ABSTRACT
Background and Aims
Stunting is a height for age Z score falls bellow ‐2 standard deviation. Untreated stunted cases have lifelong consequences like cognitive development, increased risk of infection and long‐term health and economy burden. Although stunting remains highly prevalent in Ethiopia, there has been no prior attempt to develop an individualized risk prediction model. This study will develop and validates a predictive model to improve targeted intervention in Ethiopia.
Methods
Data from 2019 Mini Ethiopian Demographic Health Survey comprised of 2079 children's below 2 years. Data analysis was done using STATA version 17 and R version 4.4.1 software. Least absolute shrinkage and selection operator were used to select variables for Multilevel Multivariable Analysis. Nomogram was developed and model's performance was assessed through the area under the receiver operating characteristic curve and calibration plots. Bootstrapping techniques were applied to internally validate the accuracy of the model. Additionally, decision curve analysis was conducted to examine its clinical and public health applicability.
Results
The prevalence of stunting was 27.8% [95% CI: 24.96, 30.89]. The developed nomogram comprised 8 predictors: Maternal education, residence, sex a child, age of a child, Current feeding status, usage of bottle feeding, twin status and marital status. The area under the receiver operating characteristic curve of the original model was (AUC = 0.722, 95% CI; 0.698, 0.747) whereas the after bootstrap model produced prediction accuracy of an AUC of 0.719 (95% CI; 0.693, 0.744). Internal validation was performed using the bootstrapping method, demonstrating reasonably corrected discriminative ability. Decision curve analysis showed that the model provided a greater net benefit than strategies of treating all or none, particularly for threshold probabilities exceeding 19%.
Conclusion
This study developed and internally validated a predictive model for stunting in children under 2 years in Ethiopia, with strong discriminatory power (AUC 0.729) and calibration. The model, incorporating eight key predictors, offers a practical tool for clinical decision‐making through a user‐friendly nomogram.
Keywords: children, Ethiopia, prediction, risk score, stunting
Abbreviations
- ANC
anti‐natal care
- NPV
negative predictive value
- PPV
positive predictive value
- ROC
receiver operating curve
- WHO
World Health Organization
1. Background
Stunting is characterized by impaired linear growth in children, defined as a height‐for‐age Z‐score less than −2 standard deviations according to the World Health Organization (WHO) Child Growth Standards. Stunting, caused by persistent nutritional deprivation, has far‐reaching consequences that go beyond physical growth, including impaired cognitive development, poor intellectual performance, and lower adult economic production. In 2022, more than 149 million children worldwide were estimated to be stunted [1, 2].
The largest burden falls on South Asia and Sub‐Saharan Africa, which together account for more than 80% of stunted children [3, 4, 5, 6]. Stunting prevalence has decreased somewhat as a result of international initiatives, but not quickly enough to meet Sustainable Development Goal (SDG) 2.2, which aims to eradicate malnutrition by 2030 [7]. In Bangladesh, the prevalence of stunting was found to be highest (42.4%) among children aged 12 to < 24 months [8].
The prevalence of stunting among under‐5 children in Africa was 30.7% in 2020 surpassing the global average of 22.0% [9]. The prevalence is as high as 35.6% in under 2 years children in Malawi [10]. Stunting affects about 35% of children across Sub‐Saharan Africa, with East Africa contributing the largest share at 37% [11].
The overall national prevalence of stunting in Ethiopia varies from 47.9% in 2000% to 35.9% in 2019 [12]. Shows a decline in recent years but still reflects the persistent challenges in combating child malnutrition. Among younger children, particularly those under 2 years, the prevalence of stunting is even more alarming, with a cohort study revealing that 57.4% of children experience stunting by the age of two [13, 14].
Key determinants of stunting severity include the child's sex and age, maternal age, household wealth, maternal education, region, and community‐level maternal education [4, 6, 13, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25].
Recent studies have shed light on the importance of predictive modeling in tackling child malnutrition [21, 22, 26]. For instance, a machine learning model conducted in Sub‐Saharan Africa effectively pinpointed crucial factors contributing to stunting, such as maternal education and household sanitation, achieving an impressive AUC of 0.82. This really highlights the promise of data‐driven strategies in preventing malnutrition [3, 16, 27]. In a similar vein, another study utilized a nomogram‐based risk prediction tool in Malawi, demonstrating that identifying high‐risk children early on boosted intervention efficiency by 30% [21].
Poor sanitation, insufficient access to clean water, and inadequate nutritional intake have been identified as major contributors to the prevalence of stunting [27]. Socioeconomic factors, such as household income levels, maternal education, and access to sanitation, emerged as key determinants of stunting, with children from lower‐income households and those with restricted access to nutritious food being more susceptible to growth impairments [23, 24, 25]. Furthermore, maternal health and nutrition, including inadequate antenatal care and suboptimal breastfeeding practices, were critical factors in exacerbating stunting outcomes. These findings highlight the importance of addressing both maternal and child health, as well as improving living conditions, to reduce the burden of stunting [13, 28, 29].
Environmental factors such as poor sanitation, unsafe drinking water, and exposure to infectious diseases also play a significant role in undernutrition. Additionally, maternal factors, including inadequate antenatal care, poor nutrition during pregnancy, and insufficient breastfeeding practices, further increase the risk of stunting in children [30, 31].
Recent advancements in predictive modeling and data analytics provide an opportunity to address stunting through targeted interventions. Machine learning approaches have been increasingly utilized to predict undernutrition, offering insights into complex interactions between risk factors. In Ethiopia, predictive models have shown promise in identifying at‐risk children, enabling timely and cost‐effective interventions [30].
Despite ongoing efforts at both national and international levels, stunting rates remain high, with significant differences across various geographic and socioeconomic groups. Traditional methods for dealing with undernutrition often use past assessments and broad interventions. These approaches can slow down timely support or overlook those who need it most. Creating a predictive model allows for early, personalized risk assessment. This enables prompt intervention and better use of resources. By predicting which children are most likely to experience stunting, healthcare providers and policymakers can develop specific strategies to prevent malnutrition and its impacts.
A validated predictive model provides a clear view of what contributes to stunting. Such a tool can improve clinical decisions and public health planning. It helps health workers identify high‐risk individuals during care and aids policymakers in crafting targeted, evidence‐based programs. Additionally, predictive analytics can strengthen health monitoring systems and improve evaluation processes. It aligns with national priorities and global objectives, such as Sustainable Development Goal 2 (Zero Hunger) and the WHO Global Nutrition Targets 2025. Ultimately, this study aims to promote a shift from reactive to preventive strategies in child health. It offers a solid, practical tool for tackling chronic undernutrition in Ethiopia.
Conceptual framework (Figure 1).
Figure 1.
Conceptual framework of predictors of stunting in Ethiopia, developed from different literatures.
2. Methods
2.1. Study Area, Design and Period
This study was conducted in Ethiopia, the second most populous country in Africa, located in the northeastern part of Africa, between 3° and 15° north latitude and 33° and 48° east longitude [32]. It has an estimated total population of 132,060,000 and is ranked as the 11th most populous country in the world. Ethiopia covers an area of 1,000,000 km² (386,102 square miles), and the population density is 115/km². The area had thirteen regional states and two city administrations. A cross‐sectional study was conducted from October to December 2024 by using 2019 MEDHS.
2.2. Source and Study Population
All under 2 age children in Ethiopia found in randomly selected enumeration areas (EAs).
2.3. Data Source, Sample Size and Sampling Procedure
The DHS data set is publicly available for all registered users. The 2019 EDHS data were derived from the DHS Program website (https://www.dhsprogram.com/data/dataset_admin) based on online request and permission. The enumeration areas (EAs) of the 2007 Ethiopian population and housing census (PHC) served as the primary sample unit for the 2019 MEDHS, and households served as the secondary sampling unit. There were 305 EAs in the initial stage (93 urban and 212 rural) chosen at random from the full list of 149,093. EAs produced for the 2019 PHC sampling frame with a probability proportional to EAs size (PPS). In the second round, 30 households per cluster were chosen using a systematic, equal‐probability process from the freshly formed household listing. In 2019 EDHS, a total of 8,885 women aged 15–49 years were interviewed. Among them, 2079 (weighted) children fell within the age group of below 2 years.
2.4. Variable
Dependent variable: The outcome variable was stunting, categorized as ‘yes = 1’ for children with a height‐for‐age Z score below −2, and ‘no = 0’ for those with a Z score above −2.
Independent variable: Sociodemographic and maternal characteristics: Maternal education status, Residence, Maternal age at first birth, Sex of household head, Wealth, Religion, Place of delivery, Iron supplemented during pregnancy, Had ANC, Household family size and Region.
Children characteristics: Age of child, Sex of child, Twin status, Vitamin A supplemented (last 6 months), Child vaccinated, Currently breastfeeding, Uses bottle feeding, Time breast feeding initiated and Postnatal checkup for baby.
2.5. Data Processing and Analysis
Data cleaning, recoding, and labeling were employed to ensure the consistency of the data. Descriptive analysis was conducted and results were presented using tables, graphs, and texts. Sampling weight (v005/1000000) was employed to correct for variations in the probability of selection. R and STATA Version 17, was used for analysis.
For the multivariable prediction model development: The theoretical framework for predicting stunting at a future time point, t, is based on prognostic factors that are measured or identified at one or more time points before the stunting outcome, specifically at t0, which represents the point of prediction.
2.6. Model Creation and Verification
The Least Absolute Shrinkage and Selection Operator (LASSO) algorithm was employed to identify the most relevant predictors. This penalized regression method was chosen to enhance feature selection, minimize overfitting, and create a parsimonious predictive model for unfavorable outcomes. Variables considered for the multilevel multivariable prediction model was selected based on availability, biological plausibility, and clinical interpretability. The LASSO model will be tuned to an optimal shrinkage factor, using the minimum cross‐validation mean deviance to include predictors with non‐zero coefficients in the multivariable logistic regression [20, 22, 24]. A reduced model was constructed using variables whose p‐value was less than 0.05. This simplified risk prediction model was shown as a nomogram, and its calibration and discriminatory power were evaluated to define its performance. By computing c‐statistics, the discriminatory capacity of the simplified risk prediction model was evaluated. The c‐statistics may fall between 0.5 and 1 (perfect discrimination and no prediction capacity) [16, 17]. Swets's criteria, which include values ranging from 0.5 to 0.6 (bad), 0.6–0.7 (poor), 0.7–0.8 (acceptable), 0.8–0.9 (good), and 0.9–1.0 (outstanding), were also used to assess the built risk prediction model.
The model was tested using a Hosmer‐Lemeshow test and a calibration plot. A value of p > 0.05 for the model calibration test indicated a good model calibration. A prediction density map was also used to evaluate the model's prediction performance. For internal model validation, the original set was bootstrapped (with 1000 repetitions) to compute a relatively corrected C‐statistic (AUC). Children were categorized as having a high or low risk of stunting based on the Youden index's ideal cut‐off point. The purpose of Decision Curve Analysis (DCA) was to produce a risk prediction model that could be clinically interpreted.
3. Result
3.1. Socio‐Demographic and Maternal Characteristics
A total of 2079 (weighted) children who had complete information were involved in this study.
The majority 73.5% children were from rural areas and 46.66% of children were from uneducated mothers. The majority 54.35% of the children were born in a health institution and 43.24% were from a poor household. Regarding ANC 73.93% of the mother had ANC and 60.13 of the mother supplemented iron during pregnancy Table 1.
Table 1.
Baseline socio‐demographic and maternal characteristics for stunting, in Ethiopia 2019 MEDHS (n = 2079).
| Characteristics | Characteristics category | Frequency | Percentage (%) |
|---|---|---|---|
| Maternal education status | No education | 970 | 46.66 |
| Primary | 823 | 39.59 | |
| Secondary and above | 286 | 13.76 | |
| Residence | Urban | 551 | 26.50 |
| Rural | 1528 | 73.50 | |
| Maternal age at first birth | < 20 | 1278 | 61.47 |
| > = 20 | 801 | 38.53 | |
| Marital status | Married | 1971 | 94.81 |
| Not married | 108 | 5.19 | |
| Sex of household head | Male | 1804 | 86.77 |
| Female | 275 | 13.23 | |
| Wealth | Poor | 899 | 43.24 |
| Medium | 396 | 19.05 | |
| Rich | 784 | 37.71 | |
| Religion | Orthodox | 744 | 35.79 |
| Protestant | 552 | 26.55 | |
| Muslim | 745 | 35.83 | |
| Other | 38 | 1.83 | |
| Place of delivery | Home | 949 | 45.65 |
| Health institution | 1130 | 54.35 | |
| Iron supplemented during pregnancy | Yes | 1250 | 60.13 |
| No | 829 | 39.87 | |
| Had ANC | Yes | 1537 | 73.93 |
| No | 542 | 26.07 | |
| Household family size | < = 3 | 311 | 14.96 |
| > 3 | 1768 | 85.04 | |
| Region | Large central | 1812 | 87.16 |
| Small peripheral | 192 | 9.24 | |
| Metropolis | 75 | 3.61 |
3.2. Children Characteristics
The study consists of children aged 5–24 months (74.84%), with a nearly balanced distribution of stunting by sex with 51% male. The majority 66.62% of children haven't received Vitamin A in the last 6 months and only 29.15% were unvaccinated. Regarding the current breast feeding status of the children only 14.05% were not breast feeding currently and 79.17% were not using bottle feeding. Finally breastfeeding were initiated after 1 h in 12.94% of children and 87.06 children hadn't postnatal care (Table 2).
Table 2.
Children characteristics for stunting, in Ethiopia 2019 MEDHS (n = 2079).
| Characteristics | Characteristics category | Frequency | Percentage (%) |
|---|---|---|---|
| Age of child | < 6 months | 523 | 25.16 |
| 6–24 months | 1556 | 74.84 | |
| Sex of child | Male | 1062 | 51.08 |
| Female | 1013 | 48.73 | |
| Twin status | Single | 2059 | 99.04 |
| Multiple | 20 | 0.96 | |
| Vitamin A supplemented (las 6 month) | Yes | 694 | 33.38 |
| No | 1385 | 66.62 | |
| Child vaccinated | Yes | 1473 | 70.85 |
| No | 606 | 29.15 | |
| Currently breast feeding | Yes | 1787 | 85.95 |
| No | 292 | 14.05 | |
| Uses bottle feeding | Yes | 433 | 20.83 |
| No | 1646 | 79.17 | |
| Time breast feeding initiated | Immediately (< 1 h) | 1522 | 73.21 |
| > 1 h | 557 | 26.79 | |
| Postnatal checkup for baby | Yes | 269 | 12.94 |
| No | 1810 | 87.06 |
3.3. Prevalence of Stunting
The findings of this study revealed that the prevalence of stunting was 579(27.83% CI: 24.96, 30.89). The mean age of the child was 11.9 (+0.16) 95% CI: 11.67, 12.3) months (Figure 2).
Figure 2.
The Prevalence of stunting in under 2 year in Ethiopia, 2019 MEDHS.
3.4. Variable Selection and Model Diagnosis
A total of 26 models were generated by the LASSO regression using a 10‐fold cross‐validation selection technique. With a minimum cross‐validation mean deviation and an optimal penalty factor (lambda) of 0.0098262, the 21st model was determined to be the most efficient. Thirteen possible features (predictors) were chosen from the 21 co‐variants that were entered into the LASSO regression.
The multivariable analysis included the potential factors that were found and chosen by lasso regression. These factors included: child age, child sex, household head sex, residence, wealth, twin status, religion, place of delivery, vitamin A supplementation, iron administered during pregnancy, child vaccination status, current breastfeeding, history of bottle feeding, mother marital status, breastfeeding initiation time, household family size, postnatal checkup for baby, region, and mothers' educational status.
3.5. Development of an Individualized Risk Prediction Model
The identified potential variables selected by lasso regression were used to create an individual stunting risk prediction model based on multilevel multivariable binomial regression analysis. At the time of patient enrollment, the majority of the predictors are readily ascertainable. At a significant level of less than 0.05, each predictor's contribution was evaluated by removing it one at a time from the entire multivariable model (Table 3).
Table 3.
Multilevel multivariable logistic regression analysis and model reduction of stunting using potential predictors of stunting in Ethiopia 2019 MEDHS (n = 2079, weighted).
| Prognostic predictors selected by lasso algorithm | Original model | Reduced model | |
|---|---|---|---|
| Maternal education status | Coef (95% CI) | Coef (95% CI) | Risk score (Total = 27.4) |
| No education | 0.48 (0.05, 0.09)* | 0.54 (0.13, 0.95)** | 2.7 |
| Primary | 0.34 (−0.07, 0.75) | 0.39 (−0.01, 0.79) | |
| Secondary and above | 1 | ||
| Residence | |||
| Urban | 1 | 1 | |
| Rural | 0.67 (0.19, 1.14)*** | 0.93 (0.51, 1.37)*** | 2.5 |
| Wealth | |||
| Poor | 0.05 (−0.28, 0.41) | ||
| Medium | 0.27 (−0.08, 0.62) | ||
| Rich | 1 | ||
| Religion | |||
| Orthodox | 1.52 (−0.5, 0.6) | ||
| Protestant | 1.31 (0.25, 0.92) | ||
| Muslim | 0.98 (−13, 2.07) | ||
| Other | 1 | ||
| Place of delivery | |||
| Home | 0.23 (−0.03, 0.5) | ||
| Health institution | 1 | ||
| Region | |||
| Large central | 0.37 (−0.48, 1.22) | ||
| Small peripheral | 0.41 (−0.52, 1.35) | ||
| Metropolis | 1 | ||
| Sex of household head | |||
| Male | 1 | ||
| Female | −0.49 (−0.88, 0.1) | ||
| Age of child | |||
| < 6 months | −0.96 (−1.25, −0.67)*** | −0.96 (−1.25, −0.67)*** | 3.9 |
| 6–24 months | |||
| Sex of child | |||
| Male | 1 | 1 | |
| Female | −0.52 (−0.74, −0.30)*** | −0.53 (−0.75, −0.31)*** | 2.2 |
| Twin status | |||
| Single | 1 | 1 | |
| Multiple | 2.62 (1.45, 3.8)*** | 2.26 (1.51, 3.86)*** | 10 |
| Currently breast feeding | |||
| Yes | 1 | 1 | |
| No | 0.31 (0.01, 0.016) | 0.28 (0.01, 0.59) | 0.5 |
| Uses bottle feeding | |||
| Yes | −0.31 (−0.61, −0.01)** | −0.34 (−0.63, −0.04)*** | 1.6 |
| No | 1 | 1 | |
| Marital status | |||
| Married | 1 | 1 | |
| Not married | 0.98 (0.45, 1.48)*** | 0.82 (0.34, 1.3)*** | 3.5 |
Note: Coef. = coefficients, CI = confidence interval, ᵅ = Variables included in the final simplified model. p‐value denoted with ‘*’ = < 0.05, ‘**’ = < 0.01, ‘***’ = 0.001.
3.6. Nomogram of the Final Model
Maternal education, residence, child sex, age, current feeding status, bottle‐feeding usage, twin status, and marital status are among the predictors utilized in the nomogram's development. It would be simple to determine each patient's risk of stunting using the created nomogram. Children under the age of two with an uneducated mother, for example, have a comorbidity score of 2.7. The child has a score of 2.5 in this category and lives in a rural area. Child sex and age were the other co‐morbidity stunting instance; their scores in this category were 2.2 and 3.9, respectively. In contrast, the children who were born twins, were not now breastfed, were not bottle fed, and came from a married mother received scores of 10, 0.5, 1.6, and 3.5. Each predictor category's overall score was 27.4. According to the nomogram, the patient's risk of stunting with this total score is 0.96 (high risk) (Figure 3).
Figure 3.
Nomogram of stunting in under 2 years in Ethiopia 2019 MEDHS. X‐axis: Score refers to the value assigned to each predictor category; Total score represents the sum of all predictor scores for a given patient; Probability of unfavorable outcome indicates the risk level associated with the calculated total score.
3.7. Performance of the Nomogram Developed
Performance of the created nomogram was assessed using the calibration plot and discriminatory power. The original model's discriminating power was determined to be (AUC = 0.722, 95% CI: 0.698, 0.747) based on the area under the curve (AUC) of the receiver operating characteristics curve (ROC‐curve) (Figure 4).
Figure 4.
ROC curve of stunting prediction model of under 2 years children in Ethiopia, 2019 MEDHS (n = 2079). The x‐axis represents the False Positive Rate (1–Specificity) and the y‐axis represents the True Positive Rate (Sensitivity). The area under the curve (AUC = 72.2%) indicates the model's ability to distinguish between stunted and non‐stunted children. AUC = area under the curve, ROC = receiver operating characteristics.
The original model exhibits an accuracy (ACC) of 0.749 (a misclassification rate of 25.1%), sensitivity (S) of 0.176, specificity (SP) of 0.966, positive predictive value (PV+) of 0.664, and negative predictive value (PV−) of 0.756, all based on the default 0.5 cut off probability. Nonetheless, the model's accuracy (ACC) was 0.706 (95% CI: 0.68, 0.72), sensitivity (S) was 0.54, specificity (SP) was 0.76, positive predictive value (PV+) was 0.47, and negative predictive value (NPV−) was 0.81 based on the optimal cut of point (Youden index) cut off point 0.316 probability (Figure 5).
Figure 5.
Optimal cut of point (Youden index) cut off point of stunting of under 2 years children in Ethiopia, 2019 MEDHS (n = 2079). X‐axis = Probability threshold, y‐axis = Sensitivity/Specificity.
The degree to which the created nomogram categorized patients with stunting as “1” and those without stunting as “0” is displayed in the prediction density plot. According to the simplified multilevel multivariate model's density plot, 27.5% of the research participants had stunting (positive instances). Children at low risk of stunting are shown by the red graph, while children at high risk are represented by the blue one. At a narrow range of threshold probabilities, the graphic displayed great overlap in the model. As a result, the plot below displays the nomogram's forecast density in general, (Figure 6).
Figure 6.
Prediction density plot of stunting of under 2 years children in Ethiopia, 2019 MEDHS (n = 2079). X‐axis = Predicted probability of stunting and Y‐axis = Density.
3.8. Model Validation
Using the “mrs” package, the bootstrapping technique was used to limit overly optimistic results and prevent over‐interpretation. The results showed a relatively corrected discriminating power of 0.719 (95% CI: 0.693, 0.744). When implemented in external situations, the created nomogram's prediction power and the possibility of optimism‐induced overfitting are guaranteed by the verified model's low optimism coefficient of 0.003. The calibration curve is almost 45°, the model fitness test obtained a p‐value of 0.090, and the stunting prediction model is well calibrated, with predicted probability that closely match actual observed probabilities (Figure 7).
Figure 7.
Observed versus predicted of stunting of under 2 years children in Ethiopia, 2019 MEDHS (n = 2079). X‐axis = Predicted probability of stunting and Y‐axis = Observed/Actual probabilities of stunting (%).
3.9. Decision Curve Analysis (DCA)
In this study, we employed Decision Curve Analysis (DCA) to evaluate the clinical utility of our stunting prediction model in Ethiopia. DCA was chosen because it provides a practical framework for weighing the benefits and harms of using a predictive model to guide interventions across a range of threshold probabilities. The DCA plot demonstrates that using our model to guide interventions yields a higher standardized net benefit compared to the “treat all” or “treat none” strategies when the threshold probability exceeds 19%. This threshold represents the point at which the expected benefit of intervening (e.g., providing nutritional support or medical care) outweighs the associated costs or risks. For example, a 19% threshold implies that clinicians or policymakers would deem it justified to intervene if the predicted probability of stunting is at least 19%, balancing the trade‐offs between missing cases (false negatives) and over‐intervening (false positives).
In practical terms, this means our model is clinically valuable for decision‐making when the perceived cost of missing a stunting case is relatively high compared to the cost of unnecessary interventions. The results suggest that the model is particularly useful in resource‐limited settings like Ethiopia, where targeted interventions can optimize resource allocation and improve outcomes. By adopting this model, stakeholders can prioritize children at higher risk (probability > 19%), ensuring efficient use of limited resources while maximizing health benefits (Figure 8).
Figure 8.
Decision curve plot showing the net benefit of the developed model for carrying out a certain intervention measure in stunting of under 2 years children in Ethiopia, 2019 MEDHS (n = 2079). The x‐axis represents threshold probability (the risk level at which an intervention would be initiated), and the y‐axis represents net benefit.
3.10. Risk Classification Using a Nomogram
For practical reasons, the final simplified model was given as a nomogram. Based on the risk likelihood determined by the nomogram, patients are categorized as having a low or high risk of stunting consequences. The nomogram's risk probability estimate is too easy for any level of health professional to perform. Therefore, patients are categorized as having a low or high risk of stunting based on the cutoff (0.316) determined using the Youden index approach. Stunting rates were 18.4% in low‐risk (< 0.316) and 47.8% in high‐risk (≥ 0.316) groups, respectively (Table 4).
Table 4.
Risk stratification of stunting of under 2 years children in Ethiopia, 2019 MEDHS (n = 2079).
| Risk category | Frequency | Incidence of poor outcome |
|---|---|---|
| Low (< 10.5 score) | 1413 | 260 (18.4%) |
| High (> = 10.5 score) | 667 | 319 (47.8%) |
| Total | 2079 | 579 (27.8%) |
4. Discussion
This study aimed to develop an individualized risk prediction tool that helps predict the probability of developing stunting. Stunting is prevalent during childhood and intervention is fruitfully at this age [26]. Developing a risk prediction model for this highly vulnerable age group of the population plays a vital role in the prevention stunting and subsequent reduction of stunting. Ultimately this will contribute global target of reducing nutritional related problem by WHO [6].
In this study, the prevalence of stunting in under 2 year in Ethiopia was 27.83% (CI: 24.96, 30.89). Which was lower than previous study conducted in Ethiopia [17]. The possible justification for this difference can be due to variation in study period, sample sizes and the policy changes on improvement of maternal child health a cross a period in Ethiopia.
The prevalence observed in our study is higher than the study conducted in Kenya 23% [18]. The possible justification for this difference can be variation between socioeconomic conditions, health care infrastructure, feeding practice and nutritional program between two countries.
The risk prediction model incorporated 8 predictors that were identified for final reduced model. Performance of the developed model was robust, with an AUC of 0.722, indicating Acceptable discriminatory power. This result is consistent with similar studies in china children settings, where AUC values above 0.678 are regarded as accurate for clinical prediction models [24]
The predictive model for stunting using both the default 0.5 probability cutoff and the optimal cutoff point based on the Youden index (0.316) were considered. With the default cutoff of 0.5, the model achieves a high overall accuracy of 74.9% of cases are correctly classified. The model identifies 17.6% of positive cases (sensitivity) and 96.6% of non‐stunting cases (Specificity). In contrast, the Youden index cutoff point of 0.316 results in slightly lower overall accuracy (70.6%) but significantly higher sensitivity (54%). However, specificity drops to 76%. The choice to use which cutoff point depends the context. If the priority issue is identifying stunting cases using Youden index cutoff point gives more sensitivity than the default one.
The nomogram integrates key predictors, such as Maternal education, residence, sex a child, age of a child, Current feeding status, usage of bottle feeding, twin status and marital status. Allowing healthcare providers to classify children's s into high‐ or low‐risk groups based on a cut‐off score of 10.5 points (corresponding to a Youden index0.316 probability) provides a clear framework for clinical decision‐making.
For instance, a 1.5‐year‐old rural male with an uneducated mother would receive scores of 3.9 (age), 2.5 (residence), 2.7 (maternal education), and 2.2 (sex). If the child is not breastfed (0.5), not bottle‐fed (1.6), and has a married mother (3.5), the total score sums to 16.9. Plotting this on the nomogram reveals a 68% stunting risk, flagging the child for priority intervention. This quick scoring system enables efficient risk stratification in clinical or community settings.
Furthermore, decision curve analysis (DCA) indicated that the model adds significant clinical and public health benefits over default strategies, such as treating all or none of the Children. When the threshold probability exceeded 19%, the model demonstrated a higher net benefit. Internal validation using bootstrapping confirmed the model's stability, with a slightly corrected AUC of 0.719. The optimism coefficient of 0.003 suggests minimal overfitting, indicating that the model is likely to perform well in external settings
The internal validation may not fully account for variations in external populations. Future studies should focus on external validation in diverse clinical settings to confirm its generalizability. Moreover, the use of national secondary data has its own uncertainty in including all possible factors and recall bias.
5. Conclusion
This study successfully developed and validated a robust predictive model for stunting for under 2 years in Ethiopia, achieving Acceptable discriminatory power (AUC 0.729) and strong calibration. The model incorporates 8 significant predictors, such as Maternal education, residence, sex a child, age of a child, Current feeding status, usage of bottle feeding, twin status and marital status, providing a practical tool for clinical decision‐making through a user‐friendly nomogram. The Youden index cutoff (0.316) prioritizes sensitivity making each suitable based on the clinical context. Decision curve analysis further highlights the model's clinical utility, demonstrating a net benefit in guiding interventions for high‐risk Children. Despite its strengths, the study recommends external validation to enhance generalizability and suggests incorporating additional variables in future models to improve predictive accuracy. This tool holds promise for optimizing child care and resource allocation in low‐resource settings.
6. Limitations of the Study
A notable strength of this study is the use of large nationwide data and robust statistical methods, including LASSO regression and internal validation with bootstrapping, ensuring the reliability and generalizability of the predictive model. Additionally, the inclusion of easily ascertainable predictors enhances the practical applicability of the model.
Future research should focus on external validation in different settings to assess the model's applicability in diverse populations.
Author Contributions
Ahmed Fentaw Ahmed: conceptualization, methodology, writing – review and editing, writing – original draft, software, validation, visualization. Tewodros Yosef: conceptualization, writing – original draft, writing – review and editing, software, investigation, validation, supervision. Cherugeta Kebede Asfaw: conceptualization, writing – original draft, writing – review and editing, methodology. Eyob Girum Weldeyes: writing – original draft, writing – review and editing. Eskindir Melese Cherinet: writing – review and editing, writing – original draft, methodology. Mohamed Abdu Oumer: writing – review and editing, writing – original draft, software. Filimon Getaneh Assefa: writing – review and editing, writing – original draft, software, methodology. Tinsae Tesfaw Tadege: writing – original draft, writing – review and editing, visualization, formal analysis, data curation. Biniyam Mequanent Sileshi: writing – original draft, writing – review and editing, supervision, data curation, formal analysis. Eyob Getaneh Yimer: writing – original draft, writing – review and editing, visualization. Fuad Seid Ebrahim: writing – original draft, writing – review and editing, formal analysis. Bemnet Yazew Abegaz: writing – review and editing, writing – original draft, software. Kalaab Esubalew Sharew: writing – review and editing, writing – original draft, conceptualization, methodology, software, formal analysis.
Ethics Statement
Written consent was taken from DHS after justifying the reason for conduction this study. All methods were carried out in accordance with the relevant guidelines of the DHS program.
Consent
The authors have nothing to report.
Conflicts of Interest
The authors declare no conflicts of interest.
Transparency Statement
The lead author Ahmed Fentaw Ahmed affirms that this manuscript is an honest, accurate, and transparent account of the study being reported; that no important aspects of the study have been omitted; and that any discrepancies from the study as planned (and, if relevant, registered) have been explained.
Acknowledgments
All authors have read and approved the final version of the manuscript [Ahmed Fentaw Ahmed or MANUSCRIPT GUARANTOR] had full access to all of the data in this study and takes complete responsibility for the integrity of the data and the accuracy of the data analysis.
Ahmed A. F., Yosef T., Asfaw C. K., et al., “Development and Validation of a Predictive Model for Individual Risk Prediction of Stunting in Ethiopia: A Predictive Modeling Study,” Health Science Reports 8 (2025): 1‐13, 10.1002/hsr2.71335.
Data Availability Statement
The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.
References
- 1. World Health Organization , Malnutrition Fact Sheet (World Health Organization, 2022). [Google Scholar]
- 2. UNICEF , UNICEF. The State of the World's Children (Children, Food and Nutrition, 2023). [Google Scholar]
- 3. Khan M. N. A. and Yunus R. M., “A Hybrid Ensemble Approach to Accelerate the Classification Accuracy for Predicting Malnutrition Among Under‐Five Children in Sub‐Saharan African Countries,” Nutrition 108 (2023): 111947. [DOI] [PubMed] [Google Scholar]
- 4. Mardiyana E., Ambarwati R., and Shifaza F. J., “The Stunting Scorecard for Early Prevention: Development and External Validation of a Novel Tool for Predicting Stunting Risk in Children Under 5 Years of Age,” International Journal of Advanced Health Science and Technology 2, no. 3 (2022): 137–144. [Google Scholar]
- 5. Ngusie H. S., Enyew E. B., Walle A. D., et al., “Employing Machine Learning Techniques for Prediction of Micronutrient Supplementation Status During Pregnancy in East African Countries,” Scientific Reports 14, no. 1 (2024): 23827. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. World Health Organization ,Reducing Stunting in Children: Equity Considerations for Achieving the Global Nutrition Targets 2025 (World Health Organization, 2018). [Google Scholar]
- 7. United Nations , “Sustainable Development Goal 2: End Hunger, Achieve Food Security and Improved Nutrition, and Promote Sustainable Agriculture,” published 2023.
- 8. Chowdhury T. R., Chakrabarty S., Rakib M., Afrin S., Saltmarsh S., and Winn S., “Factors Associated With Stunting and Wasting in Children Under 2 Years in Bangladesh,” Heliyon 6, no. 9 (2020): e04849. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Global Nutrition Report (2023). Nutrition profiles of Africa. (2023).
- 10. Mkungudza J., Twabi H. S., and Manda S. O. M., “Development of a Diagnostic Predictive Model for Determining Child Stunting in Malawi: A Comparative Analysis of Variable Selection Approaches,” BMC Medical Research Methodology 24, no. 1 (2024): 175. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Takele B. A., Gezie L. D., and Alamneh T. S., “Pooled Prevalence of Stunting and Associated Factors Among Children Aged 6–59 Months in Sub‐Saharan Africa Countries: A Bayesian Multilevel Approach,” PLoS One 17, no. 10 (2022): e0275889. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Atalell K. A., Techane M. A., Terefe B., and Tamir T. T., “Mapping Stunted Children in Ethiopia Using Two Decades of Data Between 2000 and 2019. A Geospatial Analysis Through the Bayesian Approach,” Journal of Health, Population, and Nutrition 42, no. 1 (2023): 113. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Asgedom Y. S., Seifu B. L., Mare K. U., et al., “Levels of Stunting Associated Factors Among Under‐Five Children in Ethiopia: A Multi‐Level Ordinal Logistic Regression Analysis,” PLoS One 19, no. 1 (2024): e0296451. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Goddard F. G. B., Hunegnaw B. M., Luu J., et al., “Prevalence, Incidence, and Reversal Pattern of Childhood Stunting From Birth to Age 2 Years in Ethiopia,” JAMA Network Open 7, no. 1 (2024): e2352856. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Gebeye L. G., Dessie E. Y., and Yimam J. A., “Predictors of Micronutrient Deficiency Among Children Aged 6–23 Months in Ethiopia: A Machine Learning Approach,” Frontiers in Nutrition 10 (2024): 1277048. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Bitew F. H., Sparks C. S., and Nyarko S. H., “Machine Learning Algorithms for Predicting Undernutrition Among Under‐Five Children in Ethiopia,” Public Health Nutrition 25, no. 2 (2022): 269–280. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Goddard F., Hunegnaw B. M., Luu J., et al., “Prevalence, Incidence, and Reversal Pattern of Childhood Stunting From Birth to Age 2 Years in Ethiopia,” JAMA Network Open 7, no. 1 (2024): 2352856. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18. Guyatt H., Muiruri F., Mburu P., and Robins A., “Prevalence and Predictors of Underweight and Stunting Among Children Under 2 Years of Age in Eastern Kenya,” Public Health Nutrition 23, no. 9 (2020): 1599–1608. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19. Hanieh S., Braat S., Simpson J. A., et al., “The Stunting Tool for Early Prevention: Development and External Validation of a Novel Tool to Predict Risk of Stunting in Children at 3 Years of Age,” BMJ Global Health 4, no. 6 (2019): 001801. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20. Harrison E., Syed S., Ehsan L., et al., “Machine Learning Model Demonstrates Stunting At Birth and Systemic Inflammatory Biomarkers as Predictors of Subsequent Infant Growth–A Four‐Year Prospective Study,” BMC Pediatrics 20 (2020): 498. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21. Mkungudza J., Twabi H. S., and Manda S., “Development of a Diagnostic Predictive Model for Determining Child Stunting in Malawi: A Comparative Analysis of Variable Selection Approaches,” BMC Medical Research Methodology 24, no. 1 (2024): 175. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Ndagijimana S., Kabano I. H., Masabo E., and Ntaganda J. M., “Prediction of Stunting Among Under‐5 Children in Rwanda Using Machine Learning Techniques,” Journal of Preventive Medicine and Public Health = Yebang Uihakhoe Chi 56, no. 1 (2023): 41–49. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Teferi M., Hassen H., Kebede A., et al., “Prevalence of Stunting and Associated Factors Among Children Aged 06–59 Months in Southwest Ethiopia: A Cross‐Sectional Study,” 3 no. 4 (2016): 1–6.
- 24. Xiong Y., Hu X., Cao J., Shang L., and Niu B. J. F. P., “A Predictive Model for Stunting Among Children Under the Age of Three,” Frontiers in Pediatrics 12 (2024): 1441714. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25. Zemariam A. B., Abate B. B., Alamaw A. W., et al., “Prediction of Stunting and Its Socioeconomic Determinants Among Adolescent Girls in Ethiopia Using Machine Learning Algorithms,” PLoS One 20, no. 1 (2025): 0316452. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26. Swets J. A., “Measuring the Accuracy of Diagnostic Systems,” Science (New York, N.Y.) 240, no. 4857 (1988): 1285–1293. [DOI] [PubMed] [Google Scholar]
- 27. Begashaw G. B., Zewotir T., and Fenta H. M., “A Deep Learning Approach for Classifying and Predicting Children's Nutritional Status in Ethiopia Using LSTM‐FC Neural Networks,” BioData Mining 18, no. 1 (2025): 11. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28. Regassa R., Belachew T., Duguma M., and Tamiru D., “Factors Associated With Stunting in Under‐Five Children With Environmental Enteropathy in Slum Areas of Jimma Town, Ethiopia,” Frontiers in Nutrition 11 (2024): 1335961. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29. Simachew Y., Abebe A., Yoseph A., et al., “Growth Monitoring and Promotion Service Utilization and Its Associated Factors Among Mothers of Children Under Two Years in Ethiopia: A Systematic Review and Meta‐Analysis,” BMC Pediatrics 24, no. 1 (2024): 463. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30. Bitew F. H., Sparks C. S., and Nyarko S. H., “Machine Learning Algorithms for Predicting Undernutrition Among Under‐Five Children in Ethiopia,” Public Health Nutrition 25, no. 2 (2022): 269–280. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Kassaw A., Kassie Y. T., Kefale D., et al., “Pooled Prevalence and Its Determinants of Stunting Among Children During Their Critical Period in Ethiopia: A Systematic Review and Meta‐Analysis,” PLoS One 18, no. 11 (2023): e0294689. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Central Statistical Agency (CSA) [Ethiopia], The 2007 Population and Housing Census of Ethiopia: Federal Democratic Republic of Ethiopia, Population Census Commission (Central Statistical Agency, 2010).
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.