environments with poverty and insufficient resources [7].
Accurate forecasting of brucellosis incidence is crucial for effective public health preparedness and resource allocation [8–11]. Strengthening dynamic monitoring and predictive capabilities is fundamental to reducing the disease burden in China [12]. While predictive modelling offers a pathway to improved early warning, existing research has predominantly relied on the Autoregressive Integrated Moving Average (ARIMA) model [13–15]. Although effective for linear trends and seasonality, the ARIMA framework may not fully capture complex nonlinear dynamics inherent in epidemiological time series data, potentially limiting forecasting accuracy and adaptability [16,17].
While time-series models have been applied to forecast brucellosis incidence in China, systematic comparisons of multiple models remain relatively limited [13,14]. To address this gap, this study contributes several key advancements. First, it systematically compares seven distinct models: Seasonal Autoregressive Integrated Moving Average (SARIMA), Holt‑Winters additive, Holt‑Winters multiplicative, Neural Network Autoregression (NNAR), Exponential Smoothing State Space (ETS), Trigonometric seasonality, Box-Cox transformation, ARMA errors, Trend and Seasonal components (TBATS), and Prophet model. Second, model performance is rigorously assessed using both rolling-window cross-validation on the training set and an independent test set to evaluate real‑world forecasting ability. Third, prediction accuracy is comprehensively measured with multiple error metrics, namely root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and mean absolute scaled error (MASE), enabling a robust multi‑criterion evaluation. Through this approach, the study aims to identify the most reliable forecasting framework for brucellosis incidence in China.
2. Materials and methods
2.1. Ethics statement
In accordance with the Law of the People’s Republic of China on the Prevention and Control of Infectious Diseases, brucellosis is a statutory notifiable disease incorporated into the national public health surveillance system. The collection of anonymized, population-level surveillance data for such diseases is a mandated public health function and does not require separate ethical committee review. The data used in this study were obtained from the National Infectious Disease Surveillance Network. All data were anonymized and de-identified prior to analysis, and no personal privacy or human subject research was involved [18,19].
2.2. Data source
Monthly incidence rates of brucellosis (per 100,000 population) in China from January 2011 to December 2020 were obtained from the Chinese Public Health Science Center. This dataset consists of 120 consecutive months of population-standardized surveillance data, which were reviewed and aggregated by the Centers for Disease Control and Prevention. Data continuity was confirmed, with no missing months identified. To preserve the original epidemiological signal, no smoothing or outlier removal was applied. The complete series was partitioned into a training set (January 2011-December 2019) and an independent test set (January-December 2020) for model development and evaluation.
2.3. Data processing and analysis
The monthly brucellosis incidence data were programmatically collated and structured as a time series object in R (version 4.3.1) using the ‘ts’ function with a frequency of 12, spanning from January 2011 to December 2020. Data continuity was verified by comparing the sequence of months, confirming a complete series of 120 observations with no missing values; consequently, no imputation was required. To maintain the integrity of the raw epidemiological signal, no data smoothing or explicit outlier detection and removal was performed prior to model fitting.
All time-series modeling, forecasting, and evaluation were conducted within the R environment. The analysis leveraged a suite of specific packages and versions to ensure reproducibility: forecast (version 8.21) for fitting SARIMA, ETS, TBATS, and NNAR models as well as for generating predictions; tseries for conducting the Augmented Dickey-Fuller (ADF) and Ljung-Box statistical tests; prophet (version 1.0) for building the Prophet model; and supporting packages including tidyverse, tsibble, and urca for data manipulation and analysis.
A rolling-window cross-validation scheme was implemented on the training set to evaluate model stability. The configuration used a window length of 60 months, encompassing five full annual cycles to adequately capture long-term seasonal patterns. The forecast horizon was set to 6 months, and the window was rolled forward by a step of 3 months. This step size balanced computational efficiency with the need for sufficient overlapping windows to ensure a stable assessment. This process generated 15 distinct validation windows in total.
The optimal predictive model was defined as the one with the lowest combined performance on MAPE and RMSE while maintaining competitive scores on the complementary metrics. The brucellosis incidence data in this study featured a low and stable average level of 3.29 per 100,000 population with distinct annual seasonality, which supported the selection of MAPE and RMSE as dual primary metrics for a comprehensive performance evaluation.
2.4. SARIMA model
The SARIMA model is derived from the time series model proposed by British statistician Jenkins in the 1970s. It aims to determine the order of the model using autocorrelation and partial autocorrelation functions, enabling it to better handle time series data with seasonal patterns [20,21]. The strength of the SARIMA model lies in its ability to capture both seasonal and trend components of time series data. This allows it, through reasonable selection of model parameters, to be more suitable for accurate short-term and medium-term predictions [22].
In this study, the SARIMA model was implemented using the auto.arima() function from the forecast package (version 8.21) in R. The model orders (p, d, q) and seasonal orders (P, D, Q) were automatically selected through a comprehensive stepwise search procedure configured with stepwise = FALSE and approximation = FALSE. The selection aimed to minimize the Akaike Information Criterion (AIC), with the search space constrained to maximum non-seasonal orders of p = 3 and q = 3 and maximum seasonal orders of P = 2 and Q = 2. This parameter range is conventional for monthly time series and serves to mitigate overfitting while capturing essential dynamics. As shown in Equation (1):
Where () is the seasonal delay operator, and
and
are delay polynomial operators.
2.5. Holt−Winters model
This model combines components of trend, seasonality, and random error, integrating the characteristics of single exponential smoothing, double exponential smoothing, and triple exponential smoothing [23]. It can be divided into additive and multiplicative models. It assigns higher weights to recent observations in the time series, while the weights of distant observations decrease in a geometric progression. The selection of model parameters has a great impact on model performance, and methods such as cross-validation are usually needed to optimize these parameters.
In this study, both the additive and multiplicative Holt-Winters models were fitted using the HoltWinters() function in R. The smoothing parameters (alpha for level, beta for trend, and gamma for seasonality) were optimized by minimizing the squared one-step prediction error on the training set. For the final multiplicative model selected, the optimized parameters were alpha = 0.356, beta = 0.126, and gamma = 1.000.
2.6. NNAR model
The NNAR model can be regarded as a network composed of a series of neurons or nodes, which can describe and model complex nonlinear relationships and function forms [24,25]. If the network contains no hidden layers, the generated model will only have the characteristics of linear regression, showing a simple linear relationship.
In this study, the NNAR model was fitted using the nnetar() function from the forecast package. The network architecture was configured with 12 lagged inputs (p = 12) to capture annual patterns and 1 seasonal lag (P = 1) for seasonal adjustment. The number of neurons in the single hidden layer was set to 8, which was determined as optimal through a grid search aimed at balancing model capacity and the risk of overfitting. To ensure robust training, the model was fitted with 30 random repetitions to mitigate the effect of initial weight randomness, and the training process was limited to a maximum of 150 iterations. The specific formula of the model is expressed as (2):
Where is the current value at time t;
is the neural network, which accepts historical observations and errors as inputs;
are lagged input items of
is the parameter of the model; and
is the error term of the time series. The NNAR model can be regarded as a powerful prediction tool. It combines the flexibility of neural networks with the time series prediction ability of the ARIMA model, and adapts to different time series patterns by adjusting parameters adaptively.
2.7. ETS model
The ETS model is suitable for time series data with stable trends and seasonality, and can automatically adjust smoothing coefficients according to data characteristics [26]. The main parameters of the model include the smoothing coefficient, trend coefficient, and seasonal index. The seasonal index is used to control the impact of seasonality, and its value range in the model is usually between 0 and 1. During model construction, it is necessary to select the appropriate type of ETS model according to the trend and seasonal characteristics of the data.
In this study, the ETS model was implemented using the ets() function. The model’s error, trend, and seasonal components (ETS) were automatically selected from a range of possible combinations. The selection process was based on optimizing the corrected Akaike Information Criterion (AICc). The final fitted model was of type ETS (A, N, A), indicating an additive error, no trend, and additive seasonality structure.
2.8. TBATS model
The TBATS model innovatively introduces an exponential smoothing state space modeling framework. It is an advanced method in time series analysis, which can capture trends, seasonality, and periodicity in time series data simultaneously [27]. Based on seasonal decomposition and locally weighted regression, it can smooth and predict time series data, and can smooth seasonal components to reduce noise and random fluctuations [28].
In this study, the TBATS model was fitted using the tbats function. The model configuration incorporated a Box-Cox transformation, included both a trend component and a damped trend component, and accounted for ARMA errors in the residuals. The seasonal period was explicitly set to 12 months to match the frequency of the monthly incidence data.
2.9. Prophet model
The Prophet model is a modular additive regression model designed to forecast time series data with multiple seasonal patterns. It decomposes a time series into trend, seasonal, and holiday components, and is particularly robust to missing data and abrupt changes in trend [29–31]. The model uses a curve-fitting approach that allows for flexible adjustment of seasonal patterns and trend changepoints, making it suitable for datasets with pronounced and potentially changing periodicities.
In this study, the Prophet model was implemented using the prophet package version 1.0 in R. A linear growth trend was specified for the model. Multiplicative annual seasonality was enabled to better capture the observed pattern where seasonal fluctuations in brucellosis incidence scale with the overall level of the series. The flexibility of the seasonal component was controlled by setting the seasonality prior scale to 10. The trend flexibility was regulated by setting the changepoint prior scale to 0.05. The model was fitted on the training set and used to generate forecasts for the test period.
3. Results
3.1. Descriptive statistics
A total of 450,284 brucellosis cases were reported in China from 2011 to 2020, with 8 cumulative deaths and an average annual incidence rate of 3.29 per 100,000 population. Annual incidence rates were 2.85, 2.93, 3.21, 4.22, 4.18, 3.44, 2.79, 2.73, 3.15, and 3.37 per 100,000 population in sequence, with the lowest incidence observed in 2018. Our study demonstrated that the period from 2014 to 2015 was a peak. After that, it declined briefly but rose again in 2020, showing a trend of decreasing, then increasing, then decreasing, and subsequently resurging (Table 1).
3.2. Specific distribution
Between 2011 and 2020, the distribution of human brucellosis cases in China exhibited substantial spatial heterogeneity across provinces. During this period, multiple provinces reported cases, with particularly high disease burdens observed in Inner Mongolia Autonomous Region, Xinjiang Uygur Autonomous Region, Ningxia Hui Autonomous Region, Shanxi Province, and Heilongjiang Province. The average annual incidence rates in these regions reached 43.23, 21.11, 25.82, 13.95, and 13.69 per 100,000 population, respectively. Hebei Province, Gansu Province, and Qinghai Province were also identified as relatively high-risk areas, each reporting incidence rates exceeding 5 per 100,000 population (Table 2).
Regarding age distribution, the majority of brucellosis cases occurred among young and middle-aged adults. The incidence rate demonstrated an initial increase followed by a decline with advancing age. The most affected demographic was the 40–64 age group, with the highest number of cases reported in the 45–49 subgroup (65,796 cases), followed by the 50–54 subgroup (64,305 cases). In terms of incidence, the 50–54 age group exhibited the highest average annual rate at 7.46 per 100,000, followed by the 55–59 group (5.96 per 100,000) and the 60–64 group (6.40 per 100,000). These findings indicate that middle-aged adults represent the population at greatest risk for brucellosis infection (Table 3).
3.3. Trend analysis and seasonal decomposition
The monthly incidence time series of brucellosis in China from January 2011 to December 2020 is presented in Figs 1 and 2. The series exhibited clear annual fluctuations, with a consistent seasonal peak occurring between April and June each year. Statistical tests confirmed the characteristics of the series: the ADF test indicated stationarity (p < 0.05), while the Ljung-Box test rejected the null hypothesis of white noise (p < 0.001), confirming the presence of serial correlation suitable for modeling Fig 3.
Seasonal-Trend decomposition using LOESS (STL) was applied to decompose the original series into trend, seasonal, and remainder components (Fig 4). The decomposition revealed that the incidence rate peaked around 2015, followed by a general declining trend, with a notable resurgence observed in 2020. The seasonal component clearly captured the recurring spring-summer epidemic pattern, providing a distinct seasonal signal for subsequent forecasting models.
3.4. Model performance evaluation
On the training set, the lowest error metrics were recorded by the NNAR model, with MAE, RMSE, and MAPE values of 0.001, 0.001, and 0.435%, respectively. Comparatively low error values were also observed for the TBATS, SARIMA, and ETS models, indicating a good fit to the training data (Fig 4 and Table 4).
On the independent test set (year 2020), the Holt-Winters multiplicative model demonstrated strong predictive performance, achieving the lowest RMSE (0.040) and MAE (0.034), with a MAPE of 14.88%. Although the SARIMA model yielded a slightly lower MAPE (13.48%), its higher RMSE (0.041) suggested less stability in handling extreme values. A notable discrepancy was identified for the NNAR model, whose MAPE increased substantially from 0.435% on the training set to 16.430% on the test set, suggesting a risk of overfitting. In contrast, higher test-set MAPE values were observed for the TBATS model (27.065%) and the Holt-Winters additive model (25.516%) (Fig 5).
3.5. Rolling-window cross-validation results
The cross-validation results indicated that the TBATS model achieved the lowest average MAPE (14.56%), followed by Prophet (15.49%) and ETS (15.97%). In this internal validation, the Holt-Winters multiplicative model yielded a higher average MAPE of 20.43%. This apparent discrepancy is reconciled by examining performance on the completely unseen test set. As shown in Section 3.4, the Holt-Winters multiplicative model attained the lowest MAPE (14.88%) on the independent 2020 data, demonstrating superior generalization.
This contrast suggests a key difference in model behaviour. Flexible models like TBATS may have adapted closely to specific local patterns and noise within the historical rolling windows of the validation, leading to superior in-sample scores. In contrast, the Holt-Winters multiplicative model, with its parsimonious multiplicative seasonal structure, likely captured a more fundamental and stable seasonal trend. While this resulted in a slightly poorer fit to certain historical sub-periods during cross-validation, it proved to be a more robust and transferable pattern for forecasting the truly unseen data of 2020, leading to more reliable out-of-sample predictions (Table 5).
3.6. Model prediction for 2021
Based on its superior and stable performance on the test set, the Holt-Winters multiplicative model was selected to generate a forecast for the year 2021. The model was applied to the complete time series, concluding in December 2020. The resulting forecast for the 12 months of 2021 indicated a unimodal distribution for the monthly incidence rate, with a peak predicted during the spring-summer period. The specific forecasted values are presented in Table 6. This application demonstrates the model’s utility for prospective, short-term epidemiological forecasting.
3.7. Residual analysis of model diagnosis
A comprehensive diagnostic analysis of the Holt-Winters multiplicative seasonal model’s residuals was conducted to evaluate its specification. The statistical metrics revealed a residual mean near zero (−0.002) and a standard deviation of 0.029, with a kurtosis of 4.108 indicating a slightly leptokurtic distribution. The Ljung-Box test (p < 0.000) and the Shapiro-Wilk test (p < 0.001) both rejected their null hypotheses, suggesting the presence of residual autocorrelation and a deviation from normality. Importantly, the identified autocorrelation is primarily linked to a subtle, unmodeled seasonal component, as the model demonstrably captured the dominant trend and core seasonal patterns with high accuracy, evidenced by its superior performance on the test set.
Visual inspection of the residuals, as shown in Fig 6, supports this interpretation. While the sequence plot shows fluctuations around zero without a clear trend, the autocorrelation function plot reveals a significant peak at the seasonal lag of 12 (approximately 0.4). This confirms the persistence of a minor seasonal signal within the residuals. It is noteworthy, however, that the amplitude of this residual seasonality is small relative to the model’s fitted values. This implies that the Holt-Winters model successfully captured the primary, deterministic seasonal structure of the brucellosis incidence data, with the residual autocorrelation representing a secondary, less pronounced pattern. Overall, the model demonstrates a robust and effective decomposition of the time series into its fundamental components.
4. Discussion
Brucellosis remains a significant public health concern in China, associated with a considerable disease burden as evidenced by the reported cases from 2011 to 2020 [32–34]. The observed incidence pattern, characterized by a distinct unimodal seasonal peak from April to June, aligns with findings from other regions and is consistent with known epidemiological drivers [35–39]. This seasonal surge is likely attributable to heightened livestock reproductive activities and consequent occupational exposure during spring and summer, coupled with environmental conditions that favor pathogen survival [40,41]. It is necessary to increase disease prevention measures to reduce healthcare-related burdens [42]. The resurgence of cases in 2020 highlights the persistent and dynamic nature of the epidemic, necessitating ongoing vigilance.
A primary contribution of this study lies in the systematic comparison of seven time-series forecasting models. The analysis revealed that model performance on the training set was not a reliable indicator of true predictive capability. For instance, the NNAR model achieved a near-perfect in-sample fit but exhibited a marked performance decline on the independent test set, a clear signature of overfitting. This suggests that its complex, nonlinear structure may have adapted too closely to idiosyncratic patterns or noise within the specific training period, compromising its ability to generalize [43–45]. Similarly, advanced models like TBATS and Prophet, despite their theoretical strengths in handling complex seasonality, did not achieve superior out-of-sample accuracy in this application, indicating potential over-parameterization or a mismatch with the fundamental structure of the incidence series.
Conversely, the Holt-Winters multiplicative model demonstrated robust and superior predictive accuracy on the test set. Its stability was further corroborated by its performance in the rolling-window cross-validation. The effectiveness of this model can be attributed to its structural congruence with the key characteristic of the brucellosis incidence data: a seasonal amplitude that scales multiplicatively with the local level of the series. This feature allows it to accurately capture the observed phenomenon where the absolute size of the annual spring-summer peak varies in proportion to the underlying trend, a relationship that additive seasonal models cannot represent adequately [46–48]. Thus, its optimal performance stems from a parsimonious design that correctly specifies the interaction between trend and seasonality inherent in the data.
Several limitations of this study should be acknowledged. The models were developed using historical incidence data alone, without incorporating potential external predictors such as climatic variables, livestock population data, or detailed intervention records, which might improve forecast precision. Secondly, while cross-validation was employed, the models are inherently more suited to short-term forecasting given the stable seasonal patterns in the near term. Future research could explore integrated models that combine time-series approaches with explanatory variables. Thirdly, the analysis was constrained to nationally aggregated monthly data up to 2020. Future studies could benefit from incorporating more recent data, higher-frequency (e.g., weekly) reporting, and spatially explicit modeling at the prefecture or county level to capture local heterogeneities. Finally, although the data originated from the national surveillance system, which is authoritative, the possibility of underreporting or misclassification, common in passive surveillance, cannot be entirely excluded.
In conclusion, through a comparative evaluation, the Holt-Winters multiplicative model was identified as the most accurate and stable choice for short-term forecasting of brucellosis incidence in China. Its implementation within surveillance frameworks could enhance early warning capabilities, supporting timely and targeted public health actions before the seasonal high-risk period.
5. Conclusion
This study confirms a distinct spring-summer seasonal peak in human brucellosis incidence in China, with middle-aged adults constituting the primary affected population. Among seven time-series models compared, the Holt-Winters multiplicative model demonstrated the most stable and accurate short-term forecasts. Its superior performance stems from its multiplicative seasonal structure, which effectively captures the key epidemiological pattern wherein seasonal fluctuation amplitudes scale proportionally with the underlying trend level. Integrating this robust model into the national surveillance framework would enable proactive early warnings and support targeted interventions prior to the annual high-risk period, thereby strengthening brucellosis control strategies in China.
Acknowledgments
The authors acknowledge the staff of disease control and prevention centers at all levels and medical workers who contributed to data reporting; it is their efforts that have enabled the improvement of this study.
References
- 1. Qureshi KA, Parvez A, Fahmy NA, Abdel Hady BH, Kumar S, Ganguly A, et al. Brucellosis: epidemiology, pathogenesis, diagnosis and treatment-a comprehensive review. Ann Med. 2023;55(2):2295398. pmid:38165919
- 2. Khairullah AR, Kurniawan SC, Puspitasari Y, Aryaloka S, Silaen OSM, Yanestria SM, et al. Brucellosis: Unveiling the complexities of a pervasive zoonotic disease and its global impacts. Open Vet J. 2024;14(5):1081–97. pmid:38938422
- 3. Tian Z, Wan L, Pei J, Li T, Wang X, Yuan P, et al. Brucellosis seroprevalence in cattle in China during 2014-2024: a systematic review and meta-analysis. Emerg Microbes Infect. 2024;13(1):2417859. pmid:39450582
- 4. Moriyón I, Blasco JM, Letesson JJ, De Massis F, Moreno E. Brucellosis and One Health: Inherited and Future Challenges. Microorganisms. 2023;11:2070.
- 5. Laine CG, Johnson VE, Scott HM. Global estimate of human brucellosis incidence. Emerging Infectious Diseases. 2023;29(9):1789.
- 6. Jin M, Fan Z, Gao R. Research progress on complications of brucellosis. Frontiers in Cellular and Infection Microbiology. 2023;13:1136674.
- 7. Laine CG, Johnson VE, Scott HM, Arenas-Gamboa AM. Global Estimate of Human Brucellosis Incidence. Emerg Infect Dis. 2023;29(9):1789–97. pmid:37610167
- 8. Liu S, Hu J, Zhao Y, Wang X, Wang X. Prediction and control for the transmission of brucellosis in inner Mongolia, China. Sci Rep. 2025;15(1):3532. pmid:39875488
- 9. Zhao D, Zhang H. The research on TBATS and ELM models for prediction of human brucellosis cases in mainland China: a time series study. BMC Infect Dis. 2022;22(1):934. pmid:36510150
- 10. Xian X, Wang L, Wu X, Tang X, Zhai X, Yu R, et al. Comparison of SARIMA model, Holt-winters model and ETS model in predicting the incidence of foodborne disease. BMC Infect Dis. 2023;23(1):803. pmid:37974072
- 11. Zhuang H-H, Qu Q, Long W-M, Hu Q, Wu X-L, Chen Y, et al. Ceftazidime/avibactam versus polymyxin B in carbapenem-resistant Klebsiella pneumoniae infections: a propensity score-matched multicenter real-world study. Infection. 2025;53(1):95–106. pmid:38884857
- 12. Zhao C, Liu K, Jiang C, Wei X, Song S, Wu X, et al. Epidemic characteristics and transmission risk prediction of brucellosis in Xi’an city, Northwest China. Front Public Health. 2022;10:926812. pmid:35937257
- 13. Zhao R, Sun R, Zhang F. Epidemiological characteristics and spatial clustering analysis of human brucellosis in Zibo City, Shandong Province, China, 2006-2024. Front Public Health. 2025;13:1580265.
- 14. Zhai M, Li W, Tie P, Wang X, Xie T, Ren H, et al. Research on the predictive effect of a combined model of ARIMA and neural networks on human brucellosis in Shanxi Province, China: a time series predictive analysis. BMC Infect Dis. 2021;21(1):280. pmid:33740904
- 15. Shirzadi MR, Mohammadi P, Moradi G, Goodarzi E, Khazaei S, Moayed L, et al. The Incidence and Geographical Distribution of Brucellosis in Iran Using Geographic Information System and Prediction of its Incidence in 2021. J Prev Med Hyg. 2021;62(3):E635–E634. pmid:34909491
- 16. Wan Y, Song P, Liu J, Xu X, Lei X. A hybrid model for hand-foot-mouth disease prediction based on ARIMA-EEMD-LSTM. BMC Infect Dis. 2023;23(1):879. pmid:38102558
- 17. Schaffer AL, Dobbins TA, Pearson S-A. Interrupted time series analysis using autoregressive integrated moving average (ARIMA) models: a guide for evaluating large-scale health interventions. BMC Med Res Methodol. 2021;21(1):58. pmid:33752604
- 18. Peng P-Y, Duan H-Y, Xu L, Sun J-Q, Ma L-J, Zu Y, et al. A Seasonal Autoregressive Integrated Moving Average (SARIMA) forecasting model to predict the epidemic trends of scrub typhus in China. PLoS One. 2025;20(6):e0325905. pmid:40549749
- 19. Wang Y, Xu C, Wang Z, Zhang S, Zhu Y, Yuan J. Time series modeling of pertussis incidence in China from 2004 to 2018 with a novel wavelet based SARIMA-NAR hybrid model. PLoS One. 2018;13(12):e0208404. pmid:30586416
- 20. Maipan-Uku JY, Cavus N. Forecasting tuberculosis incidence: a review of time series and machine learning models for prediction and eradication strategies. Int J Environ Health Res. 2025;35(3):645–60. pmid:38916208
- 21. Warnasekara J, Agampodi S, Nr A. SARIMA and ARDL models for predicting leptospirosis in Anuradhapura district Sri Lanka. PLoS One. 2022;17(10):e0275447. pmid:36227833
- 22. Luo Z, Jia X, Bao J, Song Z, Zhu H, Liu M, et al. A Combined Model of SARIMA and Prophet Models in Forecasting AIDS Incidence in Henan Province, China. Int J Environ Res Public Health. 2022;19(10):5910. pmid:35627447
- 23. Davidescu AA, Apostu S-A, Paul A. Comparative Analysis of Different Univariate Forecasting Methods in Modelling and Predicting the Romanian Unemployment Rate for the Period 2021-2022. Entropy (Basel). 2021;23(3):325. pmid:33803384
- 24. Intarapak S, Supapakorn T, Vuthipongse W. Classical Forecasting of International Tourist Arrivals to Thailand. J Stat Theory Appl. 2022;21:31–43.
- 25. Napoli AM, Smith-Shain R, Lin T, Baird J. The accuracy of predictive analytics in forecasting emergency department volume before and after onset of COVID-19. West J Emerg Med. 2024;25:61–6.
- 26. Naher S, Rabbi F, Hossain MM, Banik R, Pervez S, Boitchi AB. Forecasting the incidence of dengue in Bangladesh-Application of time series model. Health Sci Rep. 2022;5(4):e666. pmid:35702512
- 27. Yu C, Xu C, Li Y, Yao S, Bai Y, Li J, et al. Time Series Analysis and Forecasting of the Hand-Foot-Mouth Disease Morbidity in China Using An Advanced Exponential Smoothing State Space TBATS Model. Infect Drug Resist. 2021;14:2809–21. pmid:34321897
- 28. Perone G. Comparison of ARIMA, ETS, NNAR, TBATS and hybrid models to forecast the second wave of COVID-19 hospitalizations in Italy. Eur J Health Econ. 2022;23(6):917–40. pmid:34347175
- 29. Liao S, Yang C, Li D. Improving precise point positioning performance based on Prophet model. PLoS One. 2021;16(1):e0245561. pmid:33465150
- 30. Xie C, Wen H, Yang W, Cai J, Zhang P, Wu R, et al. Trend analysis and forecast of daily reported incidence of hand, foot and mouth disease in Hubei, China by Prophet model. Sci Rep. 2021;11(1):1445. pmid:33446859
- 31. Talkhi N, Akhavan Fatemi N, Ataei Z, Jabbari Nooghabi M. Modeling and forecasting number of confirmed and death caused COVID-19 in IRAN: A comparison of time series forecasting methods. Biomed Signal Process Control. 2021;66:102494. pmid:33594301
- 32. Zheng R, Xie S, Lu X, Sun L, Zhou Y, Zhang Y, et al. A Systematic Review and Meta-Analysis of Epidemiology and Clinical Manifestations of Human Brucellosis in China. Biomed Res Int. 2018;2018:5712920. pmid:29850535
- 33. Liu Z, Gao L, Wang M, Yuan M, Li Z. Long ignored but making a comeback: a worldwide epidemiological evolution of human brucellosis. Emerg Microbes Infect. 2024;13(1):2290839. pmid:38039063
- 34. Bayhan Gİ, Batur A, Ece İ. Pulmonary infections due to brucellosis in childhood. Tuberk Toraks. 2020;68(1):43–7. pmid:32718139
- 35. Ducrotoy M, Bertu WJ, Matope G, Cadmus S, Conde-Álvarez R, Gusi AM, et al. Brucellosis in Sub-Saharan Africa: Current challenges for management, diagnosis and control. Acta Trop. 2017;165:179–93. pmid:26551794
- 36. Bagheri Nejad R, Krecek RC, Khalaf OH, Hailat N, Arenas-Gamboa AM. Brucellosis in the Middle East: Current situation and a pathway forward. PLoS Negl Trop Dis. 2020;14(5):e0008071. pmid:32437346
- 37. Glikman D. Human Brucellosis in Israel - The Saga Continues. Isr Med Assoc J. 2019;21(1):52–3. pmid:30685907
- 38. Al-Amr M, Abasi L, Khasawneh R, Almharat S, Al-Smadi R, Abbasi N, et al. Epidemiology of human brucellosis in military hospitals in Jordan: A five-year study. J Infect Dev Ctries. 2022;16(12):1870–6. pmid:36753655
- 39. Chalabiani S, Khodadad Nazari M, Razavi Davoodi N, Shabani M, Mardani M, Sarafnejad A, et al. The Prevalence of Brucellosis in Different Provinces of Iran during 2013-2015. Iran J Public Health. 2019;48(1):132–8. pmid:30847321
- 40. Sun X, Jiang W, Li Y, Li X, Zeng Q, Du J, et al. Evaluating active versus passive sources of human brucellosis in Jining City, China. PeerJ. 2021;9:e11637. pmid:34221727
- 41. Yang Y, Li C, Fan X, Long W, Hu Y, Wang Y, et al. Effectiveness of Omadacycline in a Patient with Chlamydia psittaci and KPC-Producing Gram-Negative Bacteria Infection. Infect Drug Resist. 2025;18:903–8. pmid:39990782
- 42. Wang Q, Jiang Q, Yang Y, Pan J. The burden of travel for care and its influencing factors in China: An inpatient-based study of travel time. Journal of Transport & Health. 2022;25:101353.
- 43. Almarashi AM, Daniyal M, Jamal F. A novel comparative study of NNAR approach with linear stochastic time series models in predicting tennis player’s performance. BMC Sports Sci Med Rehabil. 2024;16(1):28. pmid:38273407
- 44. Głowacka P, Żakowska D, Naylor K, Niemcewicz M, Bielawska-Drózd A. Brucella - Virulence Factors, Pathogenesis and Treatment. Pol J Microbiol. 2018;67:151–61.
- 45. Yagupsky P, Morata P, Colmenero JD. Laboratory diagnosis of human brucellosis. Clin Microbiol Rev. 2019;33:e00073-19.
- 46. Xu W, Shao Z, Lou H, Qi J, Zhu J, Li D, et al. Prediction of congenital heart disease for newborns: comparative analysis of Holt-Winters exponential smoothing and autoregressive integrated moving average models. BMC Med Res Methodol. 2022;22(1):257. pmid:36183070
- 47. Wang N, Zhuang W, Ran Z, Wan P, Fu J. Prediction of acute onset of chronic cor pulmonale: comparative analysis of Holt-Winters exponential smoothing and ARIMA model. BMC Med Res Methodol. 2024;24(1):204. pmid:39271998
- 48. Cañedo MC, Lopes TIB, Rossato L, Nunes IB, Faccin ID, Salomé TM, et al. Impact of COVID-19 pandemic in the Brazilian maternal mortality ratio: A comparative analysis of Neural Networks Autoregression, Holt-Winters exponential smoothing, and Autoregressive Integrated Moving Average models. PLoS One. 2024;19(1):e0296064. pmid:38295029