Eating disorders, particularly anorexia nervosa and bulimia nervosa are among the most debilitating mental health conditions globally. Reference Treasure, Duarte and Schmidt1,Reference Chew and Temples2 According to the DSM-5, eating disorders represent a heterogeneous group of psychiatric conditions, each characterised by disturbed eating behaviours, significant weight fluctuations and a pervasive preoccupation with body shape and weight. 3,Reference Silén and Keski-Rahkonen4 Anorexia nervosa is characterised by persistent restriction of energy intake leading to significantly low body weight, an intense fear of gaining weight and a distorted body image, while bulimia nervosa involves recurrent episodes of binge eating followed by compensatory behaviours such as vomiting, fasting or excessive exercise. Eating disorders impair both physical and psychological health, contributing significantly to global morbidity and mortality rates, with individuals demonstrating markedly heightened mortality rates, especially among those with anorexia nervosa. Reference Krug, Liu, Portingale, Croce, Dar and Obleada5 While this study focuses primarily on anorexia nervosa and bulimia nervosa due to their well-established diagnostic criteria and extensive research base, it is important to acknowledge that other specified feeding or eating disorders and binge-eating disorder also contribute significantly to the global burden of eating disorders. Eating disorders impair both physical and psychological health, contributing significantly to global morbidity and mortality rates, with individuals demonstrating markedly heightened mortality rates, especially among those with anorexia nervosa. 3,Reference Hornberger and Lane6
The global burden of eating disorders has grown substantially over recent decades, with a notable rise in incidence, prevalence and disability-adjusted life years (DALYs) attributable to anorexia nervosa and bulimia nervosa. Approximately 6 to 8% of adolescents develop an eating disorder during their teenage years. Reference López-Gil, García-Hermoso, Smith, Firth, Trott and Mesas7 The estimated prevalence of eating disorders among young adults was about 4% for males and ranges from 11 to 17% for females. Reference Galmiche, Déchelotte, Lambert and Tavolacci8 Young adults aged 15–29 years represent a particularly vulnerable demographic, as this developmental period encompasses critical transitions from adolescence to adulthood, characterised by identity formation, peer influence and increased autonomy in health behaviours. Reference Nagata, Stuart, Hur, Panchal, Low and Chaphekar9 During adolescence, individuals face numerous developmental challenges, including societal pressures, body image concerns and the development of their own identities. These factors, combined with genetic predispositions and psychological vulnerabilities, can contribute to the emergence of eating disorders. Reference Keski-Rahkonen10 Adolescents and young adults with eating disorders often experience heightened levels of anxiety, depression and low self-esteem, which can further exacerbate their condition and hinder their ability to engage in normal social and academic activities. Reference Johnson-Munguia, Negi, Chen, Thomeczek and Forbush11
During the COVID-19 pandemic, eating disorders saw a significant increase, which could be attributed to heightened psychological stress, social isolation, disrupted routines, limited access to resources and the influence of media and information. Reference Sonne, Kildegaard, Strandberg-Larsen, Rasmussen, Wesselhoeft and Bliddal12,Reference Bevilacqua, Fox-Smith, Lewins, Jetha, Sideri and Barton13 These conditions lead to malnutrition, growth stunts and other serious medical complications, while also having profound psychological consequences. 14 Eating disorders are linked to impaired quality of life, increased healthcare costs and high mortality rates, which often become evident during adolescence and young adulthood. Reference Austin, De Silva, Ilesanmi, Likitabhorn, Miller and Sousa Fialho15
However, a detailed, data-driven analysis of evolving trends across age groups, regions and sociodemographic levels remains lacking. Reference Baker, Freestone, Cai, Silverstein, Urban and Steinberg16,Reference Davidson, Barry, Mangione, Cabana, Chelmow and Coker17 Previous studies assessing the burden of eating disorders have predominantly relied on single-model forecasting approaches, which may be inadequate to capture the complex and dynamic nature of these disorders across diverse populations and regions. To address these gaps, this study employs eight advanced machine-learning models – Facebook Prophet Time Series Forecasting (Prophet), Autoregressive Integrated Moving Average (ARIMA), Trigonometric Seasonality, Box–Cox transformation ARMA Errors Trend and Seasonal Components (TBATS), Elastic Net, Error Trend Seasonality (ETS), Vector Autoregression (VAR), Holt–Winters Exponential Smoothing and the Theta Method – to analyse historical trends (1990–2021) and project future patterns (2022–2050) of eating disorders among young adults aged 15–29 years. Each model has distinct strengths, such as the ability to account for seasonality (TBATS, ETS), handle external regressors (ARIMA, VAR) or integrate regularisation techniques for improved forecasting (Elastic Net). By comparing the performance of these models across multiple metrics, we aim to generate robust and reliable projections, providing valuable insights for public health planning, early intervention and prevention strategies.
This study represents the first comprehensive analysis of eating disorders using multiple machine-learning models to forecast their global burden across incidence, prevalence and DALYs. By leveraging global burden of disease (GBD) 2021 data and integrating population projections, this analysis provides a granular understanding of age-specific and region-specific trends, offering a deeper understanding of how these disorders may evolve by 2050. The findings not only highlight the growing global challenge posed by eating disorders but also underscore the importance of adopting innovative approaches to inform health policy and intervention strategies.
Method
Data sources and study design
This study utilised data from the GBD 2021 database (https://ghdx.healthdata.org/gbd-2021). We extracted data on incidence, prevalence and DALYs for anorexia nervosa and bulimia nervosa among young adults aged 15–29 years from 1990 to 2021. Both absolute numbers and rates per 100 000 population were collected. According to the GBD 2021 classification, anorexia nervosa (ICD-10: F50.0-F50.1) and bulimia nervosa (ICD-10: F50.2-F50.5) are categorised under non-communicable diseases, mental disorders and eating disorders. 18–20
Data preparation and feature engineering
We integrated age-standardised rates from the GBD database (1990–2021) 18 with population projections from 1990 to 2050. 21,Reference Vollset, Goren, Yuan, Cao, Smith and Hsiao22
To enhance model predictive capacity, we performed comprehensive feature engineering: (a) Standardisation: all numerical variables (year and population data) were standardised using z-score normalisation to ensure comparable scales across features; (b) Lag Features: we created lag-1 and lag-2 features representing eating disorder rates from the previous year and two years prior to capture temporal dependencies; (c) Moving Averages: we calculated 3-year moving averages to smooth short-term fluctuations and capture underlying trends and (d) Interaction Terms: we generated interaction terms between standardised population and year data (pop_year interaction) to capture the combined effects of population growth and temporal trends. No additional missing data imputation was performed in our study, as the data-set was complete for our analytical requirements.
Model development
We employed eight advanced time-series forecasting and machine-learning algorithms: Prophet, ARIMA, TBATS, ETS, VAR, Holt–Winters Exponential Smoothing and Theta Method models. Each model was implemented with specific parameters and constraints to optimise performance and address the unique characteristics of the data.
The Prophet model was implemented with logistic growth and 15 change-points. The ARIMA model utilised auto arima for parameter selection, with external regressors incorporated. The TBATS model applied a simplified BATS approach without Box–Cox transformation. The Elastic Net model combined L1 (Lasso) and L2 (Ridge) regularisation penalties with a mixing parameter (α) of 0.3 to balance L1 and L2 regularisation, where α = 0 corresponds to pure Ridge regression, α = 1 corresponds to pure Lasso regression and our chosen α = 0.3 provides a balanced approach that leverages both the feature selection capabilities of Lasso and the multicollinearity handling strengths of Ridge regression. The VAR model employed a first-order lag structure with the population as an exogenous variable. The Holt–Winters Exponential Smoothing model incorporated an additive seasonal component, while the Theta Method model added a cyclical component with constrained growth rates. Detailed algorithm specifications, parameters and comparative analysis are provided in Supplementary Table S11 available at https://doi.org/10.1192/bjp.2025.10450.
Model validation and performance assessment
We implemented a rigorous time-series validation framework to ensure reliable performance estimates and prevent overfitting: (a) training period: 1990–2010 (historical data for model fitting and hyperparameter optimisation), (b) validation period: 2011–2021 (out-of-sample testing for unbiased performance evaluation) and (c) prediction period: 2022–2050 (future projections). Within the training period, we employed ten-fold time-series cross-validation for hyperparameter tuning and model stability assessment. All reported evaluation metrics represent out-of-sample performance on the validation set, ensuring genuine predictive accuracy assessment. Model performance was assessed using multiple evaluation metrics calculated on the validation set (2011–2021): mean squared error (MSE), mean absolute percentage error (MAPE), root mean squared error (RMSE), symmetric mean absolute percentage error (SMAPE), coefficient of determination (R 2) and mean absolute scaled error (MASE). These metrics provide comprehensive assessment of forecast accuracy, bias and relative performance.
Statistical analysis
Age-standardised rates were calculated using the direct standardisation method with the GBD 2021 global age-standard population as the reference. The direct method applies age-specific rates from our study populations to the standard population age structure, producing rates that are directly comparable across different countries, regions and time periods by removing the confounding effect of different age distributions. Uncertainty intervals were generated using 1000 draws from the posterior distribution of each estimate, with the 25th and 975th ordered draws determining the 95% uncertainty intervals. Multiple evaluation metrics (MSE, MAPE, RMSE, SMAPE, R 2 and MASE) were employed to provide comprehensive assessment of forecast accuracy: MSE and RMSE provide absolute error measures sensitive to outliers and scale, while MAPE and SMAPE offer percentage-based measures that are scale-independent; MASE provides baseline comparison against naive forecasting methods, while R 2 indicates the proportion of variance explained. This comprehensive approach addresses scale sensitivity, error distribution characteristics and ensures that model performance assessment is not biased towards any particular type of error pattern. A two-sided p-value <0.05 was considered statistically significant. All statistical analyses were performed using R version 4.3.0 for Windows (R Foundation for Statistical Computing, Vienna, Austria; https://www.r-project.org) with specific packages: prophet (Facebook Prophet model), forecast (ARIMA, ETS, Theta and Holt–Winters models), bats (TBATS model), glmnet (Elastic Net regularisation), vars (Vector Autoregression), tidyverse (data manipulation and processing), zoo (moving averages and time-series operations) and ggplot2 (data visualisation). To ensure full reproducibility of our results, all analytical code, model specifications and implementation details are publicly available at https://github.com/Paperaceepted/BJPsych.
Results
Global burden of eating disorders among young adults: a concerning trend
The global burden of eating disorders, specifically anorexia nervosa and bulimia nervosa, among adolescents and young adults aged 15–29 years demonstrated substantial increases from 1990 to 2021 (Figs 1, 2 and 3). Bulimia nervosa exhibited more pronounced growth compared to anorexia nervosa across all epidemiological metrics examined. The incidence of bulimia nervosa increased by 44.68%, with rates rising from 298.24 to 351.29 per 100 000 population, representing a 17.79% increase in age-standardised rates (ASRs) and an estimated annual percentage change (EAPC) of 0.56 (95% uncertainty intervals: 0.53 to 0.58) (Table 1). Similarly, anorexia nervosa incidence grew by 38.12%, with rates increasing from 38.52 to 43.31 per 100 000, corresponding to a 12.44% rise in ASRs and an EAPC of 0.44 (95% uncertainty intervals: 0.41 to 0.46).
Global and regional trends in incidence, prevalence and disability-adjusted life years (DALYs) of eating disorders among young adults aged 15–29 years, 1990–2021. (a) Incidence of anorexia nervosa and bulimia nervosa globally and by sociodemographic index (SDI) quintile, 1990–2021. The left y-axis shows absolute numbers; the right y-axis shows age-standardised rates per 100 000 population. Blue bars represent bulimia nervosa; red bars represent anorexia nervosa. (b) Prevalence of anorexia nervosa and bulimia nervosa globally and by SDI quintile, 1990–2021. Axes and colour coding as in panel (a). (c) DALYs of anorexia nervosa and bulimia nervosa globally and by SDI quintile, 1990–2021. Axes and colour coding as in panel (a).
Distribution and ranking of eating disorders burden among young adults aged 15–29 years by sociodemographic index (SDI) quintile, 1990 and 2021. (a) Percentage distribution of incident cases of anorexia nervosa and bulimia nervosa by SDI quintile in 1990 and 2021. (b) Ranking of age-standardised incidence rates of anorexia nervosa and bulimia nervosa by SDI quintile in 1990 and 2021. (c) Percentage distribution of prevalent cases of anorexia nervosa and bulimia nervosa by SDI quintile in 1990 and 2021. Colour coding as in panel (a). (d) Ranking of age-standardised prevalence rates of anorexia nervosa and bulimia nervosa by SDI quintile in 1990 and 2021. (e) Percentage distribution of disability-adjusted life years (DALYs) for anorexia nervosa and bulimia nervosa by SDI quintile in 1990 and 2021. Colour coding as in panel (a). (f) Ranking of age-standardised DALY rates for anorexia nervosa and bulimia nervosa by SDI quintile in 1990 and 2021. ASR, age-standardised rates.
Percentage change and estimated annual percentage change (EAPC) in eating disorders burden among young adults aged 15–29 years globally and by sociodemographic index (SDI) quintile, 1990–2021. (a) Incidence: percentage change in number of cases, age-standardised rates (ASRs) and EAPC in ASRs for anorexia nervosa and bulimia nervosa. (b) Prevalence: percentage change in number of cases, ASRs and EAPC in ASRs for anorexia nervosa and bulimia nervosa. (c) Disability-adjusted life years (DALYs): percentage change in number of DALYs, ASRs and EAPC in ASRs for anorexia nervosa and bulimia nervosa.
Global incidence of anorexia nervosa and bulimia nervosa among adolescents and young adults aged 15–29 years in 1990 and 2021, with trends from 1990 to 2021
EAPC, estimated annual percentage change; ASRs, age-standardised rates; SDI, sociodemographic index.
Prevalence data mirrored these concerning trends, with bulimia nervosa demonstrating a 53.18% increase in total cases and a 22.41% rise in ASRs (EAPC: 0.72, 95% uncertainty intervals: 0.69 to 0.75), while anorexia nervosa prevalence increased by 32.56% with a 6.92% rise in ASRs (EAPC: 0.28, 95% uncertainty intervals: 0.26 to 0.31) (Supplementary Table S1). The disease burden measured in DALYs also escalated significantly, with bulimia nervosa DALYs increasing by 53.12% and ASRs by 22.39% (EAPC: 0.72, 95% uncertainty intervals: 0.69 to 0.76), compared with anorexia nervosa’s 32.29% increase in DALYs and 6.72% rise in ASRs (EAPC: 0.28, 95% uncertainty intervals: 0.26 to 0.3) (Supplementary Table S2). These findings underscore the escalating public health concern of eating disorders among young adults globally, with particularly alarming increases in bulimia nervosa burden.
Sociodemographic disparities and regional variations in eating disorder burden
The burden of eating disorders among young adults exhibited significant variations across sociodemographic index (SDI) regions, revealing complex patterns of disease distribution from 1990 to 2021 (Figs 1, 2 and 3). Low SDI regions experienced the most substantial absolute increases in both anorexia nervosa and bulimia nervosa case numbers, with increases of 160.67 and 164.12%, respectively (Table 1). However, when examining standardised rates, high-middle SDI regions demonstrated the highest EAPCs in incidence rates for both disorders (anorexia nervosa: 0.73, 95% uncertainty intervals: 0.65 to 0.80; bulimia nervosa: 0.82, 95% uncertainty intervals: 0.76 to 0.89), indicating rapid epidemiological transitions in these low- and middle-income economies.
Prevalence patterns revealed particularly striking trends, with low SDI regions showing a 179.05% increase in bulimia nervosa cases, while middle SDI areas demonstrated the highest EAPCs in prevalence rates for both disorders (anorexia nervosa: 1.00, 95% uncertainty intervals: 0.97 to 1.02; bulimia nervosa: 1.41, 95% uncertainty intervals: 1.39 to 1.43) (Supplementary Table S1). The DALY burden reflected these concerning patterns, with low SDI regions experiencing the largest percentage increases in DALY cases (anorexia nervosa: 160.97%; bulimia nervosa: 181.33%), while middle and high-middle SDI regions exhibited the highest EAPCs in DALY rates (Supplementary Table S2).
Regional analysis revealed substantial geographical heterogeneity in eating disorder trends (Figs 4, 5 and Supplementary Fig. S1). East Asia demonstrated the most pronounced increases in ASRs for both disorders, with percentage changes of 48.74% for anorexia nervosa and 77.90% for bulimia nervosa, accompanied by the highest estimated EAPCs of 1.51 (95% uncertainty intervals: 1.46 to 1.57) for anorexia nervosa and 1.96 (95% uncertainty intervals: 1.85 to 2.08) for bulimia nervosa (Supplementary Table S2). This dramatic increase in East Asia contrasts sharply with high-income North America, which showed minimal changes or slight decreases in ASRs, with EAPCs near zero for anorexia nervosa (0.02, 95% uncertainty intervals: 0.00 to 0.05) and negative for bulimia nervosa (−0.31, 95% uncertainty intervals: −0.48 to 0.14).
Age-standardised rates of eating disorders burden among young adults aged 15–29 years by Global Burden of Disease (GBD) region, 1990–2021. (a) Age-standardised incidence rates of anorexia nervosa and bulimia nervosa by GBD region, presented as a heat map with colour intensity indicating rate magnitude. (b) Age-standardised prevalence rates of anorexia nervosa and bulimia nervosa by GBD region, presented as a heat map with colour intensity indicating rate magnitude. (c) Age-standardised disability-adjusted life year (DALY) rates of anorexia nervosa and bulimia nervosa by GBD region, presented as a heat map with colour intensity indicating rate magnitude.
Regional patterns and trends in eating disorders burden among young adults aged 15–29 years, 1990–2021. (a) Ranking of age-standardised incidence rates of anorexia nervosa and bulimia nervosa by Global Burden of Disease (GBD) region in 1990 and 2021. (b) Ranking of age-standardised prevalence rates of anorexia nervosa and bulimia nervosa by GBD region in 1990 and 2021. (c) Ranking of age-standardised disability-adjusted life year (DALY) rates of anorexia nervosa and bulimia nervosa by GBD region in 1990 and 2021. (d) Estimated annual percentage change (EAPC) in age-standardised incidence rates of anorexia nervosa and bulimia nervosa by GBD region, 1990–2021, presented as a radar chart. (e) EAPC in age-standardised prevalence rates of anorexia nervosa and bulimia nervosa by GBD region, 1990–2021, presented as a radar chart. (f) EAPC in age-standardised DALY rates of anorexia nervosa and bulimia nervosa by GBD region, 1990–2021, presented as a radar chart. ASR, age-standardised rate.
South Asia experienced substantial increases across all metrics, with percentage changes in cases exceeding 120% for anorexia nervosa and 160% for bulimia nervosa (Supplementary Tables S1 and S2). Notably, Central Sub-Saharan Africa was the only region to demonstrate negative percentage changes in ASRs for both disorders, although the EAPCs were not statistically significant, suggesting stable or slightly declining standardised rates despite population growth.
Contemporary global distribution and country-specific patterns in 2021
Analysis of 2021 data revealed significant global disparities in eating disorder burden across 21 regions and 204 countries and territories (Fig. 6, Supplementary Figs S4, S5 and Supplementary Tables S3–5). High-income regions consistently exhibited the highest ASRs, with Western Europe leading in both anorexia nervosa and bulimia nervosa incidence rates. Spain demonstrated the highest anorexia nervosa incidence ASR at 111.59 per 100 000, while Sweden reported the highest bulimia nervosa incidence ASR at 1045.00 per 100 000. Australasia, particularly Australia, showed the highest ASRs for bulimia nervosa prevalence (2154.47 per 100 000) and corresponding DALYs (454.41 per 100 000).
Global distribution and trends in incidence of eating disorders among young adults aged 15–29 years, 2021. (a) Number of incident cases of anorexia nervosa (left) and bulimia nervosa (right) by country in 2021. Countries with the highest values in each of the 21 Global Burden of Disease (GBD) regions are labelled. (b) Age-standardised incidence rates of anorexia nervosa (left) and bulimia nervosa (right) by country in 2021. Countries with the highest rates in each of the 21 GBD regions are labelled. (c) Estimated annual percentage change (EAPC) in age-standardised incidence rates of anorexia nervosa by country, 1990–2021, presented as a world map with a colour gradient indicating EAPC values. (d) EAPC in age-standardised incidence rates of bulimia nervosa by country, 1990–2021, presented as a world map with colour gradient indicating EAPC values. ASR, age-standardised rate.
In contrast, lower-income regions, particularly Sub-Saharan Africa and South Asia, generally reported substantially lower ASRs across all metrics. Eastern Sub-Saharan Africa, specifically Zambia, exhibited the lowest anorexia nervosa incidence ASR (35.42 per 100 000), while Western Sub-Saharan Africa demonstrated the lowest bulimia nervosa incidence ASR (341.07 per 100 000 in Nigeria). However, when considering absolute case numbers, populous lower-middle-income countries bore disproportionately large burdens. India reported the highest incident cases globally for both anorexia nervosa (144 872.26) and bulimia nervosa (1 394 285.97), despite having relatively low ASRs, highlighting the impact of large population sizes on absolute disease burden.
Country-specific analysis revealed substantial heterogeneity in temporal trends from 1990 to 2021. For anorexia nervosa, China exhibited the highest absolute number of cases in 2021 (83 601.86, 95% uncertainty intervals: 54 414.39–122 106.21), while Singapore demonstrated the highest ASR per 100 000 (89.80, 95% uncertainty intervals: 59.18–134.58). Remarkably, Equatorial Guinea experienced the most dramatic percentage increase in anorexia nervosa cases (537.56%), with an exceptionally high EAPC of 1.66 (95% uncertainty intervals: 1.32–2.01). For bulimia nervosa, India’s case burden was substantial (1 394 285.97, 95% uncertainty intervals: 691 404.88–2 497 140.56), while Sweden maintained the highest ASR. Equatorial Guinea again demonstrated the most substantial percentage increase in bulimia nervosa cases (632.57%) with the highest EAPC of 2.12 (95% uncertainty intervals: 1.82–2.42).
Conversely, some countries experienced decreases in incidence, notably Estonia for anorexia nervosa (−29.92% change in cases) and Lithuania for bulimia nervosa (−39.88% change in cases), suggesting potential improvements in prevention or changes in diagnostic practices. The Netherlands showed remarkable increases in anorexia nervosa prevalence, with a 60.64% rise in ASRs and an EAPC of 2.05 (95% uncertainty intervals: 1.86 to 2.24), reaching 12 772.36 cases in 2021. Australia maintained consistently high prevalence rates, with bulimia nervosa cases reaching 108 582.43 and an ASR of 2 154.47 per 100 000 in 2021, representing an 82.08% increase in cases since 1990.
Age-specific patterns and demographic variations
Age-stratified analysis revealed distinct patterns in eating disorder burden across the 15–29 age range from 1990 to 2021 (Fig. 7 and Supplementary Figs S2 and S3). The 15–19 age group consistently demonstrated the highest incidence rates for both disorders, with anorexia nervosa reaching 59.73 cases per 100 000 (95% uncertainty intervals: 37.84–91.06) and bulimia nervosa reaching 462.37 cases per 100 000 (95% uncertainty intervals: 216.08–862.06) in 2021 (Supplementary Table 2). This pattern suggests that eating disorder onset predominantly occurs during mid-to-late adolescence, with early identification and intervention being crucial.
Age-specific trends in eating disorders burden among young adults aged 15–29 years globally, by sociodemographic index (SDI) quintile, and by Global Burden of Disease (GBD) region, 1990–2021. (a) Estimated annual percentage change (EAPC) in age-specific incidence rates of anorexia nervosa and bulimia nervosa, presented as a heat map. Rows represent global, SDI quintiles and GBD regions; columns represent 5-year age groups from 15 to 29 years. (b) EAPC in age-specific prevalence rates of anorexia nervosa and bulimia nervosa, presented as a heat map. Row and column structure as in panel (a). (c) EAPC in age-specific disability-adjusted life year (DALY) rates of anorexia nervosa and bulimia nervosa, presented as a heat map. Row and column structure as in panel (a).
Prevalence patterns showed a different distribution, with the 20–24 age group exhibiting the highest rates for both disorders in 2021: 144.46 per 100 000 (95% uncertainty intervals: 88.02–219.28) for anorexia nervosa and 425.66 per 100 000 (95% uncertainty intervals: 214.1–730.24) for bulimia nervosa (Supplementary Table S6). The DALY burden was similarly concentrated in the 20–24 age group, with bulimia nervosa causing substantially higher burden (90.35 DALYs per 100 000; 95% uncertainty intervals: 39.59–164.88) compared to anorexia nervosa (31.00 DALYs per 100 000; 95% uncertainty intervals: 17.67–50.9) in 2021 (Supplementary Table S7).
Temporal changes showed consistent increases across all age groups, with the 25–29 age group experiencing the largest percentage increases in bulimia nervosa prevalence (61.04%) and DALYs (61.07%) from 1990 to 2021. All EAPCs were positive across age groups for both disorders, with bulimia nervosa consistently showing higher rates of increase compared to anorexia nervosa.
Age-specific analysis across SDI levels revealed complex interactions between socioeconomic development and demographic patterns. In high SDI regions, anorexia nervosa incidence among 15–19 year-olds increased modestly by 1.55% to 67 563 cases in 2021, while low SDI regions experienced dramatic increases of 161.65% to 53 667 cases. Middle SDI regions showed particularly concerning trends for bulimia nervosa with the 25–29 age group experiencing a 53.98% increase in incidence to 381 881 cases, accompanied by an 82.95% increase in prevalence to 746 569 cases.
Regional age-specific patterns further highlighted global disparities. East Asia demonstrated the highest EAPCs across all age groups, with the 15–19 age group showing an EAPC of 1.45 (95% uncertainty intervals: 1.39 to 1.52) for anorexia nervosa despite a 12.07% decrease in absolute case numbers, indicating substantial population demographic changes. South Asia experienced consistent increases across all age groups, with EAPCs ranging from 0.70 to 0.81, while high-income regions like North America showed more moderate changes with EAPCs between 0.14 and 0.31.
Machine-learning model performance and future projections
To provide evidence-based projections for healthcare planning, we evaluated eight distinct machine-learning algorithms for forecasting eating disorder trends: Prophet, ARIMA, TBATS, ElasticNet, ETS, VAR, Holt–Winters and Theta models (Fig. 8 and Supplementary Tables S8–12). Model selection was based on comprehensive performance evaluation using multiple metrics including MSE, MAPE, RMSE and R 2.
Projected trends in age-standardised rates of eating disorders burden among young adults aged 15–29 years, 2022–2050. (a) Projected age-standardised incidence rates of anorexia nervosa, 2022–2050, based on eight machine-learning models. Each line represents a different model’s projection. (b) Projected age-standardised prevalence rates of anorexia nervosa, 2022–2050, based on eight machine-learning models. Line representation as in panel (a). (c) Projected age-standardised disability-adjusted life year (DALY) rates of anorexia nervosa, 2022–2050, based on eight machine-learning models. Line representation as in panel (a). (d) Projected age-standardised incidence rates of bulimia nervosa, 2022–2050, based on eight machine-learning models. Line representation as in panel (a). (e) Projected age-standardised prevalence rates of bulimia nervosa, 2022–2050, based on eight machine-learning models. Line representation as in panel (a). (f) Projected age-standardised DALY rates of bulimia nervosa, 2022–2050, based on eight machine learning models. Line representation as in panel (a). ARIMA, Autoregressive Integrated Moving Average; ETS, Error Trend Seasonality; TBATS, Box– Cox transformation ARMA Errors Trend and Seasonal Components; VAR, Vector Autoregression.
For anorexia nervosa incidence rate predictions, the Prophet model demonstrated superior performance with the lowest MSE (0.002243), MAPE (0.084995) and RMSE (0.047363), achieving the highest R 2 value (0.994606). The Prophet model’s strength lies in its ability to handle trend changes and seasonal patterns while being robust to missing data and outliers. The ARIMA model ranked second in performance, while the TBATS and ETS models showed inadequate performance with negative R 2 values, indicating poor model fit for anorexia nervosa data characteristics.
Bulimia nervosa incidence rate forecasting showed similar patterns, with the Prophet model again achieving optimal performance: lowest MSE (0.101725), MAPE (0.072919), RMSE (0.318943) and highest R 2 (0.997493). The ElasticNet model, which combines Ridge and Lasso regression techniques, ranked second for bulimia nervosa predictions, suggesting that regularisation approaches may be valuable for bulimia nervosa trend modelling.
Prevalence rate modelling revealed slightly different optimal approaches. For anorexia nervosa prevalence, the Prophet model maintained superior performance with MSE (0.087950), RMSE (0.296564) and R 2 (0.972392), with ARIMA showing comparable results. However, for bulimia nervosa prevalence, the ARIMA model outperformed others with MSE (0.184974), MAPE (0.093134), RMSE (0.430086) and R 2 (0.997549), suggesting that classical time-series approaches may be more suitable for bulimia nervosa prevalence patterns.
For DALY rate projections, the Prophet model excelled for anorexia nervosa with MSE (0.004826), RMSE (0.069470) and R 2 (0.963292), while the ARIMA model was optimal for bulimia nervosa with MSE (0.010152), MAPE (0.106085), RMSE (0.100756) and R 2 (0.996908).
Based on the best-performing models, projections indicate concerning continued increases in both disorders through 2050. Anorexia nervosa incidence rates are projected to rise from 43.40 to 45.50 per 100 000 population, while bulimia nervosa incidence is expected to increase more substantially from 353.10 to 391.08 per 100 000 population between 2022 and 2050 (Supplementary Table S8). Prevalence projections suggest anorexia nervosa will increase from 133.41 to 139.31 per 100 000 (Prophet model), while bulimia nervosa prevalence is projected to rise from 374.90 to 389.15 per 100 000 (ARIMA model) (Supplementary Table S9). DALY projections indicate continued burden increases, with anorexia nervosa expected to rise from 28.63 to 29.15 per 100 000 and bulimia nervosa from 79.49 to 83.29 per 100 000 population by 2050 (Supplementary Table S10).
These forecasting results highlight the persistent and growing nature of the eating disorder epidemic, emphasising the critical need for enhanced prevention strategies, early intervention programmes and healthcare system capacity building to address the anticipated increases in disease burden over the coming decades.
Discussion
Summary of main findings
This comprehensive global analysis represents the first study to systematically estimate eating disorder burden among adolescents and young adults at global, regional and national levels using GBD 2021 data, while innovatively employing eight advanced machine-learning algorithms to predict future trends through 2050. Our findings reveal concerning upward trajectories in eating disorder burden from 1990 to 2021, with bulimia nervosa demonstrating more pronounced growth than anorexia nervosa across all epidemiological metrics. The burden exhibits significant global disparities, with highest incidence rates concentrated in the 15–19 age group for anorexia nervosa and the 20–24 age group for bulimia nervosa. Regional analysis revealed substantial heterogeneity, with East Asia showing the most dramatic increases while maintaining lower absolute burden compared to high-income Western regions. Machine-learning projections indicate continued increases through 2050, emphasising the persistent and escalating nature of this global public health challenge.
Socioeconomic development and epidemiological transitions
The complex relationship between socioeconomic development and eating disorder trends reflects multifaceted epidemiological transitions occurring globally. Low SDI regions experienced the most substantial absolute increases in case numbers, likely attributable to cultural shifts associated with globalisation, rapid social change and extensive media exposure. Reference Keski-Rahkonen23 These regions face a dual burden: rising incidence coupled with limited healthcare infrastructure for adequate diagnosis and treatment, potentially leading to delayed recognition and management. Reference Linardon, Messer, Rodgers and Fuller-Tyszkiewicz24,Reference Devoe, Han, Anderson, Katzman, Patten and Soumbasis25 The scarcity of healthcare resources and diagnostic capabilities in these settings may paradoxically result in both underdiagnosis of existing cases and delayed intervention for emerging cases.
Conversely, middle and high-middle SDI regions demonstrated the highest age-standardised rate increases, suggesting that economic development correlates with eating disorder risk through multiple pathways. Higher socioeconomic-status countries often subject young adults to considerable social pressures and performance expectations, creating psychological vulnerabilities that may precipitate disordered eating behaviours as coping mechanisms. The widespread influence of electronic media in these regions exacerbates risk by promoting unrealistic body ideals and contributing to body dissatisfaction. Reference Keski-Rahkonen10,Reference Izydorczyk and Sitnik-Warchulska26 Additionally, improved healthcare access in higher SDI regions facilitates earlier detection and reporting, contributing to observed increases in documented cases through enhanced diagnostic capacity and reduced treatment-seeking barriers.
Regional heterogeneity and cultural transformations
Regional disparities reflect complex interactions between traditional cultural values and modern lifestyle changes. East Asia’s dramatic increases mirror broader patterns observed during rapid societal transformation, where traditional values intersect with Western beauty ideals and urbanisation-driven lifestyle modifications. Reference Kim, Nakai and Thomas27,28 China’s rapid economic growth since the 1990s exemplifies this phenomenon, where increased urbanisation and accelerated life pace elevated stress levels, prompting some individuals to adopt food-related coping strategies. Simultaneously, economic prosperity altered dietary patterns through the increased availability of high-calorie processed foods, while media emphasis on thinness and beauty standards heightened body dissatisfaction and eating disorder risk. Reference Wu, Lin, Liu, He, Bai and Lyu29
High-income Western regions, particularly Western Europe and Australasia, consistently exhibited the highest absolute burden, consistent with established sociocultural frameworks emphasising thinness ideals as fashion standards. Reference Burke, Hazzard, Schaefer, Simone, O’Flynn and Rodgers30 However, the relatively stable or slightly declining trends in some high-income regions may indicate either intervention effectiveness, diagnostic saturation or successful implementation of prevention programmes developed over decades of clinical experience.
The substantially lower reported burden in Sub-Saharan Africa likely represents significant underdiagnosis rather than true low prevalence, reflecting healthcare system limitations, diagnostic tool availability and cultural differences in symptom recognition and help-seeking behaviours. Reference Mikhail and Klump31 Research within South Africa revealed that many Black African adolescents maintained positive body image with fewer weight loss attempts compared to other demographic groups, suggesting cultural protective factors that may influence both actual prevalence and diagnostic patterns. Reference Morris and Szabo32
Age-related vulnerabilities and developmental considerations
Age-specific patterns align closely with established developmental psychology frameworks and clinical observations. Anorexia nervosa’s predominant impact on the 15–19 age group corresponds with critical identity formation periods and heightened body image concerns during pubertal transitions. Reference Solmi, Monaco, Højlund, Monteleone, Trott and Firth33,Reference Solmi, Radua, Olivola, Croce, Soardo and Salazar de Pablo34 Adolescents navigate complex developmental challenges where excessive preoccupation with weight and appearance can precipitate anorexia nervosa, particularly during periods characterised by maturity fears and self-perception establishment. Reference Yamamiya, Desjardins and Stice35 Media portrayal of idealised body standards places additional pressure on this vulnerable population, intensifying focus on weight and appearance.
Bulimia nervosa’s later onset, peaking in the 20–24 age group, reflects different psychological and social pressures characteristic of early adulthood transitions. Reference Yamamiya and Stice36,Reference Rohde, Stice, Shaw, Gau and Ohls37 This period involves metabolic changes that complicate weight management, potentially triggering bulimic behaviours in susceptible individuals. Additionally, complex adult expectations including academic achievement, career development and social competencies create multifaceted stressors that may lead some individuals to cope through binge-eating patterns. Reference Neale, Pais, Nicholls, Chapman and Hudson38 These age-specific patterns support targeted intervention approaches addressing developmental vulnerabilities and risk factors specific to different life stages.
Novel contributions: machine-learning forecasting and methodological advances
Our multi-model forecasting approach represents a significant methodological advancement in epidemiological prediction, demonstrating that different eating disorders require distinct analytical approaches for optimal accuracy. The Prophet model’s superior performance for anorexia nervosa predictions, contrasted with ARIMA’s effectiveness for bulimia nervosa forecasting, suggests fundamental differences in underlying temporal dynamics between these disorders. This variation in model performance underscores the critical importance of employing diverse forecasting methodologies rather than relying on single-model approaches.
The forecasting results provide crucial evidence for healthcare planning and resource allocation, indicating continued burden increases requiring proactive intervention strategies. Projections suggest anorexia nervosa and bulimia nervosa will continue rising through 2050, emphasising the persistent nature of this epidemic and the urgent need for evidence-based prevention and intervention-programme development.
Mechanistic interpretations and evidence synthesis
Several converging mechanisms likely drive observed trends, though distinguishing between empirically-supported and speculative explanations remains crucial. Well-documented factors include rapid urbanisation and lifestyle changes that increase stress levels while reducing traditional social support systems. The COVID-19 pandemic likely accelerated existing trends through social isolation, routine disruption and heightened mental health challenges. Reference Linardon, Messer, Rodgers and Fuller-Tyszkiewicz24,Reference Devoe, Han, Anderson, Katzman, Patten and Soumbasis25
However, distinguishing between true incidence increases and improved diagnostic recognition presents a fundamental challenge in eating disorder epidemiology. The DSM-5’s broadened diagnostic criteria offer more comprehensive understanding of the eating-disorder spectrum, potentially contributing to apparent increases through enhanced case identification. Reference Himmerich, Kan, Au and Treasure39,Reference Bulik, Coleman, Hardaway, Breithaupt, Watson and Bryant40 Simultaneously, reduced mental health stigma and increased awareness may encourage help-seeking behaviours, inflating observed trends through improved detection rather than actual incidence changes.
Cultural modernisation, particularly increased availability of high-calorie processed foods with addictive qualities, significantly influences dietary habits and may exacerbate binge-eating behaviours. Reference Himmerich and Mirzaei41,Reference Yu, Robinson, Bobou, Zhang, Banaschewski and Barker42 Social media proliferation creates unprecedented exposure to appearance-related pressures, though direct causal relationships require further investigation beyond our current data-set capabilities.
Public health implications and intervention priorities
The projected burden increases through 2050 demand immediate, coordinated policy responses across multiple sectors. Prevention strategies should prioritise high-risk age groups through comprehensive school-based programmes addressing body image, media literacy and healthy coping mechanisms. Early intervention systems require substantial strengthening, particularly in rapidly developing regions experiencing the steepest increases and limited diagnostic infrastructure.
Healthcare workforce training and diagnostic capacity building represent critical priorities, especially in low-resource settings where underdiagnosis persists. Mental health support systems must be enhanced to identify and treat underlying psychological conditions before they progress to clinical eating disorders. Reference Himmerich, Kan, Au and Treasure39,Reference Bulik, Coleman, Hardaway, Breithaupt, Watson and Bryant40 Educational initiatives promoting body positivity and self-esteem should challenge societal beauty standards while providing guidance for healthy social media engagement.
Global coordination mechanisms are essential to address transnational factors including social media influence and cultural globalisation effects. Policy frameworks should integrate eating-disorder prevention into broader mental health and adolescent health initiatives, recognising the interconnected nature of these challenges. International collaboration can facilitate knowledge sharing, resource allocation and coordinated intervention development across regions experiencing different epidemiological phases.
Study limitations
This analysis relies on GBD 2021 data, which may be subject to reporting biases, underdiagnosis and misclassification, particularly in regions with limited healthcare infrastructure and diagnostic resources. Cultural differences in symptom presentation and help-seeking behaviours may introduce additional uncertainties across diverse global populations. Our forecasting models assume historical trend continuity, potentially missing discontinuous changes from future interventions, policy implementations or emerging risk factors. The models cannot fully incorporate potential impacts of public health interventions, technological advances or societal changes that could significantly alter disorder trajectories. Additionally, our focus on ages 15–29 excludes other affected populations, and the analysis cannot fully account for comorbidities or indirect societal impacts. The linear assumptions inherent in our modelling approach may not capture complex non-linear relationships or threshold effects that could influence future trends.
Summary and implications
This comprehensive analysis reveals concerning upward trajectories in eating disorder burden among young adults globally, with bulimia nervosa showing particularly rapid growth patterns. Regional disparities reflect complex interactions between socioeconomic development, cultural transformations and healthcare accessibility. The application of multiple machine-learning models provides robust projections indicating sustained increases through 2050, emphasising the urgent need for evidence-based prevention and intervention strategies tailored to regional contexts and age-specific vulnerabilities. Our methodological approach demonstrates the value of multi-model forecasting in capturing disorder-specific temporal dynamics while providing actionable intelligence for public health planning. Addressing this growing epidemic requires coordinated global action integrating clinical care enhancement, comprehensive prevention programmes, policy interventions and international collaboration to mitigate projected burden increases and support holistic youth well-being.
Supplementary material
The supplementary material is available online at https://doi.org/10.1192/bjp.2025.10450
Data availability
All the data generated or analysed during this study are included in this published article and its supplementary files. The data used for the analyses in the study are publicly available at https://ghdx.healthdata.org/gbd-2021.
Acknowledgements
This study acknowledges the support of the Key Laboratory of Digital-Intelligent Disease Surveillance and Health Governance of Sichuan Province. We thank Dr Bo Tang from Sichuan Provincial People’s Hospital for his assistance with this article. We also acknowledge Professor Allan Young, Director of the Centre for Affective Disorders, Department of Psychological Medicine, Institute of Psychiatry, Psychology & Neuroscience, King’s College London, for his expertise in manuscript preparation and contributions to the field of eating disorders research. We also acknowledge Professor Chao Deng, Director of Psychotropic Drugs Research Unit, University of Wollongong, for his expertise in our study design, manuscript writing and revision.
Author contributions
Conceptualisation: J.L., N.L. and L.L.; formal analysis: L.L., K.W., M.D. and L.T.; resources: L.L., K.W., X.D., L.W. and J.L.; software: L.L., M.D., L.W. and K.W.; supervision: Y.L., J.L. and N.L.; visualisation: L.T., K.W., L.L. and M.D.; writing–original draft: L.L. and K.W.; writing–review and editing: K.W., J.L. and W.L.; revision: K.W., J.L., L.L. and N.L.; validation: L.L. and K.W.; funding acquisition: L.L., Y.L. and N.L.; project administration: N.L. and J.L. All authors read and approved the final manuscript.
Funding
This study was supported by the Sichuan Science and Technology Program (grant number 2024ZYD0272, 2024NSFSC2107, 2024YFFK0057) and the Natural Science Foundation of North Sichuan Medical College (grant number CBY24-QDA11).
Declaration of interest
None.
Open access
eLetters
No eLetters have been published for this article.