Global burden of varicella in children aged 5–9 y, relationship with vaccination policy indicators, and projected trends to 2030 Authors: Zitai Zeng, Xiaosong Ma, Huadong Zhang, Dongmei Yuan, Yunjia Cao, Yang Liu, Jingfu Qiu, Chunbei Zhou Published: 2026 ABSTRACT ABSTRACTVaricella remains a common childhood infection worldwide. Understanding long-term trends and identifying key determinants of disease burden are essential for informing vaccination strategies and public health planning. Data were obtained from the Global Burden of Disease (GBD) 2021 study and the global varicella vaccination policy information. Extreme Gradient Boosting (XGBoost) model was trained to predict the disease burden for 2030. This model estimated the incidence, prevalence, and disability-adjusted life years (DALYs) for varicella among children aged 5–9 y worldwide. Shapley Additive exPlanations (SHAP) was employed to identify key factors and inform public health strategies. From 1990 to 2021, the rate of global incidence rose from 1,028.46/100,000 to 1,048.24/100,000, prevalence from 30.81/100,000 to 31.07/100,000, and DALYs declined from 21.09/100,000 to 12.36/100,000. By 2030, these metrics are projected to fall to 897.01/100,000, 28.96/100,000, and 11.71/100,000, respectively. From 1990 through 2030, Southeast and South Asia had the highest varicella incidence, Asian and Pacific Island nations had the highest prevalence, and sub-Saharan Africa and South Asia had the highest DALYs. SHAP identified population as the strongest predictor (mean absolute SHAP value: 0.167 for incidence and 0.103 for prevalence), and higher vaccination program tier may be associated with lower DALYs. Varicella burden among children aged 5–9 y has declined globally, but model projections indicate that regional disparities will persist. The lack of vaccination coverage and cohort-level immunization data limits interpretation of vaccine effects; future studies should incorporate both coverage and cohort exposure information to strengthen policy evaluation. Introduction Varicella, an acute and highly contagious disease caused by the varicella-zoster virus (VZV), is primarily transmitted through respiratory droplets and direct contact. While generally self-limiting in children, it poses a substantial global health burden. The World Health Organization (WHO) estimates that approximately 140 million cases occur annually worldwide, of which about 4.2 million lead to severe complications.1,2 Extensive research on the global and regional burden of varicella-zoster virus infection has highlighted persistent disparities linked to geography and age. Regional inequalities, shaped by demographic and epidemiological transitions, are reflected in differing disease burdens across countries and socio-demographic index (SDI) regions.3 Age-specific patterns also revealed pronounced variation: incidence and DALYs have continued to increase among specific population groups, with an estimated annual percentage change (EAPC) of around 0.13%.4–6 Notably, children in temperate regions, particularly those aged 5–9 y, experienced the highest incidence rates.7 These findings underscore the wide-ranging impact of varicella across regions and age groups. Therefore, clarifying long-term trends in varicella disease burden and building reliable predictive models are essential for developing targeted strategies to mitigate the disease’s impact on high-risk populations worldwide. Traditional statistical models, such as age-period-cohort (APC), Bayesian age-period-cohort (BAPC), and autoregressive integrated moving average (ARIMA) models, are widely used in disease forecasting. While APC/BAPC models decompose age, period, and cohort effects, and ARIMA captures temporal autocorrelation, they share critical limitations when applied to the dynamic epidemiology of varicella, a disease strongly influenced by non-linear interactions and exogenous interventions like vaccination programs. ARIMA models assume stationarity and respond poorly to structural breaks caused by policy changes.8 APC/BAPC models, though interpretable, struggle to integrate multi-source data or capture complex interactions without extensive manual feature engineering.9 Additionally, such traditional approaches often lack the flexibility and predictive accuracy of modern machine learning ensembles when handling spatial heterogeneity and complex, high-dimensional data.10 XGBoost is a robust prediction algorithm and is well-suited for handling complex, high-dimensional symptom data. SHAP addresses the “black box” nature of machine learning models by quantifying feature importance, thereby enhancing interpretability and transparency in decision-making.11 The complementary strengths of these methods provide a solid technical foundation for infectious disease modeling. For instance, in developing an automated monkeypox detection system, researchers applied XGBoost with SHAP to quantify core feature contributions, offering statistical validation of model performance.12 Moreover, SHAP has been widely validated in ensemble learning frameworks, where it clarifies how input features influence classification outcomes and reveals latent associations between predictors and disease endpoints, thereby strengthening model interpretability.13,14 This framework also provides a systematic evaluation of the relative importance of clinical symptoms, signs, and laboratory parameters, supporting clinicians with objective, quantitative evidence for diagnostic decision-making.15 Based on GBD 2021 study, we applied the XGBoost model combined with SHAP methodology to analyze global, regional, and national trends in varicella among children aged 5–9 y and to project future burden up to 2030. This study aims to generate evidence-based insights that can inform targeted public health strategies to reduce the burden of varicella worldwide. Methods Data sources We assessed the burden of varicella in children aged 5–9 y using incidence, prevalence, and DALYs as primary metrics. Data were obtained from the GBD 2021 Results Tool (https://vizhub.healthdata.org/gbd-results/.), encompassing estimates from 204 countries and territories for 1990–2021. The dataset was stratified by year, sex, and country/regions. To ensure cross-country comparability despite variations in healthcare infrastructure, we relied exclusively on the GBD’s harmonized final estimates, which are generated through a rigorous standardized modeling framework that corrects for differential case definitions, diagnostic practices, and reporting biases across regions and over time. Population estimates of 1990–2021 and projections of 2022–2030 were obtained from the Institute for Health Metrics and Evaluation (IHME) database.16 Since this study used an open-access, publicly available dataset from the Global Burden of Disease Study and involved the analysis of secondary, aggregated data that did not contain any personal or identifiable information, it did not require institutional ethical approval. Because no human participants or individual-level data were involved, informed consent was not required. Data on varicella vaccination policy were obtained from the review by Lee et al. and from the WHO Immunization Data,17,18 which provide global records on the introduction and national adoption of varicella vaccine. By integrating these sources, we constructed a four-level categorical variable (vaccination_program_tier) representing national varicella vaccination programs. Countries were classified as follows: Tier 0: No vaccination program (WHO records the varicella vaccine as “Not introduced” or “No data available,” and the review likewise reports no documented schedule). Tier 1: Limited vaccination (WHO indicates that vaccination is implemented only for high-risk groups, case of outbreaks, or certain regions). Tier 2: National one-dose routine immunization (WHO records nationwide introduction of varicella vaccine, or the review reports adoption of a one-dose schedule). Tier 3: National two-dose routine immunization (WHO records nationwide introduction of varicella vaccine, and the review reports adoption of a two-dose schedule). See Supplemental Appendix 1 for the complete list. Model construction Based on the GBD 2021 study, we examined the disease burden patterns of varicella-zoster virus infection among children aged 5–9 y from 1990 to 2021, and predicted incidence, prevalence, and DALY rates through 2030. We constructed two machine learning models including the Regularized Linear Regression19 and XGBoost.20 Both models utilized the following input features: gender, year, vaccination program tier, and the natural logarithm of the population aged 5–9 y, stratified by gender. Model performance was evaluated using rolling-origin validation. Three expanding windows were constructed, each starting in 1990 and extending forward: W1 trained on 1990–2007 and validated on 2008–2011; W2 trained on 1990–2011 and validated on 2012–2015; and W3 trained on 1990–2015 and validated on 2016–2019. The period of 2020–2021 was reserved as an independent test set to evaluate true out-of-sample forecasting performance, and the final comparison between XGBoost and Regularized Linear Regression was made on this set. The model with the lower root-mean-square error (RMSE) was selected as optimal. Regularized linear regression strategy The elastic net combines L1 and L2 penalties, which is suitable for handling potential collinearity among predictors and for variable selection. The general form of the model is:log(Outcome Ratey,s,c+1)∼β0+β1⋅y+β2⋅s+β3⋅log(pnumy,s,c)+β4⋅(Tiery,c)⋅ε Where OutcomeRatey,s,c represents the rate of incidence, prevalence, and DALY for year y, sex s, and country/region c; Tiery,c is the vaccination_program_tier for year y and country c; [log(pnum)]y,s,c indicates the population size of children aged 5–9 y for year y, sex s, and country/region c, stabilizing demographic effects. The optimal hyperparameters for the Regularized Linear Regression model were selected using rolling-origin validation. We conducted a grid search over two regularization hyperparameters: the mixing parameter α (which controls the balance between L1 and L2 penalties) and the regularization strength λ. Candidate values for α were 0, 0.1, 0.5, and 1 (α = 0 corresponds to ridge regression and α = 1 to lasso), and for λ were 0.001, 0.01, 0.1, 1, and 10. For each parameter combination, we computed the cross-validated RMSE and selected the combination that minimized the average RMSE. The final model was then refitted on the full training set using the selected hyperparameters (see Supplemental Appendix 2 for details on the Regularized Linear Regression Model). XGBoost modeling strategy The XGBoost model iteratively builds decision trees to improve prediction accuracy. Its objective function combines a loss term, which measures prediction error, and a regularization term that limits model complexity to prevent overfitting. Each new tree corrects residual errors from previous trees, while regularization penalizes overly complex structures by constraining leaf numbers and weight magnitudes. This balance ensures the model remains accurate, stable, and resistant to overfitting (see Supplemental Appendix 3 for more details on the XGBoost model). The rates of incidence, prevalence, and DALY were chosen as outcome variables. Incidence, prevalence, or DALY rates were log-transformed [log (Incidence, prevalence, or DALY rate +1)] to manage skewness and handle zero values. The general form of the model is:log(Outcome Ratey,s,c+1)∼y+s+(Tier)y,c+[log(pnum)]y,s,c Where Outcome Ratey,s,c represents the rate of incidence, prevalence, and DALY for year y, sex s, and country/region c; Tiery,c is the vaccination_program_tier for year y and country c; [log(pnum)]y,s,c indicates the population size of children aged 5–9 y for year y, sex s, and country/region c, stabilizing demographic effects. The XGBoost model was optimized through a grid search over key hyperparameters: nround (number of boosting rounds), eta (learning rate), and max_depth (maximum tree depth), with early stopping applied so that training halted when the validation RMSE showed no improvement for 50 consecutive boosting rounds. Following the grid search, the Pareto front method was employed to balance the average validation RMSE (across the three rolling windows) with the overfitting rate. The combination with the lowest weighted score was ultimately selected as the optimal hyperparameter set. Then, the final model was retrained on the full training data and evaluated on the independent test set. Model performance was quantified using test RMSE and the Pearson correlation coefficient between observed and predicted values (see Supplemental Appendix 4 for details). Predicting disease burden Based on model performance, the optimal model was applied using IHME population projections to estimate varicella burden for 2022–2030, assuming that each country’s vaccination-program tier remains unchanged after 2021, with no new vaccine introductions or dose-schedule modifications during the projection period. To generate projections with uncertainty intervals, we performed 500 bootstrap resamples and reported the 2.5th and 97.5th percentiles of the predictions as the 95% uncertainty interval (UI). SHAP analyses We used SHAP to interpret the trained model and quantify the contribution of each predictor (year, sex, log-population, and vaccination program tier) to the estimated burden. The mean absolute SHAP value was computed to rank feature importance globally. SHAP dependence plots were examined to illustrate the direction and magnitude of each feature’s effect on the predictions for incidence, prevalence, and DALYs. This analysis enhances the interpretability of the model and identifies key drivers of varicella burden (details in Supplemental Appendix 5). Sensitivity analyses To evaluate the robustness of projected varicella burden under alternative policy assumptions, we conducted sensitivity analyses across five SDI strata (High, High-middle, Middle, Low-middle, and Low SDI) and three policy scenarios (a baseline scenario, an expansion scenario, and a conservative scenario). In the baseline scenario, vaccination program tiers were held constant at their 2021 levels for all SDI groups throughout 2022–2030. In the expansion scenario, vaccination policies were assumed to strengthen during the projection period. Specifically, program tiers increased in 2024 and again in 2027, with the magnitude of improvement varying by SDI level. High and High-middle SDI countries advanced more rapidly, Middle SDI countries experienced a single tier increase in 2027, and Low-middle and Low SDI countries showed limited or no change. In the conservative scenario, vaccination policies were assumed to weaken in selected SDI groups. Middle SDI countries experienced a reduction in 2027, while Low-middle and Low SDI countries saw tier decreases in both 2024 and 2027. High and High-middle SDI countries maintained their 2021 tiers throughout the period. For each scenario combination, the disease burden was re-projected for 2022–2030 using the trained model, allowing us to quantify how alternative policy pathways influence future varicella incidence, prevalence, and DALYs. These analyses characterized the sensitivity of long-term burden estimates to assumptions about vaccination-program evolution. Results Model performance The optimal hyperparameters for models were selected using rolling-origin validation, and the models were then evaluated on the independent test set. For the Regularized Linear Regression model, the RMSE values were 404.69/100,000 for incidence, 8.51/100,000 for prevalence, and 11.38/100,000 for DALYs, with Pearson correlation coefficients between observed and predicted values were 0.2456, 0.2701 and 0.4395, respectively (P < 0.001; Figure 1). Similarly, for the XGBoost model, the RMSE values on the test set were 354.06/100,000 for incidence, 7.38/100,000 for prevalence, and 10.21/100,000 for DALYs. Pearson correlation coefficients for this model were 0.5253, 0.5399 and 0.5589, respectively (P < 0.001; Figure 2). Therefore, the XGBoost model was used to forecast the disease burden of varicella. Figure 1.RMSE and Pearson correlation coefficient between observed and predicted values in the Regularized Linear Regression model. Incidence (A), prevalence (B), and DALYs (C).Image A: Scatter plot with r=0.2456, P <0.001, RMSE=404.69/100k. X-axis: Observation Incidence Rate (0-2,000), Y-axis: Prediction Incidence Rate (0-1,500). Diagonal line from (0,0) to upper right. Points cluster at x=300-1,400, y=700-1,150; some at x=1,900-2,000, y=700-1,050. Image B: Scatter plot with r=0.2701, P <0.001, RMSE=8.51/100k. X-axis: Observation Prevalence Rate (0-50), Y-axis: Prediction Prevalence Rate (0-40). Diagonal line from (0,0) to upper right. Points cluster at x=18-42, y=26-36; some at x=48-52, y=26-32. Image C: Scatter plot with r=0.4395, P <0.001, RMSE=11.38/100k. X-axis: Observation DALY Rate (0-60), Y-axis: Prediction DALY Rate (0-25). Steep diagonal line from (0,0). Points mainly at x=0-10, y=2-12; broad band at y=9-11 across x=10-45; few at x=60, y=9-10.Three scatter plots comparing observed and predicted incidence, prevalence and DALY rates. Figure 2.RMSE and Pearson correlation coefficient between observed and predicted values in the XGBoost model. Incidence (A), prevalence (B), and DALYs (C).Image A: Scatter plot with r=0.5253, P <0.001, RMSE=354.06/100k. X-axis: Observation Incidence Rate (0-2000), Y-axis: Prediction Incidence Rate (0-1500). Diagonal reference line present. Dense points from x=300-1500, y=600-1300; vertical cluster near x=2000, y=650-1650. Image B: Scatter plot with r=0.5399, P <0.001, RMSE=7.38/100k. X-axis: Observation Prevalence Rate (0-50), Y-axis: Prediction Prevalence Rate (0-40). Diagonal reference line present. Points cluster x=18-42, y=22-40; vertical group near x=50, y=25-45. Image C: Scatter plot with r=0.5589, P <0.001, RMSE=10.21/100k. X-axis: Observation DALY Rate (0-60), Y-axis: Prediction DALY Rate (0-25). Steep diagonal reference line. Concentrated points x=0-5, y=0-6; scattered points up to x=60, y=5-17; several points near y=25, x=20-25.Three scatter plots comparing observed and predicted incidence rate, prevalence rate and DALY rate. RMSE and Pearson correlation coefficient between observed and predicted values in the Regularized Linear Regression model. Incidence (A), prevalence (B), and DALYs (C). RMSE and Pearson correlation coefficient between observed and predicted values in the XGBoost model. Incidence (A), prevalence (B), and DALYs (C). Global and SDI-level trends in varicella burden among children aged 5–9 y from 1990 to 2021 From 1990 to 2021, the global incidence and prevalence of varicella among children aged 5–9 y showed similar patterns. The global number of incident cases increased from 6.00 million in 1990 to 7.20 million in 2021 (a 20% increase), while prevalent cases rose from 0.18 million to 0.21 million (an 18% increase). Correspondingly, the incidence rate increased slightly from 1,028.47/100,000 (95% UI: 558.19–1,609.64) to 1,048.25/100,000 (95% UI: 562.87–1,639.90), while the prevalence rate increased slightly from 30.81/100,000 (95% UI: 20.33–42.31) to 31.07/100,000 (95% UI: 20.92–42.74). In contrast, the global DALY burden declined substantially, with the number of DALYs decreasing from 0.12 million in 1990 to 0.08 million in 2021 and the DALY rate falling from 21.09/100,000 (95% UI: 17.14–25.60) to 12.35/100,000 (95% UI: 10.27–14.89). The varicella burden among children aged 5–9 y displayed non-uniform patterns across SDI regions. Except for the Low-SDI region, which showed no clear peak, incidence and prevalence rates in most regions followed an inverted-U trajectory, rising through the 1990s, reaching their highest levels between the late 1990s and early 2000s, and subsequently declining toward 2021. This pattern was evident globally and in the High-middle, Middle, and Low-middle SDI regions. Specifically, the highest incidence rate occurred in 1995 in the Low-middle SDI region (1277.62/100,000; 95% UI: 783.48–1829.73). The High SDI region consistently had the highest prevalence, with two peaks in 2000 (38.44/100,000; 95% UI: 26.76–51.59) and 2015 (38.13/100,000; 95% UI: 24.63–53.28). For DALYs, all regions showed a downward trend throughout the study period, but the Low SDI region maintained the highest rates across all SDI strata and experienced the smallest reduction, declining by only 35% (Figure 3). Figure 3.Time trends of varicella across different SDI regions. Incidence (A), prevalence (B), and DALY rate (C).Three time-series graphs depict varicella burden globally and across five SDI regions: Global, High-middle, High, Low-middle, Low and Middle SDI. Each graph shares an x-axis from 1990 to 2021. Image A shows incidence rates per 100,000, ranging from 800 to 1300. Low-middle SDI starts highest near 1180 in 1990, peaks at 1278 in 1995, then declines to 1190 by 2021. High SDI rises from 1000 in 1990 to 1200 in 2000, then drops to 1080 by 2021. Other regions peak around 2000, with Low SDI consistently lowest at 750-780. Image B displays prevalence rates per 100,000, ranging from 24 to 40. High SDI peaks near 38 in 2000 and 2015, declining to 35 by 2021. Low SDI remains lowest at 24-26. Image C shows DALY rates per 100,000, ranging from 2 to 41, with all regions declining from 1990 to 2021. Low SDI starts highest at 41, ending at 27, while High SDI remains lowest, ending near 2.Graphs of varicella incidence, prevalence and DALY rates globally and in five SDI regions, 1990-2021. Time trends of varicella across different SDI regions. Incidence (A), prevalence (B), and DALY rate (C). Global varicella burden among children aged 5–9 y projections to 2030 Model-based forecasts suggested that the global disease burden of varicella will experience slight reductions by 2030 relative to 2021. Specifically, the global incident cases of Varicella were projected to reach 5.92 million by 2030 (a 17% decrease from 2021). prevalent cases were projected to reach 0.19 million by 2030 (a 10% decrease from 2021).The projection for 2030 indicated an incidence rate of 897.01/100,000 (95% UI: 683.17–1,170.62), a prevalence rate of 28.96/100,000 (95% UI: 22.31–37.58) and a DALY rate of 11.71/100,000 (95% UI: 8.63–15.82). Both the number and rate of DALYs are expected to continue their slight downward trend, with only modest fluctuations through 2030 (Figure 4). Figure 4.Global temporal trends in the disease burden of varicella from 1990 to 2030. (A) incidence rate; (B) incidence cases; (C) prevalence rate; (D) prevalence cases; (E) DALY rate; (F) DALYs.Image A: Incidence Rate (1/100,000) rises from 1,000 in 1990 to 1,080 in 2000, falls to 1,000 by 2020 and is forecasted to decline to 900 by 2030 with uncertainty. Image B: Incidence Cases (Thousands) increase from 6,000 in 1990 to 6,800 in 2000, dip to 6,500 by 2010, rise to 7,400 by 2020 and are expected to slightly decrease to 6,800 by 2030. Image C: Prevalence Rate (1/100,000) peaks at 33 in 2000, declines to 31 by 2020 and is forecasted to drop to 28 by 2030. Image D: Prevalence Cases (Thousands) rise from 180 in 1990 to 200 in 2000, dip to 190 by 2010, rise to 220 by 2020 and are predicted to decrease to 200 by 2030. Image E: DALY Rate (1/100,000) declines from 21 in 1990 to 12 by 2020, with forecasted stability around 12-13 by 2030. Image F: DALY Cases (Thousands) decrease from 125 in 1990 to 85 by 2020, with a forecasted slight decline to 80-85 by 2030.Six line graphs showing varicella incidence, prevalence and DALY trends from 1990 to 2030. Global temporal trends in the disease burden of varicella from 1990 to 2030. (A) incidence rate; (B) incidence cases; (C) prevalence rate; (D) prevalence cases; (E) DALY rate; (F) DALYs. National-level heterogeneity of varicella burden through 1990–2030 The varicella burden among children aged 5–9 y exhibited substantial geographic variation across different indicators. In 1990, high incidence rates were observed not only in Southeast Asia, but also in several Pacific island nations, and a similar pattern emerged for prevalence rates. In contrast, the highest DALY rates were overwhelmingly concentrated in Low-SDI settings. By 2021, these spatial patterns had remained largely stable. For incidence, the Philippines recorded the highest burden in both 1990 (1,962.56/100,000) and 2021 (1,966.07/100,000). For prevalence, Thailand, the Maldives, and Seychelles were consistently among the highest-burden countries. Similarly, DALY rates remained highest in Low-SDI settings, with the Central African Republic registering 53.90/100,000 in 1990 and 56.01/100,000 in 2021. By 2030, incidence and prevalence are forecast to remain elevated in High and High-middle SDI settings, particularly in more developed and island settings. The highest incidence rates are expected in Saint Vincent and the Grenadines (1,355.71/100,000), and the highest prevalence rates are projected in Maldives (38.95/100,000). In contrast, the highest DALY rates are projected in Nigeria (23.21/100,000), indicating that the health burden of varicella will persist in Low-SDI countries, particularly in sub-Saharan Africa and South Asia. This divergent pattern underscores the persistent inequality in varicella outcomes (Table 1, Supplemental Table 1 and Figure 5). Sensitivity analyses revealed that the projected burden estimates remained consistent with the baseline scenario under both alternative policy assumptions for the majority of countries, with relative changes in incidence, prevalence, and DALY rates falling within ±5% of baseline predictions (Supplemental Table 1, 2 and 3). Figure 5.Disease burden of varicella nationwide in 1990, 2021, and 2030. (A-C) incidence rates (A: 1990; B: 2021; C: 2030); prevalence rates (D-F) (D: 1990; E: 2021; F: 2030); DALY rates(G-I) (G: 1990; H: 2021; I: 2030).A) A world map showing varicella incidence rates in 1990. High rates over 1900 per 100,000 are in parts of Southeast Asia. B) Incidence rates in 2021 show similar patterns, with high rates in Southeast Asia. C) By 2030, incidence remains high in Southeast Asia, with some reduction. D) Prevalence rates in 1990 are highest in North America, over 50 per 100,000. E) In 2021, high prevalence persists in North America. F) By 2030, prevalence decreases slightly but remains high in North America. G) DALY rates in 1990 are highest in sub-Saharan Africa, over 15 per 100,000. H) In 2021, high DALY rates continue in sub-Saharan Africa. I) By 2030, DALY rates remain elevated in sub-Saharan Africa, indicating persistent health burdens. Each map uses a legend to categorize rates, showing regional disparities in varicella impact over time.Nine world maps show varicella incidence, prevalence and DALY rates for 1990, 2021 and 2030.Table 1.Top 5 countries or regions with highest incidence, prevalence and DALY rates of varicella in 1990, 2021 and 2030.Measure199020212030Incidence Rate(/100,000)Philippines(1962.56)Philippines(1966.07)Saint Vincent and the Grenadines(1355.71)Seychelles(1950.52)Thailand(1963.60)Seychelles(1353.50)Thailand(1950.44)Timor-Leste(1963.07)Guam(1337.60)Maldives(1947.87)Myanmar(1956.52)Tonga(1307.94)Mauritius(1945.74)Lao People’s Democratic Republic(1949.25)Barbados(1263.60)Prevalence Rate(/100,000)Taiwan (Province of China)(51.74)Thailand(50.77)Maldives(38.95)Seychelles(50.53)Timor-Leste(50.76)Montenegro(38.50)Thailand(50.52)Republic of Korea(50.67)Saint Vincent and the Grenadines(38.44)Maldives(50.47)Myanmar(50.64)Brunei Darussalam(38.41)Mauritius(50.43)Lao People’s Democratic Republic(50.50)Seychelles(38.36)DALY Rate(/100,000)Afghanistan(85.37)Central African Republic(56.01)Nigeria(23.21)Haiti(74.96)Eritrea(49.17)Pakistan(22.37)Rwanda(68.63)Somalia(43.57)Democratic Republic of the Congo(19.36)Guinea-Bissau(59.22)Haiti(42.02)Philippines(16.86)Ethiopia(57.33)Afghanistan(39.74)Mozambique(16.28)(For complete information on regional differences, please refer to Supplementary Table 1). Disease burden of varicella nationwide in 1990, 2021, and 2030. (A-C) incidence rates (A: 1990; B: 2021; C: 2030); prevalence rates (D-F) (D: 1990; E: 2021; F: 2030); DALY rates(G-I) (G: 1990; H: 2021; I: 2030). Table 1.Top 5 countries or regions with highest incidence, prevalence and DALY rates of varicella in 1990, 2021 and 2030.Measure199020212030Incidence Rate(/100,000)Philippines(1962.56)Philippines(1966.07)Saint Vincent and the Grenadines(1355.71)Seychelles(1950.52)Thailand(1963.60)Seychelles(1353.50)Thailand(1950.44)Timor-Leste(1963.07)Guam(1337.60)Maldives(1947.87)Myanmar(1956.52)Tonga(1307.94)Mauritius(1945.74)Lao People’s Democratic Republic(1949.25)Barbados(1263.60)Prevalence Rate(/100,000)Taiwan (Province of China)(51.74)Thailand(50.77)Maldives(38.95)Seychelles(50.53)Timor-Leste(50.76)Montenegro(38.50)Thailand(50.52)Republic of Korea(50.67)Saint Vincent and the Grenadines(38.44)Maldives(50.47)Myanmar(50.64)Brunei Darussalam(38.41)Mauritius(50.43)Lao People’s Democratic Republic(50.50)Seychelles(38.36)DALY Rate(/100,000)Afghanistan(85.37)Central African Republic(56.01)Nigeria(23.21)Haiti(74.96)Eritrea(49.17)Pakistan(22.37)Rwanda(68.63)Somalia(43.57)Democratic Republic of the Congo(19.36)Guinea-Bissau(59.22)Haiti(42.02)Philippines(16.86)Ethiopia(57.33)Afghanistan(39.74)Mozambique(16.28)(For complete information on regional differences, please refer to Supplementary Table 1). Top 5 countries or regions with highest incidence, prevalence and DALY rates of varicella in 1990, 2021 and 2030. (For complete information on regional differences, please refer to Supplementary Table 1). SHAP analysis of variables affecting varicella incidence, prevalence, and DALY rates SHAP analysis of the XGBoost predictive model revealed distinct patterns in how gender, population of children aged 5–9 y, vaccination_program_tier, and year influence incidence, prevalence, and DALYs (Figures 6–8). Figure 6.SHAP summary plot and correlation plot of features predicting varicella incidence rates by the XGBoost model. (A) summary plot. (B-E) the dependence plots showing contributions of vaccine program (B), sex (C), year (D), and logarithm (population) (E).Image A displays a SHAP summary beeswarm plot with SHAP Values from -1.0 to 0.5. Features listed are Log Population, Sex(Male), Year and Program tier, with mean absolute SHAP values of 0.167, 0.061, 0.019 and 0.012 respectively. A reference line is at 0.0 and a legend indicates high and low feature values. Image B shows a scatter plot of SHAP Value vs. Program tier, with tiers 0 to 3 and SHAP Values from -0.3 to 0.3, forming vertical bands around 0.0. Image C depicts SHAP Value vs. Sex, with Female values below 0.0 and Male values above, separated by a dashed line at 0.0. Image D illustrates SHAP Value vs. Year from 1990 to 2020, with a curve rising from negative to positive values. Image E presents SHAP Value vs. Log Population, with values from -0.8 to 0.4. The curve declines from positive near Log Population 4 to a minimum near 13, then rises toward positive by 17, with a dashed line at 0.0.A mixed plot showing SHAP summary and four SHAP dependence scatter plots for varicella incidence features. Figure 7.SHAP summary plot and correlation plot of features predicting varicella prevalence rates by the XGBoost model. (A) summary plot. (B-E) the dependence plots showing contributions of vaccine program (B), sex (C), year (D), and logarithm (population) (E).Image A′s SHAP beeswarm plot highlights four features: Log Population, Sex Male, Program tier and Year, with SHAP Values from -0.50 to 0.50. Log Population shows a mean SHAP value of 0.103, dense between 0 to 0.40. Sex Male clusters near 0, Program tier ranges from slightly below 0 to 0.30 and Year clusters around 0. Image B′s scatter plot for Program tier (0 to 3) shows SHAP Values from -0.1 to 0.3, with tier 0 near 0, tier 1 mostly negative, tier 2 mostly positive and tier 3 positive. Image C′s scatter plot for Sex (Female and Male) shows SHAP Values from -0.08 to 0.04, with Female points mostly below 0 and Male points mostly above 0. Image D′s scatter plot for Year (1990 to 2020) shows SHAP Values from -0.10 to 0.10, with a line slightly negative early and positive later. Image E′s scatter plot for Log Population (4 to 16) shows SHAP Values from -0.50 to 0.25, peaking near 4 to 6, dipping near 12 to 13, then rising toward 16 to 17.A mixed plot showing SHAP summary and four SHAP dependence scatter plots for varicella prevalence features. Figure 8.SHAP summary plot and correlation plot of features predicting varicella DALY rates by the XGBoost model. (A) summary plot. (B-E) the dependence plots showing contributions of vaccine program (B), sex (C), year (D), and logarithm (population) (E).Image A displays a SHAP summary beeswarm for four features: Log Population, Program tier, Year and Sex(Male). The x-axis shows SHAP Values from -1 to 1, with a vertical reference line at 0. Mean |SHAP| values are 0.419 for Log Population, 0.121 for Program tier, 0.083 for Year and 0.007 for Sex(Male). Image B is a scatter plot of SHAP Value by Program tier, with x-axis categories 0 to 3 and y-axis from -1.00 to 0.00. Points stack vertically at each tier. Image C shows SHAP Value by Sex, with x-axis categories Female and Male and y-axis from -0.10 to 0.05. Points stack around 0.00. Image D is a scatter plot of SHAP Value by Year, with a smooth line declining from 0.15 in 1990 to -0.10 in 2020. Image E shows SHAP Value by Log Population, with a smooth line starting near 0.5 at 4, dropping below 0.0 around 7-8, reaching -0.5 at 12, then rising above 0.0 after 13, approaching 0.5 by 16.A mixed plot showing SHAP summary and four SHAP dependence scatter plots for varicella DALY features. SHAP summary plot and correlation plot of features predicting varicella incidence rates by the XGBoost model. (A) summary plot. (B-E) the dependence plots showing contributions of vaccine program (B), sex (C), year (D), and logarithm (population) (E). SHAP summary plot and correlation plot of features predicting varicella prevalence rates by the XGBoost model. (A) summary plot. (B-E) the dependence plots showing contributions of vaccine program (B), sex (C), year (D), and logarithm (population) (E). SHAP summary plot and correlation plot of features predicting varicella DALY rates by the XGBoost model. (A) summary plot. (B-E) the dependence plots showing contributions of vaccine program (B), sex (C), year (D), and logarithm (population) (E). The patterns of influence of each factor on incidence and prevalence were relatively similar. “Logarithm of population” (log_pop) was the most important feature for both incidence and prevalence (average absolute SHAP value of 0.167 for incidence and 0.103 for prevalence). Its dependency plots showed a negative correlation until log_pop reached 14, after which it turned positive. “Gender” (s) was the second most important feature for incidence and prevalence (average absolute SHAP value of 0.061 for incidence and 0.017 for prevalence). Dependency plots for both metrics indicated that male children were assigned higher incidence and prevalence risk weights in the model. The dependency plot for the “Year” (y) feature showed an increasing trend in its impact on incidence, a stable effect on prevalence, and a gradually decreasing influence on DALYs. “Vaccination_program_tier”(Tier) was the second most important predictor for DALYs, with a mean absolute SHAP value of 0.121. The SHAP distribution indicates that higher Vaccination program tiers generally correspond to lower predicted DALY values within the model. Since vaccination coverage was not included in the analysis, the model’s output should be interpreted as showing that program tier serves as a contextual proxy correlated with lower DALY predictions. Discussion This study systematically assessed and projected the incidence, prevalence, and DALY rates of varicella among children aged 5–9 y globally, along with the temporal trends of these indicators, based on GBD 2021 data using a combined XGBoost and SHAP modeling approach. Overall varicella burden declined globally from 1990 to 2021, yet significant regional disparities persisted. Low-SDI countries maintained persistently high DALY rates, while High-SDI countries exhibited lower varicella burden. SHAP analysis indicated that the population size of children aged 5–9 y was the most influential predictor in the model, with gender and year also contributing to prediction patterns. “Vaccination_program_tier” appeared as an important variable associated with lower predicted DALY values. Model-based projections for 2030 suggest that higher levels may persist in certain nations, underscoring the need for region-specific surveillance and intervention strategies. The study findings indicated that males carried a higher risk weighting for incidence and prevalence in this model, which is consistent with multiple epidemiological studies. For instance, Al-Tawfiq et al. reported 19,577 varicella cases recorded in Eastern Saudi Arabia from 1994 to 2011, with 10,248 male cases, yielding a male-to-female ratio of approximately 1.15:1.21 Similarly, Jablonka et al. found a male prevalence as high as 77.9% in a serological assessment of a refugee cohort in Germany.22 The combined effects of exposure environment distribution, vaccine accessibility, and healthcare service availability may explain the higher male prevalence observed in several studies and populations. First, refugee, immigrant, and migrant populations typically have higher male proportions, and these groups often experience lower vaccination coverage and reduced access to healthcare, leading to a concentration of susceptible individuals among males and consequently higher clinical case rates. Second, behavioral and social structural factors may amplify disparities, and men are more frequently engaged in high-exposure occupations or reside in collective living environments (e.g., military barracks, training camps), and may exhibit delays or differences in seeking medical care and reporting, thereby influencing the distribution of detected and registered cases.23,24 To reduce the burden of varicella among male children, risk communication targeting boys should be conducted through communities and schools. This should emphasize the importance of childhood vaccination, promote household prevention measures, and clearly outline medical consultation and isolation procedures for suspected cases to minimize exposure and transmission risks. Regional disparities in the global burden of varicella remained substantial from 1990 to 2021. Lower-burden and higher-burden regions showed clear and persistent contrasts: some areas experienced consistently elevated prevalence, DALY rates, and mortality, whereas others maintained comparatively low levels or showed more rapid declines. Studies on other infectious diseases support this pattern. An epidemiology study has found that regions with lower SDI bore higher incidence and mortality rates for diseases such as cysticercosis.25 A global childhood tuberculosis study showed that Low-SDI regions consistently had the highest incidence and DALY rates, with a strong negative correlation between SDI and disease burden, indicating that poorer regions faced disproportionately higher transmission and worse outcomes.26 Similarly, the global rabies analyses reported that Low-SDI regions bore the highest rate of incidence, prevalence, and DALY, whereas High-SDI regions approached zero burden.27 Broader neurological disease assessments also showed that infectious neurological conditions, such as malaria, meningitis, and rabies, remain heavily concentrated in Low-SDI regions, while High-SDI regions have largely controlled these threats.28 Furthermore, research on neglected tropical diseases and malaria demonstrated that Low-SDI regions exhibited the highest incidence, mortality, and DALY rates.29 In summary, the geographic patterns of varicella burden in this study are broadly consistent with those of other major communicable diseases, highlighting persistent regional disparities. SHAP analysis indicated that the population size of children aged 5–9 y was the most influential predictor across all models. Vaccination program tier appeared as a contextual feature associated with lower predicted DALY values, with both “limited vaccination tier” and “national one/two-dose routine immunization tier” showing lower SHAP values than the “no vaccination program.” Previous observational studies have reported similar trends. Varela et al.’s review summarized global experiences with varicella vaccination programs.30 In Japan, pediatric varicella notifications declined after the introduction of a two-dose schedule, accompanied by milder breakthrough presentations.31 In New Zealand, implementation of a funded single-dose program coincided with reductions in varicella-related hospitalizations and complications.32 However, this pattern reflects an association learned by the model rather than a causal effect, as vaccination coverage was not included in the analysis. Overall, vaccination program characteristics and temporal patterns may collectively relate to the observed variations in varicella burden. The dependence plot of DALY rates showed that higher vaccination program tiers were associated with lower predicted DALY values in the model. Observational studies have reported similar epidemiological trends. From a broader public health perspective, the advantages of promoting varicella vaccination are well recognized. These include direct clinical benefits such as reductions in severe complications, hospitalizations, and mortality, as well as indirect population-level benefits like outbreak prevention and herd immunity when coverage is high. For example, surveillance data from Shanghai described declines in varicella incidence following the introduction of a second dose.33 A seroprevalence study in Guangzhou, China, demonstrated increasing antibody levels with age and vaccination history, supporting the recommendation for two doses of varicella vaccine in children.34 In the United States, declines in varicella-related mortality were observed during the early years of routine two-dose vaccination.35 Economic evaluations have further suggested that routine vaccination may be cost-effective in many high‑income and middle‑income countries, particularly when long-term savings from reduced healthcare use and caregiver burden are considered.36,37 However, program costs and infrastructure demands pose substantial feasibility barriers in low‑resource settings. Vaccine acceptance can be undermined by parental hesitancy, which directly threatens coverage and herd protection.38 Safety profiles of the varicella vaccine are generally favorable, with most adverse events being mild and localized, although rare serious events warrant continued pharmacovigilance.39 In addition, moderate coverage levels in some settings have raised concerns about a potential shift in the age distribution of cases toward older children and adults.40 These observations collectively highlight the need for maintaining high coverage, strengthening surveillance, and adapting vaccination strategies to local epidemiological contexts. Because our vaccination variable is a simplified vaccination policy indicator rather than achieved population coverage, the cited observational studies should be interpreted as describing trends observed in specific high-coverage settings, rather than providing causal evidence that can be validated within our modeling framework. Future research could incorporate longitudinal vaccination coverage data and dynamic transmission models to more rigorously evaluate how changes in immunization strategies may influence varicella burden across different demographic contexts. Beyond these general public health considerations, a specific observation from our SHAP analysis warrants further discussion. Notably, in the SHAP dependency plots for incidence and prevalence, countries implementing higher vaccination program tiers exhibit higher predicted values. This counterintuitive association likely reflects confounding between our vaccination policy variable and surveillance intensity: High-SDI countries, which predominantly adopt high-tier vaccination programs, typically possess more robust surveillance systems and sensitive diagnostic tools, enabling more thorough case identification and reporting. Consequently, they appear in the model with higher recorded incidence and prevalence rates, rather than the vaccine strategy itself increasing disease burden. Furthermore, it is important to reiterate that SHAP attribution values reflect association rather than causation, and thus cannot establish that a feature has a causal effect on the outcome.41 In summary, if more robust policy conclusions are needed, supplementary data on vaccination coverage and testing intensity are required. This study has several limitations. First, the lack of high-quality primary data may introduce uncertainty, as GBD estimates rely on heterogeneous sources and extrapolation in data-poor regions. To mitigate this, the GBD applies a rigorous standardized framework that enhances cross‑country comparability through systematic bias correction and quantifies uncertainty via 95% UIs to address underreporting and variations in surveillance intensity. Nevertheless, substantial differences in healthcare infrastructure, surveillance system sensitivity, and diagnostic practices across countries may still affect the consistency and accuracy of the final estimates. Second, our vaccination policy variable is a simplified categorical indicator of program adoption rather than a measure of achieved population coverage. The absence of coverage data and cohort‑level immunization records limits our ability to rigorously evaluate the causal impact of vaccination strategies, as we cannot distinguish policy implementation from actual vaccine uptake. Finally, SHAP reflects associations rather than causal effects. Future work should integrate primary surveillance data, include vaccination coverage, and apply quasi-experimental methods to better evaluate the impact of varicella vaccination policies. Conclusions This study assessed the global burden of varicella among children aged 5–9 y using GBD 2021 data and applied XGBoost and SHAP to evaluate trends from 1990 to 2021 and project outcomes to 2030. Globally, incidence and prevalence rates remained relatively stable over the study period, whereas DALY rates showed a sustained decline. Low-SDI countries experienced persistently high DALY rates, while incidence and prevalence peaks were more heterogeneous across settings, consistent with differences in health system capacity. Among the four features included in the model, population size of children aged 5–9 y showed the strongest SHAP importance across all outcomes. Higher vaccination program tiers were associated with lower SHAP values for DALY rate, suggesting a relationship between higher-tier policies and reduced predicted burden, although causality cannot be established given the absence of coverage data. Overall, our study demonstrates the value of combining XGBoost and SHAP for infectious disease prediction and interpretation. While these findings should be interpreted as associations rather than causal evidence, they underscore the importance of strengthening surveillance systems, expanding vaccine access in high-burden regions, and implementing targeted interventions for children aged 5–9 y to reduce varicella-related health loss. Supplementary Material Supplementary Table 3.docx Supplemental Appendix.docx Supplementary Table 1.docx Supplementary Table 2.docx