The Divina Providencia Hospital in Luanda is located in the Kilamba Kiaxi Municipality, a peri-urban area of the city. Over the years, the hospital has emerged as a city-wide benchmark in the treatment of acute malnutrition, for both severe and moderate cases. Following the Angolan National Protocol for the Treatment of Acute Malnutrition, children with SAM are hospitalized to stabilize their clinical conditions. Subsequently, the children are transferred to the OTP where their nutritional status is periodically reassessed while receiving highly specialized ready-to-use therapeutic food. In this study, two different sources of data were analyzed. The registers data gather information of children hospitalized with SAM. A total of 1,699 children with SAMwere hospitalized at the Divina Providencia Hospital of Luanda between January 2019 and June 2021. After data cleaning, the final number of children included in the analysis is 1,259.¹ The dataset is periodically compiled by the clerks of the hospitals and comprises 15 variables: the gender of the child, the anthropometric measurements on both the day of admission and discharge including weight, height and middle-upper arm circumeference (MUAC), the rate of weight gain and the gain in millimeters of the MUAC, the treatment duration, the outcome (dead / survived), the presence of edema and the area of residency of the family. Due to the hospital reporting requirements, the age of the children was aggregated in age groups: < 6 months, from 6 months to 59 months, and > 60 months. The main limitation of this data set is that there was no record of nutritional therapy administered to each child, nor if there were comorbidities. The average weight gain rate for each child was calculated with the following formula:

Since the registers did not include information about the socio-economic conditions of families, a survey was implemented to show how deprivations and lack of capabilities relate to the well-being and food consumption of families who referred to the hospital in August 2021 to seek hospital services. A sample of 84 families whose children were in the outpatient treatment program for malnutrition were interviewed in August 2021.

The hospital board granted formal authorization to conduct the survey. Data were collected through interviews conducted in the hospital’s nutrition ward during scheduled nutritional follow-up visits for children. Caregivers were recruited using a convenience sampling strategy based on their availability and willingness to participate and an original questionnaire was redacted and used to the scope of answering to the research question. Prior to each interview, the scope of the research was presented to the caregivers both in Portuguese and in local language. Also it was clearly specified that participation in the study was voluntary and that it would not affect the nutritional service, nor their right to receive food and the entitlements provided by the hospital. Verbal consent was asked to ensure caregivers with limited literacy or writing skills. A total of 84 families provided verbal consent to participate in the reasearch. The original questionnaire included quantitative and qualitative questions about SEC, anthropometric measurements of children, and weekly food consumption at the household level. By its nature, the capability approach requires a large informational basis to enable a comprehensive analysis of direct and sometimes hidden drivers of food security. Hence the reason of the authors to include in the questionnaire a range of multidimensional variables including education, literacy, age, family size, assets ownership, religious affiliations, income, health conditions, employment, nutritional, livelihoods and head of the household sex. This approach allows the study of food security to move beyond the “mechanicistic” perspectives of Income-based or the food availability approaches, to a broader area of human development studies. As a proxy indicator of food security we used an index of food consumption, computed by multiplying the frequency of weekly consumption of each food group by scores that underline the differences in the bromatological composition of food groups and then summed together to obtain a single value for each household (43). The indicator measures the food consumption of families in terms of quantity and quality. It has been used to assess food access (44) and has been validated to conduct in-depth food security assessments elsewhere (45). Our research is based on three main assumptions. First, food consumption is intended as a proxy measure for food security at the household level assuming intra-household distribution equity; second, we asked about food consumption at the household level assuming it to be positively associated with child food consumption (46). Third, we assumed that diet appropriateness is positively associated with the socioeconomic condition of households (47, 48).

We used classification and regression trees (CART). Classification trees use recursive binary splitting based on the variables that have the least level of impurity (weighted Gini Index) in assigning samples to the right terminal node of the tree. Weighted Gini impurities are calculated for all independent variables: those having the least value of impurity represent the direct follower of the root.

Regression trees, instead, estimate the mean value of a quantitatively dependent outcome variable, given multiple independent factors. As well as classification trees, regression trees are built in a top-down direction. For choosing the root variable, each independent factor is used singularly to estimate the outcome variable through the process of least squares residuals. The variable that results in the minimum sum of squared residuals (RSS) is chosen as the top variable of the three. The same procedure is generated for the following splits. The tree stops when each leaf represents a minimum number of observations established a priori.

CART analysis usually produces models that are easy to analyze due to their tree-shaped output. Other studies have applied CART and regression analysis in malnutrition studies. Yin et al. have applied CART to develop a model for classifying the severity of malnutrition in cancer patients and accelerating pretreatment classification to achieve higher survival rates (49). CART are also likely to overfit the data due to extra ramifications. To reduce bias, fully grown trees have been pruned by controlling for the cost complexity parameter that minimizes the error rate of the prediction of each tree (50).

Climate conditions in Luanda must be considered when studying malnutrition. In fact, The weather in Luanda is dry from May to September and rainy from October to April (51) and Figure 3 shows that the level of malnutrition mortality in the Divina Providencia hospital peaked during the dry season in both 2019 and 2020, thus providing a connection between household food security (possibly due to the lack of water for agricultural use) and higher levels of malnutrition mortality. In 2020, the hospital did not accept new admissions for several weeks from February to March due to Covid-19 restrictions. However, children with SAM who were admitted to the hospital prior to February were allowed to continue treatment within the hospital. This partially explains why in 2020 the mortality peak was reached in the month of September, when families were finally allowed to seek nutritional help at the hospital.

Since the survey was implemented in August, during the dry season, analyzing how deprivations affect households in different areas is of paramount importance. To the scope, we applied CART analysis to our survey data. The hospital admits children from five main Municipalities namely Belas, Luanda, Kilamba Kiaxi, Talatona and Viana. These geographic areas were set as the dependent variable and all SECs as independent variables. SEC includes information about both the mother of the child and the head of the household, including age, education, literacy, and employment; it also includes information about the size of the family, the rooms available, the type of material for construction, income, access to water and electricity, the cooking fuel, the number of meals per day, the food expenditure, the place from where they usually buy the food, the month of interruption of breastfeeding of the malnourished child, the index of food consumption, and a series of assets owned by the family such as the fridge, the cell phone, the bicycle, the television, and the radio.

By sampling, all the families interviewed shared the problem of child malnutrition. Figure 4 classifies families according to household characteristics and deprivations that they more frequently experience in the municipality where they live. Each node also provides the estimated percentage of the families who belong to each municipality in each split. The root node (the box at the top of the tree) provides the approximated percentages of families sampled per municipality: Belas (23%), Cacuaco (1%), Kilamba Kiaxi (43%), Luanda (5%), Talatona (12%) and Viana (17%). Thus, the majority of families tend to refer to the hospital from Kilamba Kiaxi, the municipality where the hospital is located. At each split the tree provides different percentages referring to each Municipality. For example, the first split identified the lack of a gas system as the best classifier of the entire survey sample. In this case, families living in Viana seem to be affected the most with a probability of 70% (gray box). The second split only applies to families that have access to the gas system. Families from Belas have a 44% probability of having a room/person ratio smaller than 0.45, while families from Kilamba Kiaxi have apparently better housing conditions. Furthermore, in Luanda, families seem to live in similar housing conditions as families in Belas, while they tend to have heads of household of different age groups. On the other hand, Kilamba Kiaxi families have lower gas systems and housing problems, but families living there have a 79% probability of experiencing anxiety in relation to the availability of water. The lack of a gas system could affect the ability to cook safely for entire communities. In fact, relying on substitutive fuels for cooking, such as wood or charcoal, can cause major health problems such as burnt and respiratory diseases (52). Furthermore, the price volatility of such materials can undermine the ability of families to prepare enough food (53). Families having limited access to fuel may exclude long-cooking ingredients from their diet and therefore limit their weekly food intake (54, 55). In Luanda, close to three quarters of urban citizens live in informal settlements called ‘musseques’ where house overcrowding is a spreading phenomenon (56), and can double the probability of cutting down the size of children’s meals (57). Furthermore, mental health markers such as anxiety and depression are known to increase when families experience water insecurity (58). Studies highlight that despite recent government actions, only 50.3% of the peri-urban population attain the minimum 40 liters per day of water consumption (59), which implies serious difficulties in food preparation. On the one hand, the informal system of water provision was seen as able to provide water to areas that are not secured by the formal system. This implies that quality controls are rarely or never conducted. Water quality has been associated with the suffering of environmental enteric dysfunction (EED), leading to intestinal inflammation, malabsorption, and therefore malnutrition, especially in children of overcrowded households (60). Furthermore, Mkupete et al. showed that unequal access to water and provision implied unequal child growth levels in Tanzania (61).

The CART in Figure 5 provides an analysis of the issue of water security in Luanda. It shows that income, buying tanks/cisterns of water (which in Luanda implies depending on the informal water system), mother education, and inaccessibility to the electrical systems are the capabilities that affect to the probability of households experiencing worry about water unavailability. Families that earn less than 3,375 Kwanzas (7.5 $) per week (11% of the sample) are shown to experience anxiety in relation to water, due to lack of purchasing power even against the informal market. Additionally, if families that earn more than 3,375 Kwanzas per week are required to rely on the informal water supply system and, at the same time, have no access to electricity (17% of the sample) they will experience anxiety in relation to water. On the other hand, families who can rely on tap water or public wells or buy bottles of water experience significantly lower stress in relation to water availability. Another study highlights the mechanisms of interdependency between water and energy security in determining FS (62).

To study the relationship between household characteristics and level of food consumption, we built a regression tree (Figure 6) setting the index of food consumption as the dependent variable and the SEC variables as independent variables. The index of food consumption ranges from 0 to 15 indicating increasing diet appropriateness.

Each box shows the change in the average index of food consumption of families depending on SEC and gives information on the number and percentage of families that are included in each split. Depending on the number of meals to which families have access (lower or greater than 2.5 per day), the score decreases to 5.5 or increases to 7.6. Therefore, Figure 6 highlights that the daily frequency of meals is a proxy for the weekly food consumption score. In addition, family income and maternal education resulted as factors that further differentiate the levels of household food consumption. In particular, families earning less than 30.000 Kwanzas are classified as having the lowest level of FS (4.9), while a level of mother education greater than 9.5 years implies the highest FS score (9.2). These results are in line with other studies in the literature. Several studies consider; several studies consider 3 meals per day as a conventional threshold of FS both in Zambia and Kenya (53, 63). Furthermore, other scholars found that the ability of families to purchase food was related to the lack of jobs and income in urban areas in Yemen (64). The positive relationship between maternal education and the levels of food consumption detected here has also been highlighted in other studies in which educated women from Burkina Faso, Mali, Niger and Chad had fewer children (65) and ultimately experienced fewer malnutrition events (66). Berlie et al. found that the educational level of the household is a strong determinant of FS in Ethiopia (6) while specifically through CART, the level of schooling of parents was found to be a strong indicator in classifying the nutritional status of children in Bangladesh and Cambodia (67, 68). We must recognize that all the households interviewed experienced child malnutrition, thus we cannot infer the causes of it.

In this section, we connect the two data sources used in this study. Since the hospital register data did not contain information on the socioeconomic characteristics of families, we assume that temporary anthropometric wasting of the hospitalized children reflects the multidimensional poverty mechanisms as discussed in the earlier section. Indeed the findings from the survey data framed the inability to access enough meals per day, the inability to access to enough income and mother education as main deprivations that affected food security in the sample of families interviewed. Other studies have highlighted how malnutrition indicators can gage food system policies’ capacities to ensure the ability of a society to be food secure, avoiding the need for intricated monetary conversions (69). Also, the mere existence of hospital facilities in Luanda collecting data on the survival of malnourished children should already sound alarming in terms of urban food system dysfunctionality. However, we find it useful to show how data can shed light on nutritional patterns. Indeed, data-driven methods might reveal patterns that can complement policy formulation (70) including those aimed at eradicating malnutrition. In this case, we use a regression tree to explore how the severity of malnutrition indicators at hospital admission influence children’s ability to gain weight during nutritional treatment and consequently the survival rate. We set the AGW as the dependent variable while the anthropometry of children at admission to the hospital including weight, height, MUAC, together with the treatment outcome (survived/dead), gender, age, and the Municipality of provenance were set as independent variables.

The final leaves of the regression tree in Figure 7 illustrate different values of AGW during hospitalization, depending on the anthropometry of the children at admission. In particular, it estimates that the children who died (13% of the average estimated) weighed between 2.4 and 7.4 kg and accounted for a loss of weight averaging −3.5 g / kg / day during hospitalization. On the other hand, survivors tended to gain at least 2 g / kg / day.² From the tree, the AGW of children varies greatly depending on the MUAC at admission, highlighting that a lower MUAC implies a lower gain weight during an average length of stay (LOS) of ~8 days (not shown). In another study conducted in Cunene (Angola), the AGW of malnourished children varied from 0 to 20 g/kg/day with a LOS of 7.9 days and a recorded death rate of 11.4% (71) highlighting a continuum among Angolan hospitals. In Egypt, the mean AGW was detected to be 10.4 g/kg/day with a LOS of 15.47 days (41), while in Ghana a weight gain per diem of 28 g / kg was measured during a LOS of 8 days (72). Differences in AGW could be due to varying clinical conditions at admission to the hospital. HIV, EED, diarrhea, or pneumonia might be the primary cause of weight loss (73).

In our data, no other symptoms other than edema were recorded in the registers. Nevertheless, the survival curve in Figure 8 shows that edema is a significant risk factor for mortality (74). Trends show different survival rates depending on the persistence of edema during each day of hospitalization. Unhealed edema on day 8 of hospitalization reduces survival probabilities to <50%. The exact biological pathway between mortality and complicated SAM has not been established and was not the objective of this study. However, it is relevant to note that prolonged starvation is associated with a wide range of metabolic changes (24) that are difficult to measure in daily practice in hospitals in developing countries.

After hospitalization, the children begin the outpatient therapeutic care program (OTP). To identify household-level deprivations that may extend the length of OTP treatment and hinder timely recovery, the regression tree in Figure 9 was built with the duration of the treatment as the dependent variable while all SEC were set as independent variables. The tree shows that the levels of schooling, myths and taboos about food, and the literacy of the mother are determine separate degrees of duration of the treatment. The partition showed in Figure 7 highlights that maternal education is the most influential independent variable associated with the duration of the OTP treatment. Children whose mothers have more than 4.5 years of schooling tend to have a shorter treatment duration, averaging 2.3 months, compared to 5.7 months for children whose mothers have less than 4.5 years of schooling. In the group of children whose mothers had better access to school, the absence of cultural / religious beliefs is another influencial determinant of the duration of OTP of their children. In this case, the average treatemt varies from 2 months when families do not have any belief that might influence food consumption to 3.8 months on average when the families do belong to religious group who predicate abstinence from certain foods. Furtherly, when the mother lacks the ability to read and write the average duration of treatment increase from 1.1 months on average to 3.3 months on average. This analyses shows that deprivations such as low mother education and literacy as well as myths and taboos have a layered influence in determining the ability of the child to recover from malnutrition during OTP. This analysis also speculate that the nexus between mother education and the duration of treatment might be similar to the link highlighted in Figure 6 that connect slow mother education to low food consumption at the household level. Indeed, mother illiteracy has been associated with children underweight (75), stunting (76), worse Infant and Young Child Feeding Practices (77) and as one of the most important factors in combating malnutrition in at least three countries such as Ghana, Nigeria and Ivory Coast (78). Likewise, we argue that food restrictions imposed by religious or traditional beliefs may undermine children’s ability to recover, thereby extending the duration of treatment (79).