https://orcid.org/0009-0008-8877-3169
Figures
Abstract
Background
The prevalence of under-five stunting in Ethiopia remains above 30%, which, according to World Health Organization (WHO), constitutes a major public health concern. Stunting has long-term consequences for child growth, development, and overall health. Accurate forecasting of its prevalence is therefore essential to guide policymakers and inform targeted interventions. This study aimed to forecast the prevalence of under-five stunting in Ethiopia for the period 2025–2030 using historical data and time series modeling.
Methods
Annual under-five stunting prevalence data for Ethiopia from 2000 to 2024 were retrieved from the WHO Global Health Observatory. Time series forecasting models, including Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing (ETS), Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM), were developed and evaluated. Model performance was assessed using time series cross-validation with mean absolute error (MAE), mean absolute percentage error (MAPE), and R² as evaluation metrics. The best-performing model was applied to forecast stunting prevalence for 2025–2030.
Results
The ETS model demonstrated the best predictive performance (MAE = 1.09, MAPE = 2.71%, R² = 0.903) and was selected for forecasting. Forecasting indicated a gradual decline in under-five stunting prevalence in Ethiopia from 33.95% in 2025 to 31.95% in 2030. The projected prevalence for 2029 is 32.35% (95% CI: 26.90–37.79%), above the national target of 19%, and the 2030 forecast remains well above the SDG target of ending all forms of malnutrition.
Citation: Gessese RT, Kelkay JM, Arage FG, Tsegaw TK, Tadese ZB, Alemayehu MA, et al. (2026) Forecasting under-five stunting in Ethiopia using classical and machine learning time series models. PLoS One 21(4): e0345000. https://doi.org/10.1371/journal.pone.0345000
Editor: Zaher Mundher Yaseen, KFUPM: King Fahd University of Petroleum & Minerals, SAUDI ARABIA
Received: November 14, 2025; Accepted: February 25, 2026; Published: April 22, 2026
Copyright: © 2026 Gessese et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the paper and its Supporting Information file.
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Background
Malnutrition refers to deficiencies, excesses, or imbalances in a person’s intake of energy and/or nutrients. It encompasses undernutrition (including stunting, wasting, underweight, and micronutrient deficiencies) as well as over nutrition (overweight and obesity) [1]. Among the forms of undernutrition, stunting is defined as a height-for-age Z-score below −2 standard deviations from the World Health Organization (WHO) Child Growth Standards and represents chronic nutritional deficits in children under five [2].
In 2024, about 150.2 million children under five were stunted worldwide, with most cases occurring in Asia (51%) and Africa (43%) [3]. In Ethiopia, the 2019 Ethiopia Mini Demographic and Health Survey reported that 37% of children under five were stunted, while a 2023 nationwide nutrition survey conducted by Ministry of Health and partners estimated the prevalence at 39% [4,5]. More recently, a 2024 systematic review and meta-analysis reported a pooled prevalence of 43% [6]. Overall, these findings show that the prevalence of stunting in Ethiopia ranges between 37% and 43%. All estimates exceed the WHO threshold of 30%, highlighting the very high public health significance of child stunting [2].
Stunting is associated with multiple adverse outcomes in children. Health-related complications include increased susceptibility to infections such as diarrhea, pneumonia, and measles due to impaired immunity, as well as higher child mortality, with 45% of all child deaths linked to undernutrition worldwide [7,8]. Stunted children also experience delayed physical growth, which may result in permanent short stature, and impaired cognitive development, leading to lower learning capacity, poorer school performance, and reduced educational attainment. In addition, stunting increases the risk of chronic diseases such as obesity, and cardiovascular conditions later in life [8–12]. In Ethiopia, 28% of all cases of child mortality are associated with the higher risk of undernutrition. Childhood undernutrition also imposes an economic cost equivalent to 0.5% of the total public health budget, while productivity losses linked to reduced educational attainment and diminished adult earnings are estimated to cost the country up to 16% of its annual GDP in Ethiopia [13,14].
Addressing stunting is central to achieving the Sustainable Development Goals (SDGs), particularly SDG 2 (Zero Hunger) and SDG 3 (Good Health and Well-being). Recognizing this, the Ethiopian government has implemented programs and strategies such as the National Nutrition Program (NNP), the Seqota Declaration, and community-based interventions including growth monitoring and micronutrient supplementation [15–19]. While these initiatives have yielded some gains, persistent socioeconomic, environmental, and behavioral barriers continue to slow progress [6,20,21].
Accurate forecasting of under-five stunting is essential to guide Ethiopia’s response. Forecasts enable monitoring of progress toward national and global nutrition targets, improve resource allocation, and support evidence-based policy design. For forecasting, traditional statistical methods, such as ARIMA and Exponential Smoothing (ETS) as well as deep learning machine learning approaches, including MLP, and LSTM networks, offer promising tools for capturing temporal dynamics and generating reliable projections [22–24]. By comparing these approaches, robust forecasts can be generated to inform policymakers, strengthen intervention strategies, and assess whether ongoing efforts are sufficient to meet nutrition targets. Against this background, the current study seeks to evaluate and compare statistical and machine learning models for forecasting under-five stunting in Ethiopia, with the ultimate goal of identifying the most accurate and practical approach for predicting the future burden.
Method
Data preparation and preliminary assessment
The analysis was conducted on a univariate time series dataset comprising 24 annual observations of under-five stunting prevalence in Ethiopia, spanning from 2000 to 2024. The data were obtained from the WHO Global Health Observatory repository and extracted programmatically using Python, with the pandas and requests libraries to facilitate automated retrieval. The raw data were loaded into a pandas DataFrame, with the Year column converted to a datetime object and set as the index to maintain chronological order. Missing data were checked.
Prior to modeling, the dataset underwent a preliminary assessment to ensure readiness for time series forecasting. Line plots were inspected to identify trends or anomalies. Stationarity is a fundamental requirement for many time series models, including ARIMA, because it ensures that the statistical properties of the series such as mean and variance remain constant over time, enabling reliable forecasts [25]. To assess stationarity, the series was first visually inspected using rolling mean and rolling standard deviation over a three-year window. Formal statistical tests were then applied using the Augmented Dickey-Fuller (ADF) test (p < 0.05) indicates stationarity). Differencing was applied where non-stationarity was detected, and the tests were repeated to confirm stationarity. Finally, the series was decomposed using STL (Seasonal-Trend decomposition using Loess) to isolate trend, seasonality and residual components (S1 Fig).
For the deep learning models (MLP and LSTM), additional preprocessing was applied to the dataset. The series was scaled to the range [0, 1] using MinMaxScaler to improve convergence during model training. Input sequences were then constructed using lagged values of the series (the previous three years of prevalence to predict the subsequent year) to serve as features.
Modeling approaches
To forecast under-five stunting prevalence, multiple candidate models were implemented, including classical statistical models (ARIMA and ETS), deep learning models (MLP and LSTM), and a hybrid ETS–LSTM approach. These models were systematically assessed to identify the best-performing model for forecasting.
Classical time series methods
ARIMA is a widely used time series model that captures both autoregressive (AR) and moving average (MA) components, with differencing applied to ensure stationarity [26]. The ARIMA model is expressed as ARIMA (p,d,q)), where p represents the autoregressive order, d the degree of differencing, and q the moving average order. ARIMA is a widely used time series approach that captures both autoregressive (AR) and moving average (MA) components, with differencing applied to ensure stationarity. For this study, the series was first tested for stationarity using the Augmented Dickey–Fuller (ADF) test using the statsmodels library. Candidate values for p and were guided by inspection of the autocorrelation (ACF) and partial autocorrelation (PACF) plots, and set as p = 0,1,2,3,5,6 and q = 0,1,2. A systematic grid search was then conducted across these ranges, with model comparison based on the Akaike Information Criterion (AIC) (S2 Fig).
The ETS approach was employed using statsmodels as a second classical statistical model. ETS models decompose a time series into three components: Error (E), Trend (T), and Seasonality (S), which can each be specified as either additive (A) or multiplicative (M), or set to none (N) if not present [27]. The ETS framework was applied to capture level and trend components of the annual series. Candidate models considered included ANN (Additive Error, No Trend, No Seasonality), AAN (Additive Error, Additive Trend, No Seasonality), AMN (Additive Error, Multiplicative Trend, No Seasonality), MNN (Multiplicative Error, No Trend, No Seasonality), MAN (Multiplicative Error, Additive Trend, No Seasonality), and MMN (Multiplicative Error, Multiplicative Trend, No Seasonality). Seasonality was excluded based on the results of the STL analysis. A grid search over the level and trend smoothing parameters was conducted, with model selection based on the Akaike Information Criterion (AIC).
Deep learning models
Deep learning models are a subset of machine learning algorithms capable of automatically learning complex, hierarchical representations from data. They are particularly well-suited for time series forecasting, as they can capture nonlinear patterns, temporal dependencies, and long-term trends that traditional statistical models may fail to detect [28,29]. In this study, two deep learning architectures Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM) networks were implemented to explore their potential in forecasting under-five stunting prevalence. These models were chosen based on their demonstrated effectiveness in prior studies and their strong performance in capturing temporal patterns in health and epidemiological time series data [30,31]. Both models used the dataset prepared and assessed in the Data Preparation and Preliminary Assessment phase.
MLPs were implemented using scikit-learn. Various architectures were evaluated, including single hidden-layer and shallow two hidden-layer networks with configurations: (5,), (8,), (5,3), (8,4), and (10,5). Activation functions tested included ReLU and tanh. Key hyperparameters and their search ranges were: L2 regularization α (0.001–0.5), learning rate (0.001–0.1), batch size (2–8), and maximum iterations (500–1500).
Hyperparameter tuning was performed using RandomizedSearchCV, allowing efficient exploration of the hyperparameter space. Early stopping with patience between 20–50 epochs was applied to prevent overfitting. The final MLP model selected was a shallow two-layer network with 8 and 4 units in the first and second layers, respectively, ReLU activation, L2 regularization α = 0.1, learning rate 0.1, batch size 8, and a maximum of 1000 iterations. This configuration provided optimal validation performance while minimizing model complexity, which is important given the small dataset.
LSTM networks were implemented using TensorFlow Keras. A single LSTM layer with 2–5 units was followed by a dropout layer (0.2–0.5) and a dense output layer for regression. L2 regularization (l1_l2) was applied to both kernel and recurrent weights. The Adam optimizer was used with mean squared error as the loss function.
Given the higher computational cost and the small size of the dataset, hyperparameter tuning for LSTM was conducted via a controlled manual random search across the following ranges: number of units (2–5), dropout rate (0.2–0.5), L2 regularization strength (0.001–0.1), learning rate (0.001–0.1), batch size (2–8), and number of epochs (50–150). Early stopping (EarlyStopping) and learning rate reduction callbacks were applied to ensure stable convergence and prevent overfitting. The final LSTM model used 4 units, dropout rate 0.2, L2 regularization 0.001, batch size 8, 50 epochs, and a learning rate of 0.01. This shallow, single-layer architecture with regularization strategies was chosen to maintain model simplicity and robustness, while still capturing temporal dependencies in the annual stunting data
Based on the performance of the statistical model, a hybridization approach with LSTM was evaluated. Residuals from the best-performing statistical model were extracted and assessed for remaining patterns. An LSTM model was then trained on these residuals using the same hyperparameter tuning, dropout, L2 regularization, early stopping, and time-series cross-validation procedures applied to the standalone LSTM model. The predictive performance of the hybrid model was compared with a naive baseline model, which predicts the mean of the residuals, to determine whether the LSTM captured additional meaningful patterns beyond those already modeled by the statistical approach.
Model performance evaluation
Predictive performance of all models was assessed using time series cross-validation, implemented via the TimeSeriesSplit function from scikit-learn, ensuring that training data always preceded testing data chronologically. For each fold, models were trained on the training subset and evaluated on the test subset. Performance metrics included mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R²). For deep learning models, early stopping and learning rate reduction were applied during training to prevent overfitting and ensure stable convergence. Model selection was based on cross-validated performance across these metrics.
Results
Temporal patterns of under-five stunting in Ethiopia
Over the 24-year period from 2000 to 2024, the prevalence of stunting among children under five in Ethiopia shows a clear declining trend, decreasing from approximately 57% to around 35%. Minor fluctuations observed in certain years reflect short-term variability (Fig 1).
Stationarity test
Visual inspection of the rolling statistics (rolling mean and standard deviation) of the original under-five stunting series revealed noticeable trends and fluctuations over time, indicating that the series was non-stationary. This observation was formally confirmed using the Augmented Dickey-Fuller (ADF) test, which yielded a p-value of 0.62, supporting the presence of a unit root. To stabilize the series, first-order differencing was applied. Post-differencing, the ADF test indicated stationarity (p < 0.001), and the rolling mean and standard deviation plots showed a stable mean and variance throughout the series (Figs 2, 3).
Modeling approaches for forecasting under-five stunting
Classical model.
For ARIMA, the series was first differenced to achieve stationarity. After evaluating candidate autoregressive and moving average orders guided by autocorrelation and partial autocorrelation patterns, ARIMA (1,1,2) was selected as the best-fitting model based on the lowest AIC. This model demonstrated predictive performance of MAE = 1.15, MAPE = 2.82%, and R² = 0.88. Among the candidate ETS models, the AAN (Additive Error, Additive Trend, No Seasonality) specification had the lowest AIC (20.00). This ETS (A,A,N) model demonstrated predictive performance with MAE = 1.09, MAPE = 2.71%, and R² = 0.90.
Deep learning model.
For the multilayer perceptron (MLP) model, the final input feature set after hyperparameter tuning included prevalence, lag_1, lag_2, lag_3, rolling_mean_2, rolling_std_2, rolling_mean_3, rolling_std_3, and time_index. The model was selected based on the lowest cross-validation MAE of 0.426. The optimized network consisted of two hidden layers with 8 and 4 neurons, respectively, using the ReLU activation function. Other selected hyperparameters were: n_iter_no_change = 20, max_iter = 1000, learning_rate_init = 0.1, early_stopping = True, batch_size = 8, and alpha = 0.1. Predictive performance metrics were MAE = 0.426, MAPE = 7.14%, and R2 = 0.49% LSTM model achieved its best performance with 4 units, a dropout rate of 0.2, L2 regularization of 0.001, a batch size of 8, 50 epochs, and a learning rate of 0.01. When evaluated using time-series cross-validation, this configuration yielded MAE of 4.52, MAPE of 11.27% and R² of 0.39%.
Hybrid model.
Based on model performance evaluation, ETS (A,A,N) was identified as the best-performing statistical model. The ETS model produced residuals with a near-zero mean (0.0047) and standard deviation of 1.10, indicating only small fluctuations around the fitted trend. Analysis further showed no significant autocorrelation at lags 1–3 (0.078, 0.012, –0.230), confirming that the residuals essentially represented random noise and that the ETS model had already captured the predictable patterns in the series.
Despite the absence of residual patterns, the LSTM was trained using time-series cross-validation with hyperparameters optimized to minimize MAE (units = 4, dropout rate = 0.2, L2 regularization = 0.001, batch size = 8, epochs = 50, learning rate = 0.01). The resulting hybrid model achieved MAE = 1.69, RMSE = 1.82, and MAPE = 309.8%. In comparison, a naive predictor using the mean of the residuals achieved a lower MAE of 1.04, indicating that the hybrid ETS–LSTM model performed worse than a simple baseline.
These findings confirm that, although the hybrid approach was performed, adding an LSTM to the ETS residuals did not capture additional meaningful patterns and likely overfit the noise. Therefore, the ETS (A,A,N) model remains the most reliable forecasting model for under-five stunting prevalence (S3 Fig).
Six-year forecast of under-five stunting in Ethiopia
Based on model evaluation, the ETS (A, A, N) model was selected as the most reliable for forecasting under-five stunting prevalence. Using this model, the prevalence was projected for the next six years. By 2029, the stunting prevalence is projected at 32.35% (95% CI: 26.90–37.79), and by 2030, it is projected at 31.95% (95% CI: 25.98–37.91), indicating a gradual decline over the forecast period if current trends continue (Fig 4).
The six-year forecast of under-five stunting prevalence is summarized in Table 1, showing the projected values and their 95% confidence intervals (Table 1).
Discussion
This study projected the prevalence of under-five stunting in Ethiopia from 2000 to 2030 using classical time series and machine learning models. Historical data showed a declining trend, with stunting prevalence decreasing from approximately 57% in 2000 to 35% in 2024. The observed decline in under-five stunting prevalence in Ethiopia from 2000 to 2024 can be attributed to multiple factors, including sustained public health interventions, expanded access to primary healthcare services, and multisectoral initiatives such as the Sekota Declaration, which foster collaboration across sectors addressing food security, sanitation, and education [32–35].
Among the models evaluated, ETS was the most reliable, capturing the trend effectively (MAE = 1.09, MAPE = 2.71%, R² = 0.9028). This suggests that classical statistical methods can perform well when dealing with relatively small datasets, where deep learning or more complex machine learning models may risk overfitting or require larger training samples to achieve stable results. The superior performance of the ETS model in forecasting under-five stunting prevalence can be explained by several methodological factors. First, the relatively small sample size of the annual stunting series increases the risk of overfitting for complex machine learning and deep learning models, whereas ETS, with its parsimonious parameter structure, can produce stable and reliable forecasts [36] Second, ETS is particularly well suited to univariate forecasting tasks, as it relies solely on historical values and does not require additional explanatory covariates. In the absence of such covariates, machine learning and deep learning models may struggle to identify meaningful patterns, whereas ETS remains robust and interpretable [37]. Collectively, these factors highlight why ETS can outperform more complex models in settings with limited, smooth, and univariate health outcome data. This finding is also consistent with previous studies, which have demonstrated that exponential smoothing methods are often robust, interpretable, and computationally efficient for forecasting health-related outcomes with limited data [38,39].
The prevalence of under-five stunting is projected to be 32.35% (95% CI: 26.9–37.79) by 2029, which remains above Ethiopia’s national target of 19% set by the Ministry of Health’s National Nutrition Program [16]. By 2030, it is expected to decline further to 31.95% (95% CI: 25.98–37.91); however, this projected prevalence still exceeds the SDG 2.2 aim of ending all forms of malnutrition by 2030 [40].
Despite ongoing nutrition initiatives, several factors may hinder the achievement of these targets. Ethiopia has experienced decades of challenges stemming from drought, armed conflict, climate change, political and economic instability, the COVID-19 pandemic, and the recent surge in living costs these challenges, along with poor diet quality, unsafe complementary feeding practices, continue to exacerbate children’s vulnerability to stunting and malnutrition [41–48]. Indicating that current interventions alone may be insufficient.
While these findings provide valuable insights into Ethiopia’s progress toward national and global nutrition targets, it is important to interpret them in light of the study’s strengths and limitations. A key strength is the use of 24 years of nationally representative data from the WHO Global Health Observatory, which provides a robust evidence base for examining long-term stunting trends. In addition, the application and comparison of multiple statistical and machine learning models enhanced the reliability of the projections. Benchmarking the results against Ethiopia’s national nutrition targets and the global SDG 2.2 also ensured that the findings are directly relevant to both national policy and global commitments. However, the analysis was conducted using national-level annual data, which may obscure important regional and sub-population disparities in stunting prevalence. Furthermore, the dataset did not include socioeconomic or other contextual variables, limiting the ability to account for factors that could influence stunting trends.
Conclusion and recommendation
This study projected the prevalence of under-five stunting in Ethiopia using 24 years of nationally representative data from the WHO Global Health Observatory. The findings show a gradual decline in stunting prevalence, with levels projected to reach 32.35% by 2029 and 31.95% by 2030. Although this indicates progress, the projected prevalence remains above Ethiopia’s national target of 19% by 2029 and the global SDG 2.2 aim of ending all forms of malnutrition by 2030. The projections indicate that, despite ongoing nutrition initiatives, Ethiopia is not on track to achieve its targeted reduction in under-five stunting under current conditions.
We recommend conducting a national-level study to assess whether current nutrition programs are on track, identify regional or local discrepancies, and inform interventions tailored to specific community needs. This approach will help ensure strategies are effective, context-specific, and capable of supporting progress toward achieving stunting reduction targets.
Supporting information
S1 Fig. Autocorrelation and partial autocorrelation plots of under-five stunting prevalence in Ethiopia.
https://doi.org/10.1371/journal.pone.0345000.s001
(TIF)
S2 Fig. STL (Seasonal-Trend Decomposition) of under-five stunting time series, showing observed, trend, seasonal, and residual components.
https://doi.org/10.1371/journal.pone.0345000.s002
(STIF)
S3 Fig. Residual statistics of forecasting models for under-five stunting, illustrating model fit and unexplained variability.
https://doi.org/10.1371/journal.pone.0345000.s003
(TIF)
References
- 1.
WHO. Malnutrition. https://www.who.int/news-room/fact-sheets/detail/malnutrition. 2023.
- 2.
WHO. Malnutrition in children. https://www.who.int/data/nutrition/nlis/info/malnutrition-in-children. 2025.
- 3.
WHO. Latest estimates. https://www.who.int/teams/nutrition-and-food-safety/monitoring-nutritional-status-and-food-safety-and-events/joint-child-malnutrition-estimates/latest-estimates?utm_source=chatgpt.com
- 4.
ICF EPHIEE a. Ethiopia Mini Demographic and Health Survey 2019: Final Report. 2021; Available from: https://dhsprogram.com/pubs/pdf/FR363/FR363.pdf
- 5.
Institue EPH. National food and nutrition strategy baseline survey: Key findings preliminary survey 2023. 2023. https://ephi.gov.et/wp-content/uploads/2023/03/FNS_baseline_survey_preliminary_findings.pdf
- 6. Abdilahi SA, Osman MO, Abate KH. Epidemiology of stunting in children aged 6-59 months, an unresolved chronic nutritional problem in Ethiopia: A systematic review and meta-analysis. SAGE Open Med. 2024;12. pmid:38881594
- 7. Candra A, Handini LS. The Relationship between Stunting and Body Resistance to Infection in Children. J Diverse Med Res. 2025;2(3):110–7.
- 8.
WHO. Child mortality (under 5 years). https://www.who.int/news-room/fact-sheets/detail/child-mortality-under-5-years#:~:text=Malnourished%20children%2C%20particularly%20those%20with,under%205%20years%20of%20age 2022.
- 9. De Sanctis V, Soliman A, Alaaraj N, Ahmed S, Alyafei F, Hamed N. Early and Long-term Consequences of Nutritional Stunting: From Childhood to Adulthood. Acta Biomed. 2021;92(1):e2021168. pmid:33682846
- 10. Moradi S, Mirzababaei A, Mohammadi H, Moosavian SP, Arab A, Jannat B, et al. Food insecurity and the risk of undernutrition complications among children and adolescents: A systematic review and meta-analysis. Nutrition. 2019;62:52–60. pmid:30852458
- 11. Budiputri GL, Gunawijaya E, Yantie NPVK, Ariawati K, Widnyana AANKP, Lastariana KAY. Association between Anemia and Stunting among Children with Congenital Heart Disease: A Retrospective Study. Academic and Clinical Pediatrics. 2025;2(2):35–9.
- 12. Dewey KG, Begum K. Long-term consequences of stunting in early life. Matern Child Nutr. 2011;7 Suppl 3(Suppl 3):5–18. pmid:21929633
- 13.
Union A. The cost of hunger in Africa: Social and economic impact of child undernutrition in Egypt, Ethiopia, Swaziland and Uganda background paper. Abuja, Nigeria. 2014.
- 14.
AGENCY AUD. Ethiopia. 2022.
- 15. Grosso G, Mateo A, Rangelov N, Buzeti T, Birt C. Nutrition in the context of the Sustainable Development Goals. Eur J Public Health. 2020;30(Suppl_1):i19–23. pmid:32391903
- 16.
Ethiopia, F. D. R. O. national nutrition program 2016-2020. 2020. https://www.nipn.ephi.gov.et/sites/default/files/2020-05/NNP2_pdf.pdf
- 17.
Health FM o. Seqota Declaration. https://www.moh.gov.et/en/initiatives-4-col/Seqota_Declaration?language_content_entity=en. 2024.
- 18. Kang Y, Kim S, Sinamo S, Christian P. Effectiveness of a community-based nutrition programme to improve child growth in rural Ethiopia: a cluster randomized trial. Matern Child Nutr. 2017;13(1):10.1111/mcn.12349. pmid:27549570
- 19. Ahmed KY, Ogbo FA, Tegegne TK, Dalton H, Arora A, Ross AG. Interventions to improve the nutritional status of children under 5 years in Ethiopia: a systematic review. Public Health Nutr. 2023;26(12):3147–61. pmid:37905557
- 20. Fikadu K, Yihune M, Boynito WG, Hailemariam Z. Exploring Multiple Barriers to Proper Child Feeding Practices in Rural Districts of Ethiopia. Food Sci Nutr. 2025;13(3):e4757. pmid:40008237
- 21. Fantay Gebru K, Mekonnen Haileselassie W, Haftom Temesgen A, Oumer Seid A, Afework Mulugeta B. Determinants of stunting among under-five children in Ethiopia: a multilevel mixed-effects analysis of 2016 Ethiopian demographic and health survey data. BMC Pediatr. 2019;19(1):176. pmid:31153381
- 22.
Lewiński M. Statistical forecasting methods and machine learning models in hierarchical forecasting for supply chain applications. 2024;286–95.
- 23. Tosepu R, Ningsi NY. Forecasting of diarrhea disease using ARIMA model in Kendari City, Southeast Sulawesi Province, Indonesia. Heliyon. 2024;10(22):e40247. pmid:39605824
- 24.
Morid M, Sheng O, Dunbar J. Time series prediction using deep learning methods in healthcare. 2021.
- 25. Greunen J. The prominence of stationarity in time series forecasting. Journal for Studies in Economics and Econometrics. 2014;38:1–16.
- 26. Adubisi OD, David IJ, James FE, Awa UE, Terna AJ. A predictive Autoregressive Integrated Moving Average (ARIMA) Model for forecasting inflation rates. Res J Bus Econ Manag. 2018;1(1):1–8.
- 27. Jain G, Mallick B. A study of time series models ARIMA and ETS. In: 2017. https://ssrn.com/abstract=2898968
- 28.
Mohammed A, Mohammed A. A survey on deep learning for time-series forecasting. 2021;:365–92.
- 29.
Alonge M, Isreal O. Time Series Forecasting with Deep Learning: New Frontiers in Predictive Analytics. 2025.
- 30. Qureshi M, Daniyal M, Tawiah K. Comparative Evaluation of the Multilayer Perceptron Approach with Conventional ARIMA in Modeling and Prediction of COVID-19 Daily Death Cases. J Healthc Eng. 2022;2022:4864920. pmid:36406332
- 31. Kaushik S, Choudhury A, Sheron PK, Dasgupta N, Natarajan S, Pickett LA, et al. AI in Healthcare: Time-Series Forecasting Using Statistical, Neural, and Ensemble Architectures. Front Big Data. 2020;3:4. pmid:33693379
- 32. Ahmed KY, Ogbo FA, Tegegne TK, Dalton H, Arora A, Ross AG. Interventions to improve the nutritional status of children under 5 years in Ethiopia: a systematic review. Public Health Nutr. 2023;26(12):3147–61. pmid:37905557
- 33. Assefa Y, Hill PS, Gilks CF, Admassu M, Tesfaye D, Van Damme W. Primary health care contributions to universal health coverage, Ethiopia. Bull World Health Organ. 2020;98(12):894-905A. pmid:33293750
- 34. Zelalem M. The seqota declaration: From proof of concept to expansion phase. UN Nutr. 2022;19:134.
- 35.
Odhiambo LK. Multi-sector programmes at the sub-national level: A case study of the Seqota Declaration in Naedir Adet and Ebinat woredas in Ethiopia. 2019.
- 36. Cerqueira V, Torgo L, Soares C. Machine learning vs statistical methods for time series forecasting: Size matters. arXiv. 2019.
- 37. Makridakis S, Spiliotis E, Assimakopoulos V. Statistical and Machine Learning forecasting methods: Concerns and ways forward. PLoS One. 2018;13(3):e0194889. pmid:29584784
- 38.
Anh N. Comparative study of classic vs deep learning models. 2024.
- 39. Wei W, Wang G, Tao X, Luo Q, Chen L, Bao X, et al. Time series prediction for the epidemic trends of monkeypox using the ARIMA, exponential smoothing, GM (1, 1) and LSTM deep learning methods. J Gen Virol. 2023;104(4). pmid:37022959
- 40.
WHO. SDG Target 2.2 Malnutrition. 2025.
- 41.
UNICEF. Nutrition, for every child: UNICEF nutrition strategy 2020–2030. UNICEF Publ. 2020.
- 42. Tizazu W, Laillou A, Hailu BA, Chitekwe S, Baye K. Complementary feeding and food-group level inequality among Ethiopian children 6-23 months of age (2011-2019). Matern Child Nutr. 2024;20 Suppl 5(Suppl 5):e13375. pmid:35599292
- 43. Simane B, Beyene H, Deressa W, Kumie A, Berhane K, Samet J. Review of Climate Change and Health in Ethiopia: Status and Gap Analysis. Ethiop J Health Dev. 2016;30(1 Spec Iss):28–41. pmid:28867919
- 44. Gebregziabher H, Kahsay A, Gebretsadik GG, Bereket T, Gebregziabher TN, Weldemichael MY, et al. Food crises and coping strategies in war-affected communities in Tigray, Ethiopia: A community-based cross-sectional study among households. BMC Nutr. 2025;11(1):82. pmid:40259413
- 45. Telila HF, Sima EA. Quantifying food insecurity in Ethiopia: Prevalence, drivers, and policy implications. Cogent Social Sciences. 2024;10(1).
- 46.
Nation U. The 17 Goals. https://sdgs.un.org/goals. Accessed 2023 October 1.
- 47.
Bachewe FN. The rising costs of nutritious foods in Ethiopia. 2017.
- 48. Abeshu MA, Lelisa A, Geleta B. Complementary feeding: review of recommendations, feeding practices, and adequacy of homemade complementary food preparations in developing countries–lessons from Ethiopia. Frontiers in Nutrition. 2016;3:41.