Temporal disruption in tuberculosis incidence patterns during COVID-19: a time series analysis in China

Data source and preparation

This study analyzed monthly PTB incidence data obtained from the National Health Commission of the People’s Republic of China (http://www.nhc.gov.cn/), covering the period from January 2007 to December 2023. The dataset included reports from 31 provinces, municipalities, and autonomous regions in China. Data quality was rigorously checked to ensure accuracy and consistency. Missing values were handled using linear interpolation, a suitable approach for time series data. Outliers were detected via Z-scores and adjusted using winsorizing to minimize their impact on model accuracy without distorting the overall trends. Population data for each year were used to calculate annual incidence rates per 100,000 individuals. The study design and methodology are illustrated in Fig. 1, which provides an overview of the time series models (SARIMA, Prophet, and LSTM) used to predict PTB incidence in China based on pre-COVID-19 data (2007–2019). Detailed monthly incidence and mortality data for pulmonary tuberculosis from 2007 to 2023 are available in Table S1.

Descriptive statistical analysis

A descriptive statistical analysis was performed to explore trends in PTB incidence over time. The annual incidence rate was calculated by dividing the total number of reported PTB cases each year by the population for that year and multiplying by 100,000 to express the rate per 100,000 people. Line charts were generated to visualize these trends over time, and the annual percentage change (APC) was calculated using the formula:

$$APC=\left({\displaystyle \frac{CurrentYearValue-PreviousYearValue}{PreviousYearValue}}\right)\times 100\mathrm{\%}$$

This measure enabled precise year-over-year comparisons of PTB incidence, with particular attention to fluctuations caused by the COVID-19 pandemic.

Model training and forecasting

Model evaluation

The accuracy of the models was assessed using the following metrics:

• Root mean square error (RMSE): Measures the average magnitude of the forecast errors.

• Mean absolute error (MAE): Provides the average absolute forecast errors.

• Mean absolute percentage error (MAPE): Expresses forecast accuracy as a percentage, with values between 0% and 10% indicating high accuracy (Lewis, 1982; Wang et al., 2019). The lower the MAPE, the better the model’s prediction performance.

Residual analysis, including the Ljung-Box test, was conducted to ensure the robustness of the models. These metrics provided a comprehensive evaluation of model performance across SARIMA, Prophet, and LSTM models.

Statistical analysis

All data analysis and modeling were conducted using R (version 4.2.2; R Core Team, 2022) for the SARIMA and Prophet models and Python (version 3.12 with TensorFlow/Keras) for the LSTM model. The analysis focused on evaluating model fit, forecast accuracy, and residual patterns to ensure reliability in predictions.

Ethics statement

This study utilized publicly available data from the National Health Commission of the People’s Republic of China, aggregated at a national level and without personally identifiable information. Therefore, ethical approval and informed consent were not required. Data handling followed ethical guidelines to maintain data integrity and confidentiality.

National trends in pulmonary tuberculosis incidence (2007–2023)

From January 2007 to December 2023, China reported a total of 19,840,536 PTB cases, with an average of 97,257.53 cases per month. Over this 17-year period, the annual incidence rate steadily declined from 88.55 per 100,000 in 2007 to an estimated 37.24 per 100,000 in 2023 (Fig. 2, Table 1). This consistent decline reflects significant improvements in PTB control measures, including enhanced diagnosis, treatment, and prevention strategies nationwide.

The most notable reduction occurred between 2020 and 2022, coinciding with the COVID-19 pandemic. Public health interventions, such as lockdowns, social distancing, and reduced mobility, likely contributed to a reduction in PTB transmission. In 2019, China reported 1,034,760 cases (55.55 per 100,000), which dropped to 876,576 cases (47.76 per 100,000) in 2020—a 15.25% reduction in case numbers and a 14.05% decrease in the incidence rate.

This downward trend continued through 2021 and 2022, with cases decreasing further to 712,586 and the incidence rate falling to 39.76 per 100,000. However, in 2023, PTB cases rebounded to 773,512, reflecting an 8.55% increase from 2022. Despite this rise in case numbers, the incidence rate continued to decline slightly to an estimated 37.24 per 100,000 (a 6.34% decrease), possibly due to the increasing population. This divergence, where the number of cases increases while the incidence rate decreases, suggests a shift in post-pandemic PTB dynamics that may warrant further investigation.

Pre-pandemic pattern and seasonality of pulmonary tuberculosis cases (2007–2018)

From January 2007 to December 2018, China reported a total of 15,636,118 PTB cases, with an average monthly incidence of 108,584.15 cases. During this period, the annual incidence rate steadily declined from 88.55 per 100,000 in 2007 to 59.27 per 100,000 by 2018 (Fig. 3A, Table 1). This decline reflects the sustained public health efforts, including enhanced PTB diagnostic techniques and improved access to treatment.

The analysis of the ACF and PACF for the original series (Fig. 4) revealed that the autocorrelation and partial autocorrelation values peaked when the lag was set at 12 months, exceeding the boundary limits. This suggests strong seasonality in the data, with a periodic cycle of 12 months, indicating that PTB incidence follows an annual pattern.

To further explore these trends, we applied LOESS smoothing to decompose the incidence data into trend, seasonal, and irregular components (Fig. 3B). The trend component confirmed a steady decline in cases over time, while the seasonal component revealed recurring peaks in March and December. These seasonal peaks were likely driven by factors such as increased human mobility during these periods and the heightened prevalence of respiratory infections, which could contribute to higher PTB transmission rates.

In addition, we plotted monthly and seasonal trends in Figs. 3C and 3D. The monthly subseries plot and seasonal plot demonstrated a similar downward trend across most months, with the exception of January and February, where no significant pattern was observed. Notably, January, February, and October consistently had the lowest number of cases each year, while the incidence rate was highest from March to May. Two annual peaks were identified: a more pronounced peak during the spring months (March to May) and a secondary peak in December. These findings strongly support the existence of seasonal characteristics in PTB transmission, with the spring season being the most critical period for high PTB incidence.

Model construction and performance analysis

We used monthly PTB case data from January 2007 to December 2018 to train three different models—SARIMA, Prophet, and LSTM—aimed at capturing the pre-COVID-19 pattern of PTB incidence in China. These models were chosen for their complementary strengths: SARIMA for handling seasonal trends, Prophet for its flexibility with irregular patterns and handling of holidays, and LSTM for modeling complex non-linear dependencies in time series data.

The models were trained on the 2007–2018 data, and subsequently, they were employed to forecast PTB incidence for the period from January to December 2019. The predicted values from each model were then compared with the actual PTB incidence data for 2019, allowing us to assess model fit and forecast accuracy. This comparison served as a cross-validation of the models’ ability to predict PTB incidence trends, providing insights into the forecasting performance of each model.

SARIMA model

Using monthly PTB incidence data from 2007 to 2018, we developed a seasonal autoregressive integrated moving average (SARIMA) model to account for both the declining trend and seasonal fluctuations in the data, with a seasonality period of 12 months (s = 12) (Fig. 3). The original time series was differenced once to account for both seasonal and non-seasonal components, producing a stationary series (Fig. 5). Stationarity was confirmed using the augmented Dickey-Fuller (ADF) test (ADF = −7.824, P = 0.01), indicating that the series was suitable for modeling.

ACF and PACF plots of the differenced series (Fig. 6) were used to guide the selection of SARIMA model parameters. Through this analysis, we determined that p = 0 and $q=1$ were optimal for the non-seasonal component, while seasonal $P$ and $Q$ were tested up to a maximum of 2. A total of five candidate models were evaluated based on parameter estimation and model diagnostics, with particular focus on the AIC and the Box-Ljung test for residual autocorrelation (Table 2).

Table 2:

SARIMA model candidate parameters and Box-Ljung test results.

SARIMA model	Model parameters							Ljung-Box (Q) test
	MA1	SAR1	SAR2	SMA1	SMA2	AIC	BIC	${\mathit{\chi }}^2$	P
(0,1,1)(0,1,1)12	−0.8731	–	–	−0.7260	–	2,717.070	2,725.980	3.411	0.065
(0,1,1)(0,1,2)12	−0.8726	–	–	−0.7314	0.0080	2,719.066	2,730.946	3.416	0.065
(0,1,1)(1,1,1)12	−0.8728	−0.0049	–	−0.7235	–	2,719.068	2,730.948	3.413	0.065
(0,1,1)(1,1,2)12	−0.8685	−0.6968	–	0.1319	−0.7107	2,714.129	2,728.978	2.022	0.155
(0,1,1)(2,1,2)12	−0.8453	−0.8222	−0.3853	0.1653	−0.4198	2,708.854	2,726.673	2.022	0.155

DOI: 10.7717/peerj.18573/table-2

The SARIMA (0,1,1)(2,1,2)12 model was selected as the best-performing model, with an AIC of 2,708.854 and a BIC of 2,726.673. The Box-Ljung test result ( ${\chi }^2$ = 2.022, P = 0.155) confirmed that the residuals of the model were white noise, indicating no significant autocorrelation and a good model fit. This model was subsequently used to forecast PTB incidence from January to December 2019 (Fig. 7 and Table S2).

Analysis of post-COVID-19 pulmonary tuberculosis patterns

Analysis of post-COVID-19 pulmonary tuberculosis patterns using time series models

To further explore the impact of COVID-19 on PTB incidence, we applied SARIMA, Prophet, and LSTM models to analyze both the pre- and post-COVID-19 periods. First, these models were trained using monthly PTB case data from January 2007 to December 2019, and performance was evaluated using RMSE, MAE, and MAPE. As shown in Fig. 8, Tables 5 and S3, all three models captured the trend in PTB incidence effectively, with RMSE values of 6,660.49 for SARIMA, 6,265.33 for Prophet, and 8,205.52 for LSTM. The MAPE values for SARIMA, Prophet, and LSTM were 4.712%, 4.293%, and 5.628%, respectively, indicating stable pre-pandemic performance.

Table 5:

Forecast performance of SARIMA, Prophet, and LSTM models for pulmonary tuberculosis (PTB) incidence during the COVID-19 pandemic (2020–2023).

		Model
		SARIMA	Prophet	LSTM
Fitting with raw data from 2007 to 2019
	RMSE	6,660.490	6,265.326	8,205.523
	MAE	4,773.484	4,364.737	5,640.628
	MAPE(%)	4.712	4.293	5.628
Forecasts from 2020 to 2023
	RMSE	8,931.543	11,366.453	8,773.820
	MAE	6,866.618	13,117.370	5,927.995
	MAPE(%)	11.477	19.032	10.365
Forecasts from 2020 to 2022
	RMSE	9,593.856	14,068.800	8,971.01
	MAE	7,528.183	12,019.408	5,676.999
	MAPE(%)	12.808	20.402	10.520
Forecasts from Jan. to Dec. 2023
	RMSE	6,554.660	9,719.462	8,153.698
	MAE	4,881.923	9,407.580	6,680.980
	MAPE(%)	7.481	14.923	9.899

DOI: 10.7717/peerj.18573/table-5

Next, the models were used to forecast PTB incidence from January 2020 to December 2023, and the predicted values were compared with the actual data. As shown in Fig. 8, Tables 5 and S3, there were significant deviations between the predicted and actual values during the forecast period. The RMSE, MAE, and MAPE values for the 2020–2023 period were notably higher than those for the pre-pandemic period, indicating a decline in predictive accuracy during COVID-19. This suggests that the pandemic introduced factors not accounted for by the models, such as disruptions in healthcare access and behavioral changes due to COVID-19 containment measures.

To better understand the deviations in PTB incidence patterns before and after the COVID-19 outbreak, we divided the forecast period into two stages: Stage 1 (January 2020 to December 2022) and Stage 2 (January 2023 to December 2023). During Stage 1, all three models exhibited higher RMSE and MAPE values, reflecting greater discrepancies between predicted and actual values during the peak of the pandemic. In contrast, Stage 2 saw a reduction in RMSE and MAPE values, approaching pre-pandemic levels (Tables 3 and 5), suggesting that the predictive performance of the models improved as PTB incidence began to return to pre-pandemic patterns.

Furthermore, we calculated the error values between the predicted and actual PTB incidence for each model across both stages (Fig. 9 and Table S4). Notably, larger fluctuations in error values were observed in Stage 1, particularly in February 2020 and December 2022, where all three models—SARIMA, Prophet, and LSTM—showed peak discrepancies between predicted and actual values.

These findings indicate that while PTB incidence patterns were significantly disrupted during the COVID-19 pandemic, they began to stabilize in 2023. The improved performance of the models in Stage 2 suggests that the effects of COVID-19 on PTB transmission may be temporary, and PTB patterns are gradually returning to pre-pandemic norms.

Impact of COVID-19 on tuberculosis incidence in China

TB remains a significant public health challenge in China, which is one of the 30 high TB burden countries globally (World Health Organization (WHO), 2023). Despite ongoing efforts, the COVID-19 pandemic introduced new variables into the TB control landscape, disrupting diagnostic and treatment pathways and altering transmission dynamics (Zhou et al., 2023). Public health measures implemented to contain COVID-19, such as lockdowns, social distancing, and enhanced hygiene practices, significantly impacted the transmission of TB and other infectious diseases (Namgung et al., 2023; Song et al., 2022; Tian et al., 2020). A study from Shantou, China, reported that these measures led to a substantial reduction in TB incidence, especially among older adults and certain occupations like agriculture and the unemployed, underscoring the indirect effects of COVID-19 interventions on TB transmission (Su et al., 2024).

This reduction in TB cases during the pandemic is consistent with other studies that have explored the broader impact of COVID-19 on infectious diseases (Li et al., 2023b; Nash et al., 2022). A national-level study in China found that both TB incidence and mortality decreased significantly at the onset of the pandemic (Zhang et al., 2023). However, it remains unclear whether this reduction reflects a true decline in TB transmission or whether it is partly due to delays in diagnosis and interruptions in healthcare services. For example, some findings suggest that healthcare access was restricted during the pandemic, reducing opportunities for TB diagnosis and treatment (Morrison et al., 2023).

As COVID-19 control measures relaxed in 2023, TB incidence began to show signs of recovery, yet the overall incidence rate remained below pre-pandemic levels. This highlights the need for further research into the long-term effects of the pandemic on TB transmission, including how changes in healthcare-seeking behavior and service delivery may have influenced TB trends (Zhou et al., 2023). Studies using time series analysis, such as the one from Guizhou, China, which observed prolonged declines in TB incidence even after the initial COVID-19 wave, suggest that some of the pandemic’s effects on TB transmission may be long-lasting (Zhou et al., 2023).

Pre-COVID-19 PTB patterns and seasonality

Before the COVID-19 outbreak, the incidence of PTB in China exhibited a distinct seasonal pattern, with two notable annual peaks: one in March and a smaller secondary peak in December. This pattern is likely linked to increased human mobility following the “Spring Festival” and a rise in respiratory infections during this period, both of which may facilitate TB transmission. Similar seasonal patterns have been observed in other respiratory diseases, further validating this finding (Wang et al., 2020).

Our results align with previous studies, such as those by Li et al. (2013) and Luo et al. (2014), which emphasized the “Spring Festival Effect” as a key factor influencing TB incidence during these months. The spring peak is positively correlated with rising temperatures, which may enhance the survival of Mycobacterium tuberculosis in aerosols (Wang et al., 2020). This effect is particularly pronounced in areas with cooler spring climates, where the post-winter surge in TB transmission is more evident (Luo et al., 2014).

Conversely, the summer months, particularly June and July, saw a noticeable decline in TB cases. This trend could be attributed to the inhibitory effects of high temperatures on bacterial survival, especially in regions like Xinjiang and Tibet, where temperatures often exceed 37 °C (Wang et al., 2021). Additionally, environmental factors such as air pollutants and varying levels of humidity may further influence the seasonal transmission of TB, as observed in the study by Wang et al. (2021), which found that PM10 concentrations had a significant effect on TB incidence in Shijiazhuang, China.

These findings underscore the importance of understanding the climatic and cultural factors that contribute to the seasonality of TB in China, particularly in the context of public health planning and targeted interventions during high-risk periods.

Post-COVID-19 PTB patterns and disruptions

Initially, we used data from 2007 to 2018 to construct SARIMA, Prophet, and LSTM models to predict the incidence of PTB in 2019. Models with a MAPE between 0% and 10% are considered highly accurate in capturing data trends. The MAPE for the three models ranged from 3% to 6% for fitting and from 6% to 9% for predictions (excluding December 2019, Table 3), indicating strong predictive performance before the onset of COVID-19.

Next, we applied these models (trained on pre-pandemic data) to forecast PTB incidence from January 2020 to December 2023. We observed significant deviations between predicted and actual values, with actual TB incidence markedly lower than predicted, consistent with findings from studies such as Ge et al. (2024), who observed delays in TB diagnosis during the pandemic in Eastern China. This pattern mirrors the results seen in other infectious diseases like hepatitis E and pertussis, where COVID-19 countermeasures significantly reduced transmission (Qin et al., 2024). The time series models revealed substantial deviations, demonstrating the collateral impact of public health interventions for COVID-19 on TB transmission.

Two key policy shifts in China’s COVID-19 response significantly influenced these deviations. First, in January 2020, the Chinese central government implemented emergency control measures, launching a Level-1 public health emergency response across more than 30 provinces. Wuhan was locked down, effectively curbing population movement (The State Council Information Office of the People’s Republic of China, 2020). Second, in December 2022, the National Health Commission downgraded COVID-19 from Category A to Category B, placing it on par with TB in terms of prevention and control, signaling the post-pandemic transition (National Health Commission of the People’s Republic of China, 2022). Our analysis shows that the prediction deviations were largest during the strictest COVID-19 measures (2020–2022), particularly during Wuhan’s lockdown and the policy shift in December 2022. These large deviations can be attributed to disruptions in TB diagnosis and reporting, delayed healthcare access, and reduced TB service capacity, as also highlighted by Zhang et al. (2022) in their study on TB services in Tianjin. Xie et al. (2022) similarly noted that diagnostic delays during the pandemic led to missed or delayed TB diagnoses.

Interestingly, our models indicated smaller prediction errors during May 2020 to November 2022, reflecting a reduction in the disruptive effects of the pandemic on TB transmission as public health measures became more normalized. The error peaks in February 2020 and December 2022 correspond with the early lockdown phase and the major policy transition, highlighting periods where TB control efforts were most disrupted.

Post-2023, as COVID-19 management moved to Category B and restrictions were relaxed, TB cases rebounded, surpassing model predictions. The relaxation of strict COVID-19 measures, the restoration of routine healthcare services, and the identification of delayed TB cases likely contributed to this resurgence. Similar findings were reported by Li et al. (2023a), who noted a rebound in TB incidence as healthcare services normalized after the pandemic. Our analysis confirms that the temporary impact of COVID-19 on TB transmission has begun to reverse, with PTB incidence gradually returning to pre-pandemic levels.

Overall, these findings suggest that the effects of COVID-19 on TB transmission in China may have been temporary, with the TB incidence pattern stabilizing as healthcare services resumed normal operations. As highlighted by Wang et al. (2021), ongoing vigilance is necessary to ensure that future disruptions to TB services are minimized.

Limitations and future directions

While this study provides valuable insights into the impact of COVID-19 on TB trends in China, several limitations should be acknowledged. First, the primary limitation is that we only used case numbers and incidence rates as quantitative measures of COVID-19’s impact on TB. This approach did not directly account for additional sociological factors, such as healthcare accessibility, public health system disruptions, and changes in population behavior during the pandemic. Future studies could benefit from incorporating a broader set of indicators, including health service availability and access, diagnostic delays, and population movement, using multivariate time series models to provide a more comprehensive understanding of these dynamics.

Second, our analysis focused on the national level, which may obscure regional disparities. Areas with high TB burdens, such as Xinjiang and Tibet, may have experienced the pandemic’s effects on TB transmission differently. Future research should aim to investigate these localized impacts, allowing for more tailored public health interventions. As noted in the studies by Zhang et al. (2023) and Fei et al. (2020), the pandemic may have exacerbated regional inequalities in healthcare access, further affecting TB outcomes.

Third, as with the studies by Zhang et al. (2023) and Fei et al. (2020), the study is limited by potential underreporting and delays in diagnosis during the pandemic. The reduction in reported TB cases during the pandemic may reflect not only reduced transmission but also disruptions in TB services, such as delayed or missed diagnoses, as patients may have avoided healthcare facilities due to COVID-19-related fears. Thus, this study cannot definitively distinguish between a true decrease in TB transmission and a decline in diagnosis and reporting.

Regarding the time series models, the evaluation of models pre-COVID-19 (with the 2007–2018 data as the training set and 2019 as the validation set) indicated that the Prophet model had the lowest MAPE, demonstrating the best overall performance for fitting and forecasting. While both SARIMA and LSTM models performed well in capturing TB trends in China, the SARIMA model required a more complex modeling process and parameter selection. The LSTM model, though advantageous in capturing long-term dependencies, was prone to overfitting, especially on smaller datasets, limiting its generalization capacity. Additionally, LSTM models may encounter irregularities, such as sudden interruptions in data patterns caused by external factors like the COVID-19 pandemic, due to their heavy reliance on historical data for predicting future trends (Chen et al., 2024). In contrast, the Prophet model proved more efficient by automatically capturing seasonality and trend parameters and offering superior performance with a simpler modeling process.

The optimization of LSTM models has been a key factor in improving prediction accuracy for various infectious diseases, including COVID-19 and TB. For instance, Sembiring, Wahyuni & Sediyono (2024) reported significant improvements in model performance when adjustments were made to the output gates and hidden states. Our study similarly benefited from fine-tuning LSTM parameters, which allowed it to model non-linear dependencies more effectively. This is consistent with findings by Wahyuni et al. (2022), who demonstrated that LSTM models generally outperform SARIMA and Prophet when larger datasets are available. Despite its complexity, the LSTM model in our study provided robust results, validating the use of advanced neural networks for time series forecasting. However, given Prophet’s performance in this study, future research could explore its application for predicting other infectious diseases in China.

As TB control strategies continue to evolve in the post-pandemic era, time series models will remain crucial in forecasting disease trends and guiding public health interventions. This study also highlights the importance of developing adaptive healthcare systems that can swiftly respond to external shocks, such as pandemics, ensuring minimal disruption to critical disease control efforts.