An institutional based retrospective follow-up study was conducted in hospitals found in the Amhara Region, from August 1, 2023 to August 2, 2024. In Amhara Region there is 100 public hospitals (eight comprehensively specialized, 20 generals, and 72 primary) served for 23, 558, 385 population. All hospital has separate room pediatric ward as used as a treatment center for SAM children. Health personnel follow an updated and a standardized form of treatment protocol of SAM guideline.

Records of all children age 0–59 years admitted stabilization center hospitals in Amhara Region were taken as the source population, whereas all under-five children admitted with cSAM to stabilization centers of in the selected hospitals considered as the study population.

All complicated under-five children with SAM admitted to therapeutic feeding unit (TFU) at, the selected hospital. Children who were admitted multiple times were only assessed for the most recent admission. However, incomplete records of under-5 children at therapeutic feeding program registration logbook with incomplete data (such as treatment outcomes, age, sex date of admission and discharge) were excluded from the study.

The minimum sample size that represents the source population was ensuring to consider the number of events per parameter (EPP) of ≥10, (32). By using the following formula EPP = n ϕ p , Where EPP is the number of events per parameter, n is the sample size, 𝜙 is the overall rate of the outcome and P is the number of parameters to be involved in the model development (32). Sample size estimation of the study followed the following assumptions using is the most frequently used conventional rule of thumb formula for minimizing the problem of overfitting in multivariable prediction modelling, Using EPP of ≥10, 21 parameters, and 𝜙 of (0.3787) which was taken from the previous studies (33) and design effect of (1.5). Therefore, the final adequate sample size considering a 10% non-response rate was 915, under-five children admitted to stabilization centers in selected hospitals within the Amhara Region.

A multistage stratified random sampling method was used to select children with SAM in hospitals across the Amhara Region. From a total of 28 general and comprehensive specialized hospitals, 9 were randomly selected. Additionally, 19 out of 69 primary hospitals were chosen, representing 28% of all hospitals in the region. The sample size was proportionally distributed among the selected hospitals and across different professional strata. Subsequently, a simple random sampling technique was applied using the serial numbers from the therapeutic feeding registration logbook of SAM children. The unique medical record numbers were then matched to the randomly selected serial numbers. Finally, the medical charts of the chosen children were retrieved from the card room for data collection.

The dependent variable was unfavorable treatment outcome (Yes/No) of under-5 children admitted for SAM. Meanwhile, the independent variables also included socio-demographic factors (age, sex of the child, place of residence), clinical conditions (loss of appetite), presence of nutritional oedema, co-morbidities (pneumonia, HIV status, diarrhoea, anaemia, malaria and tuberculosis), routine medication intake (intravenous (IV) antibiotic treatment, folic acid, vitamin A supplementation, deworming, Intake of F75 and F100) and level of anthropometric deficits at admission.

Unfavorable treatment outcome: The treatment outcome of under-5 children admitted for SAM being death, and/or treatment failure and/or non-respondent and/or defaulter, and/or staying longer time in the hospital (34).

Length of stay: The number of days the child stayed in the hospital from admission until death or recovered, (35).

Severity of oedema: was defined as grade + if involving only the feet, ++ if also involving the legs or hands, and +++ if also involving the area around the eyes (36, 37).

Complete record: if age in months, sex of the child, date of admission and discharge, type major complications, treatment outcome and other parameter well recorded.

Saver acute malnutrition (SAM): was defined as the presence of any bilateral pitting oedema (kwashiorkor), MUAC <11.5 cm (marasmus) or weight-for-height or length Z-scores < −3 (marasmus), or Marasmic-kwashiorkor (1, 38).

Complicated SAM: Children (6–59 months) who are acutely malnourished with associated medical complications and/or poor appetite; and infants less than 6 months with SAM, need to be treated in an inpatient care facility until they are well enough to continue nutritional rehabilitation (39).

Co-morbidities: In children with severe acute malnutrition, having TB and/or HIV and/or malaria and/or pneumonia, and/ or diarrhoea, and/or severe anaemia co-infection on admission to stabilization center.

Stabilization phase: children with cSAM are initially admitted to an inpatient facility for stabilization. In this phase: Life-threatening medical complications are treated, Routine drugs are given to correct specific deficiencies, Feeding with F-75 milk (low caloric and sodium) is begun (39).

Anaemia was defined with a haemoglobin level below 11 gm/dl (haematocrit level < 33%) at admission.

Severe Anaemia: If the haemoglobin concentration is less than 40 g/L or the packed–cell volume is <12% the child has very severe anaemia.

Rehabilitation Phase: Children that progress through phase 1 and transition phase enter phase 2 (rehabilitation phase) when they have a good appetite and no major medical complication. In this phase: Routine drugs are continued, Feeding with RUTF or F100 is started (40).

Cure from SAM: was achieved when weight-for height/length is ≥ − 2 Z-scores and they have had no oedema for at least 2 weeks, or MUAC is ≥12.5 cm, and they have had no oedema for at least 2 weeks without any acute medical complications (39).

Recovered: is defined as when the child is treated for acute medical complications at a stabilization center and transferred to OTP (not cured yet) for continuous SAM treatment.

A structured data extracting tool (checklist) was developed using different literature. Predictors of unfavorable treatment outcomes such as sociodemographic factors, clinical-related factors, medical-related factors, comorbidities, and anthropometric related factors were extracted. Therapeutic feeding program registration logbook and under-5 IMNCI registration were used for data extraction. The Data were collected by 6 trained Nurse and two master’s holder health professionals were recruited to manage the data collection processes. Data quality was managed by training and appropriate supervision of data collectors. Overall supervision was made by the supervisor and principal investigator. The collected data were checked for completeness, clarity, and accuracy. The quality checking was done daily after data collection and correction was made before the next data collection measures.

The data are collected through KoboCollect data collection toolbox, and then it was exported to STATA version 17 and RStudio version 4.3.3 statistical software for analysis. Variables with missing data was managed by multiple imputation techniques using the “mice” package in R-software assuming that the data were missed at random (MAR). The imputation process was performed in the whole dataset and 5 imputed datasets were generated. Sensitivity analysis was done to investigate the plausibility of the MAR assumption. Descriptive statistics, frequencies, and percentages were done for categorical variables. The normality distribution test was done using the Kolmogorov–Smirnov test. Incidence was calculated to determine the occurrence of unfavorable treatment outcomes.

For the multivariable prediction model development: The theoretical design of the incidence of unfavorable treatment outcome at a future time “t” is a function of prognostic determinants like demographic factors, clinical factors, and behavioral factors measured or ascertained at one or more time points before the occurrence of the unfavorable treatment outcome, i.e., “t₀” which is the moment of prognostication.

The Least Absolute Shrinkage and Selection Operator (LASSO) algorithm was used to select the most potent predictors. Penalized regression method was preferred for feature selection to develop unbiased and most parsimonies unfavorable treatment outcome risk prediction model by minimizing overfitting (41).

Variables were considered for multivariant unfavorable treatment outcome prediction model development based on their easily obtainability, biologically plausible relationship with the outcome, and ease of interpretation in clinical practice. The lasso model with optimum shrinkage factor and minimum cross-validation mean deviance was selected to take predictors with non-zero coefficients for multivariable logistic regression analysis.

The most potent predictors selected by LASSO regression were incorporated into the multivariable analysis. Then, variables were removed from the multivariable model step by step to build a simplified reduced model at a significance level of p-value<0.15 just to be more liberal. The final simplified risk prediction model was presented in the form of a nomogram and its performance was described by assessing its discriminatory power and calibration. The discriminatory power of the simplified risk prediction model was determined by calculating c-statistics. The c-statistics might range from 0.5 (no predictive ability) to 1 (perfect discrimination) (42, 43). The developed risk prediction model was also assessed qualitatively by using Swets’s criteria, which values range from 0.5–0.6 (bad), 0.6–0.7 (poor), 0.7–0.8 (satisfactory), 0.8–0.9 (good), and 0.9–1.0 (excellent) (44).

The calibration of the model was presented graphically using the calibration plot and Hosmer-Lemeshow test. For the model calibration test, a value of p > 0.05 suggested a good model calibration. The prediction performance of the model was also assessed by a prediction density plot. Bootstrap resampling (with 1,000 repetitions) of the original set was performed for internal model validation to calculate a relatively corrected C-statistics (AUC). Patients were classified as high and low risk of unfavorable treatment outcomes using the optimal cut-off point identified by the Youden index. The traditional risk prediction model performance measures (discrimination and calibration) could not address the issue of how useful the developed nomogram would be in clinical practice. Therefore, Decision Curve Analysis (DCA) was done to provide a clinically interpretable risk prediction model which could show the clinical and public health impact of using it. It was plotted by putting the net benefit of carrying out a specified intervention using the developed risk prediction model or intervention on all of the patients or none of them in the y-axis and the threshold probabilities in the x-axis. The study was reported per the TRIPOD (transparent reporting of a multivariable prediction model for individual prognosis or diagnosis) statement (45).

The majority 493 (57.1%) admitted SAM cases in stabilization centers were came from rural area. Similarly, 491(56.8%) admitted SAM under five children were male. The median (±IQR) age of a children were 11 (±14) years old. On the other hand, 238 (27.5%) of admitted SAM cases in stabilization centers were had oedematose. Beside of this the complete blood count (CBC) result verified that 239(27.7%) of SAM case were had Anemia. Adjacent to this 147(17%) of complicated SAM case were develop Pneumonia. From the total cases above 2/3^rd (86.5%) were marasmus clinical characteristics (Table 1).

The findings of this study revealed that the proportion of treatment failure or not-recovered, defaulter, death rates and staying longer in the hospital were 98(11.4%), 63(7.3%), 51(5.9%) and 28(3.2%) respectively. However, the overall incidence of unfavorable treatment outcome among the study participants was 27.8% (95% CI: 25, 31%) (Figure 1).

A total of 39 models were generated using LASSO regression with a 10-fold cross-validation selection method and LASSO estimator. The 34th model was found to be the most parsimonious model with an optimum penalty factor (lambda) of 0.008602 and minimum cross-validation mean deviance. Among 19 co-variants entered in LASSO regression, 12 potential features (predictors) were selected (Table 2).

The identified 10 potential predictors selected by lasso regression were incorporated into the multivariable analysis. These variables were; Oedematous, Diarrhoea, complete blood count (CBC) or Haemoglobin test result, Pneumonia, Folic acid supplementation, kidney function test, Vitamin-A supplementation, intake of F75, IV antibiotics treatment and intake of F100.

An individualized unfavorable treatment outcome risk prediction model was developed based on multivariable binomial regression analysis using the identified potential predictors selected by lasso regression. Most of the predictors are easily ascertainable right at patient enrolment time. The role of each predictor was assessed by reducing them one by one from the full multivariable model at a significant level of p-value <0.15 (Table 3).

Predictors used in the construction of the nomogram are; Baseline Oedematous, Diarrhoea, CBC test result, Pneumonia, Folic acid supplementation, Vitamin A supplementation and IV antibiotics treatment. The developed nomogram could be used to calculate the risk of individual patients for unfavorable treatment outcome easily. For instance; The risk of under-5 admitted SAM cases who had baseline oedema with co-morbidity of Pneumonia, Anemia, Diarrhea, without supplementation of folic acid, Vitamin-A supplementation and intravenous antibiotics as follows.

SAM under-5 children who had a baseline oedema and the score for this category of comorbidity is 6. The SAM case who had Pneumonia and the score for this category is 8.5. The other co-morbidity SAM case was Anemia and Diarrhea and its score for this category is 7 and 8, respectively. On the other hand, admitted SAM cases get folic acid and Vitamin-A supplementation and IV antibiotic treatment and its score for this category is zero. The total score of each predictors category was 29.5. The risk of the patient for unfavorable treatment outcome with this total score would be read from the nomogram and it is 0.94 (high risk) (Figure 2).

Based on the discriminatory power and calibration plot the developed nomogram performance was evaluated. The area under the curve (AUC) of receiver operating characteristics curve (ROC-curve) of the original model found to have a discriminatory power of (AUC = 0.913, 95% CI: 0.890, 0.935) (Figure 3). The prediction role of individual prognostic determinants; Baseline Oedematous (Yes), Diarrhoea (Yes), Pneumonia (Yes), CBC test result (presence of Anaemia), IV antibiotics treatment (No), Folic acid supplementation (No) and Vitamin-A supplementation (No) was assessed and it was found to be 0.699, 0.714, 0.674, 0.669, 0.658, 0.621 and 0.555, respectively.

The model fitness test had a p-value of 0.4449 the calibration curve is nearly 45 degrees, showing that there is no difference between predicted and the observed probabilities (Figure 4).

Based on the default 0.5 cut off probability, the original model has accuracy (ACC) of 0.872 (12.8% misclassification rate), sensitivity(S) 0.671, specificity (SP) 0.949, positive predictive value (PV+) 0.834, and negative predictive value (PV-) 0.882. However, based on the optimal cut of point (Youden index) cut off point 0.6955 probability, the model has accuracy (ACC) was 0.863 (95% CI: 0.839, 0.886), sensitivity(S) 0.813 [95% CI: 0.757, 0.859], specificity (SP) 0.883 [95% CI: 0.855, 0.907], positive predictive value (PV⁺) 0. 728 [95% CI: 0.676, 0.791] and negative predictive value (NPV^-) 0.925[95% CI: 0.898, 0.941], (Figure 5).

The prediction density plot shows the extent to which the developed nomogram classified patients with the unfavorable outcome as “1” and those without unfavorable outcome as “0.” The density plot of the reduced multivariate model indicated that 27.8% of the study subjects were with poor outcomes of SAM (positive cases). The graph with red one represents children with low risk of SAM, and the blue one was children at high risk of unfavorable treatment outcome of SAM. The plot showed some overlap in the model at a limited range of threshold probabilities. It is not 100% perfect, no model is 100% perfect. Therefore, the plot below shows the general insight of the prediction density of the nomogram, (Figure 6).

The model was validated internally using the bootstrapping method to avoid over-interpretation and minimize too optimistic results from the original model by using “mrs” package. Bootstrap method is preferred over split-half method and cross-validation for model more stable. It was performed by drawing bootstrap samples of 1,000 repetitions with replacement and was found to have a relatively corrected discriminatory power of 0.9086 (95% CI: 0.886, 0.929). The β coefficients from the bootstrapped model produced marginally the same results as the original β coefficients. The optimism coefficient for the validated model was 0.008, which is minimal and ensures the less likelihood of optimism induced over fitting of the developed nomogram and its prediction capability when it is applied in external settings. The finding is also supported by the model calibration test (p-value 0.009), indicates a very good agreement between predicted and observed probabilities; very slightly that the apparent curve seems to outperform the bias-corrected curve between 0.18 and 0.5 predicted probabilities. Therefore, given the limited optimism, and excellent calibration, the model might perform well in a new sample (Figure 7).

To predict an individual estimated risk for unfavorable treatment outcome of cSAM based on the identified predictors of validated regression coefficients as: the estimated probability of risk for unfavorable treatment outcome = 1/1 + exp-(1.68 + 1.401) * Baseline Oedematous (Yes) + 1.80* Diarrhoea (Yes) + (−1.58) * CBC test result (presence of Anaemia) +1.96* Pneumonia (Yes) + (−1.79) * Folic acid supplementation (No) + 0.36* Vitamin A supplementation (No) + (−2.03) * IV antibiotics treatment (No).

Regarding the decision to use our model, the decision curve outperforms the default strategies (referring all and none) across the entire range of threshold probabilities. The purple line represents the developed risk prediction nomogram, the thin black line represents the assumption that all patients are at risk of unfavorable treatment outcomes and the thick black line represents that none of the patients are at risk of unfavorable treatment outcomes.

The net benefit was calculated by subtracting the proportion of all patients who are false positive from the proportion who are true positive, weighting by the relative harm of not taking an intervention compared with the negative consequences of unnecessary intervention. The relative harm was calculated as; ( pt 1 − pt ) (46). Here, “p_t” stands for the threshold probability, where the expected benefit of a certain intervention is equal to the expected benefit of avoiding an intervention. The DCA plot shows, the net benefit of using the model to carry out a certain intervention, which could be determined according to the status of the clinical setting, to tackle the unfavorable treatment outcome in admitted SAM cases at stabilization center among under-5 children was found to be higher than intervening on all or none of the SAM cases.

The decision curve showed that if the threshold probability is greater than 21%, using the developed risk prediction nomogram in this study to predict unfavorable treatment outcomes in admitted SAM cases at stabilization center among under-5 children adds more benefit than the intervention on all or none of the patients’ strategies. This implies that our model has the highest clinical and public health importance. Therefore, decisions made using the model such as safely discharging children with some medications or keeping children for more intensive care in the hospitals has a higher net benefit (Figure 8).

The final simplified model was presented in the form of a nomogram just for practical utility. Patients are classified as at low, and high risk of unfavorable treatment outcomes based on the risk probability identified using the nomogram. The risk probability calculation using the nomogram is too simple that any health professional at any level can do. Hence, using the cutoff (0.6955) identified by the Youden index method, patients are classified as at low, and high risk of unfavorable treatment outcomes. The proportion of unfavorable treatment outcome in low (<0.6955), and high-risk groups (≥0.6955) were; 2.6 and 89.4%, respectively. On the other way for simple interpretation in the clinical settings, we categorized risk scores into less than twenty-one points (< 21) (low-risk group), and greater or equal to twenty-one points (≥ 21) points (high-risk group) based on Youden index (optimal cut-off point) which corresponds to the probability of 0. 6,955 in the model. Therefore, a child can have a minimum and maximum risk score of 0 and 33, respectively. The incidence of high-risk groups of unfavorable treatment outcome of SAM cases were 170(19.7%) and the rest 694(80.3%) cases were categorized in the low-risk group, (Table 4).

These study implies that using the model to admit high-risk children in stabilization center yields a greater net benefit than not admitting/providing specialized care for all SAM children, regardless of risk thresholds.