IntentionThe research aimed to establish the key determinants of food insecurity inside the Arab area during the COVID-19 pandemic via a novel strategy, ML algorithms.SettingThe “Corona Cooking Survey April 2020” [35] performed by 38 completely different countries during the COVID -19 pandemic was used for this research. The survey was distributed to over 300 individuals per nation in every of the following Arab countries: (Palestine, Lebanon, Jordan, Kuwait, Oman, Qatar, Saudi Arabia, United Arab Emirates, and Bahrain. The survey knowledge was collected by the analysis workforce between April seventeenth and June twenty fifth, 2020.DesignData supplyThe dataset was extracted from the worldwide “Corona Cooking Survey April 2020” [29]. The survey was designed as a multi-language survey together with the Arabic language to facilitate knowledge assortment among completely different countries. The knowledge assortment instrument assessed the impact of the COVID-19 lockdown on adults’ well being and vitamin. The survey included a number of varieties of info, akin to COVID-19 Lockdown Measures, COVID-19 Feelings (Kessler K6 scale), Food Literacy Scale, Shopping Experiences and Behavior, Cooking Behavior and Attitudes, Seeking Recipes & Food Content, Eating Behaviors (Food Frequency Questionnaire), Food Advice Sources, E-drinking and E-dining, anthropometric measurements, and life-style and consuming habits. The survey was distributed via completely different social media platforms and thru the associate universities’ networks. The last pattern included over 300 individuals per nation. Countries that acquired a lesser quantity of samples have been excluded from the evaluation. The knowledge referring to food insecurity in Arab countries have been extracted from the International Survey. Overall, 13,443 individuals aged over 18 years have been included in the ML mannequin improvement.Model optionsThe options have been extracted from the most important research variables, accessible through https://osf.io/nz9xf/files/ [29]. The research options comprise respondents’ knowledge from earlier than and after the COVID-19 lockdown. The ML fashions used food insecurity as the most important goal variable for assessing the efficiency of ML in predicting and figuring out associated risk factors. Food insecurity was outlined in reference to the Food Consumption Score (FCS), which is an indicator used to evaluate dietary variety earlier than and after the pandemic. The FCS was categorized into two teams: Low FCS if the FCS < 42 (Unacceptable), and high FCS score if the FCS scores > = 42 (Acceptable). Thus, food insecurity was decided based mostly on the FCS classification. The low FCS group was thought of a food insecure group, whereas the excessive FCS group was thought of a traditional group. Detailed details about the research variables’ definition and calculation might be discovered in Hoteit et al. [29].The research used interconnected nature of social determinants of well being, scientific options, and food insecurity. By understanding these relationships, the research makes an attempt to develop an early detection system for food insecurity utilizing ML fashions.The social determinants of well being, which embody factors like socioeconomic standing, schooling, and bodily surroundings, among others, have been well-documented to affect entry to adequate and nutritious food [9, 31]. For occasion, low-income households would possibly battle to afford wholesome meals, contributing to food insecurity. On the opposite, communities with food insecurity could face restricted instructional and financial alternatives, thus making a suggestions loop. By figuring out such patterns, the ML mannequin can predict potential food insecurity based mostly on present social determinants of well being [32, 36]. As for scientific options, factors akin to physique mass index (BMI) or presence of sure well being circumstances could present precious insights into a person’s dietary standing, which is intrinsically linked to food safety [10]. For occasion, an unusually low BMI would possibly counsel malnutrition because of insufficient food entry, a doable signal of food insecurity [37]. Meanwhile, sure ailments could impression an individual’s capability to acquire or eat nutritious food, thereby exacerbating food insecurity.In our research the scientific options and social determinants of well being have been used as predictors of food insecurity, the existence of these factors gives info that a person or a neighborhood is at risk. Therefore, by coaching ML fashions to acknowledge these options, we will probably establish food insecurity early, even earlier than extra extreme signs manifest. This allows well timed interventions to alleviate food insecurity and stop its dangerous results on bodily and psychological well being.Data preprocessingData underwent a preprocessing process previous to constructing the ML fashions. The preprocessing part included knowledge cleansing, formatting, lacking knowledge remedy, and knowledge categorization. The knowledge cleansing course of included the null worth knowledge, the text-to-numeric conversion, and the lacking knowledge remedy.In the knowledge set, imbalanced knowledge was encountered as 1529 individuals have been categorized as food insecure, and 11,914 individuals have been categorized as food safe. An imbalanced knowledge set would possibly bias the ML mannequin’s estimation by offering extra weight to the dominant class [38]. The easy and efficient Min–Max Normalization method was used to scale options to a typical vary to make sure that all the options have been on the similar scale, and to permit the mannequin to make extra correct predictions. The minimal and most values for every characteristic have been first calculated, adopted by subtracting the minimal worth from all characteristic values and dividing the ensuing tables by the vary (i.e., the distinction between the most and minimal values) to make sure that all the options have been scaled between 0 and 1.ML fashions descriptionSeven ML fashions have been used in this research to evaluate the efficiency of ML in predicting food insecurity among Arab countries following the COVID-19 pandemic. Logistic Regression (LR), Gradient Boosting (GB), Support Vector Machine (SVM), Random Forest (RF), Artificial Neural Network (ANN), Naïve Bayes (NB), and k-nearest neighbors’ algorithm (k-NN) have been constructed and evaluated contemplating their efficiency measures.
Logistic Regression (LR)
Logistic regression serves as a supervised machine learning algorithm for classification, facilitating the prediction of the probability of an final result variable [23]. This final result, or dependent variable, inherently has a binary nature, signifying that solely two potential outcomes are doable. The algorithm’s in depth utilization arises from its capability to assemble a mannequin reflecting the affiliation between the dependent and unbiased variables by using the logistic operate. Essentially, logistic regression is a linear mannequin that computes the probability of an statement being half of a selected class. The algorithm develops a linear equation aligned with the gathered knowledge, subsequently making use of a sigmoid operate to transform the resultant worth right into a chance vary of 0 to 1. This strategy is rooted in the most probability estimation precept that goals to find out the best-fitting values for the coefficients tied to the unbiased variables. Regularization power, denoted as (C), is a parameter that balances the excellent match to the coaching knowledge and mitigates overfitting. Another parameter, the penalty (L1 or L2), decides the sort of regularization utilized, the place L1 encourages sparsity whereas L2 governs the dimension of the coefficients.
One of the interesting options of logistic regression is its capability to be interpreted. The coefficients tied to the unbiased variables depict each the extent and the path of their impression on the dependent variable. The algorithm performs optimally when the affiliation between unbiased variables and the dependent variable is linear or might be adjusted right into a linear kind. However, its effectiveness could lower when dealing with intricate non-linear relationships.
Gradient Boosting (GB)
Gradient Boosting is a supervised machine learning algorithm used for classification and regression issues. It is a strong boosting algorithm that mixes a number of weak learners into sturdy learners, in which every new mannequin is educated to attenuate the loss operate akin to imply squared error or cross-entropy of the earlier mannequin utilizing gradient descent. In every iteration, the algorithm computes the gradient of the loss operate with respect to the predictions of the present ensemble after which trains a brand new weak mannequin to attenuate this gradient. The predictions of the new mannequin are then added to the ensemble, and the course of is repeated till a stopping criterion is met.
Gradient Boosting is an ensemble learning methodology that mixes a number of weak learners, normally choice bushes, to create a robust predictive mannequin. GB builds the mannequin in an iterative method, specializing in the samples with greater errors in every iteration. It updates the mannequin by including weak learners that appropriate the errors of the earlier ones [39].The GB algorithm works by becoming the weak learners sequentially, the place every subsequent learner learns from the errors of its predecessors. It assigns greater weights to the misclassified situations, thereby boosting their significance in subsequent iterations. The learning charge (learning_rate) determines the contribution of every weak learner to the last prediction.GB has the benefit of having the ability to seize advanced relationships between the options and the goal variable. It handles high-dimensional knowledge properly and may successfully deal with non-linear relationships. The most depth of particular person regression estimators (max_depth) controls the complexity of the bushes, whereas the quantity of boosting phases (n_estimators) determines the quantity of weak learners to be mixed.
Support Vector Machine (SVM)
Support Vector Machine is a strong supervised learning algorithm used for classification and regression duties. SVM constructs a hyperplane or a set of hyperplanes in a high-dimensional house to maximise the separation between completely different lessons [23, 39]. The alternative of the kernel determines the choice boundary form.
SVM works properly when the lessons are well-separated, and the quantity of options is comparatively small. The kernel trick permits SVM to implicitly map the knowledge right into a higher-dimensional house, the place the lessons might be linearly separated. The regularization parameter (C) controls the trade-off between maximizing the margin and minimizing the classification error.SVM goals to search out the optimum hyperplane that maximizes the margin between the help vectors, that are the closest situations to the choice boundary. The kernel coefficient (gamma) controls the affect of particular person coaching samples on the choice boundary. A low gamma worth outcomes in a easy choice boundary, whereas a excessive gamma worth makes the choice boundary extra versatile.
Random Forest (RF)
Random Forest is an ensemble learning methodology that constructs a number of choice bushes and combines their predictions to make a last choice. Each tree is constructed utilizing a random subset of the coaching knowledge and a random subset of the options [23, 39]. RF reduces overfitting and improves accuracy by averaging the predictions of a number of bushes.
RF handles high-dimensional knowledge properly and may seize non-linear relationships between options and the goal variable. The quantity of bushes in the forest (n_estimators) determines the quantity of choice bushes to be constructed. The most depth of the tree (max_depth) controls the depth of every particular person tree, whereas the most quantity of options thought of for splitting (max_features) determines the quantity of options to contemplate for every cut up. The criterion for splitting (gini or entropy) determines the measure of impurity used to judge the high quality of a cut up.Random Forest combines the predictions of the particular person bushes by both majority voting (classification) or averaging (regression). This ensemble strategy helps cut back the impression of particular person bushes’ biases and improves the total accuracy and robustness of the mannequin.One benefit of Random Forest is its capability to deal with noisy knowledge and outliers. It also can present insights into characteristic significance, because it calculates the common lower in impurity for every characteristic throughout all bushes. Furthermore, it’s computationally environment friendly and may deal with massive datasets with high-dimensional options.
Artificial Neural Network (ANN)
Artificial Neural Network is an adaptable and highly effective ML mannequin impressed by the construction and performance of the human mind. ANN consists of interconnected nodes (neurons) organized in layers. Each neuron performs a weighted sum of its inputs, applies an activation operate, and passes the output to the subsequent layer [23].
ANN can be taught advanced patterns and seize non-linear relationships between options and the goal variable. The structure of an ANN consists of the quantity of hidden layers, the quantity of neurons in every hidden layer, and the activation operate used. The learning charge (alpha) controls the step dimension in updating the weights during coaching, whereas the regularization parameter (lambda) helps forestall overfitting [39].Training an ANN includes ahead propagation, the place the inputs are handed via the community to generate predictions, and backward propagation, the place the errors are propagated again to replace the weights. The alternative of activation operate, akin to sigmoid, ReLU, or tanh, impacts the mannequin’s capability to mannequin non-linearities.ANN might be computationally costly, particularly with massive datasets and sophisticated architectures. However, it has the benefit of having the ability to be taught hierarchical representations of the knowledge, making it appropriate for duties involving picture recognition, pure language processing, and sophisticated sample recognition.
Naïve Bayes (NB)
Naïve Bayes is a probabilistic classifier based mostly on Bayes’ theorem with an assumption of characteristic independence given the class label. Despite its simplicity, NB performs properly in many classification duties, particularly when the independence assumption holds. It is quick and requires a small quantity of coaching knowledge [40].
NB calculates the chance of an occasion belonging to a sure class by multiplying the conditional chances of every characteristic given the class. The algorithm assumes that the options are conditionally unbiased, which permits for environment friendly coaching and classification.NB doesn’t have many parameters to tune. However, it will probably deal with each categorical and numerical options by assuming completely different chance distributions. For categorical options, NB makes use of the frequency of occurrences, whereas for numerical options, it applies chance density estimation [40].One benefit of NB is its interpretability. It gives simple explanations of the classification selections based mostly on the chance calculations. NB is especially efficient in textual content classification duties, akin to spam filtering and sentiment evaluation, the place the independence assumption aligns properly with the nature of textual knowledge.
k-nearest neighbors’ algorithm (k-NN)
The k-nearest neighbors’ algorithm is a non-parametric, lazy learning algorithm that classifies knowledge based mostly on their proximity to different knowledge factors. In k-NN, the “ok” represents the quantity of nearest neighbors used to find out the class of a given knowledge level [41]. The algorithm assigns the class that’s commonest among its ok nearest neighbors.
The alternative of ok determines the stability between bias and variance in the mannequin. A smaller worth of ok (e.g., ok = 1) results in low bias however excessive variance, that means the mannequin is extra vulnerable to noise. On the different hand, a bigger worth of ok (e.g., ok = 5) reduces the impression of particular person knowledge factors however could introduce greater bias [41].k-NN depends on distance metrics, akin to Euclidean distance or Manhattan distance, to measure the similarity between knowledge factors. The algorithm searches the coaching dataset to search out the ok nearest neighbors and assigns the class based mostly on majority voting.k-NN is an easy and intuitive algorithm that may be utilized to each classification and regression duties. It doesn’t require coaching and is strong to noisy knowledge. However, its most important disadvantage is its computational complexity, particularly with massive datasets, because it must calculate distances for every question occasion.ML fashions parametersThe fashions’ options have been categorized into two teams: 1) the dependent variable, represented by the food consumption rating, and a couple of) the unbiased variables, together with all associated options. The options are additional represented in Table 1.
Table 1 The checklist of variables used in the machine learning mannequinThe ML fashions have been constructed based mostly on the knowledge ratio of 70:20:10 for coaching, testing, and validation. The grid search methodology and cross-validation with 10-folds have been used for parameters’ optimizations. The following parameters have been set for the ANN, RF, and SVM:
◦
In Artificial Neural Networks, the hidden layer had 100 neurons, with a 600-maximum quantity of iterations in reference to the logistic activation operate.
◦
The Random Forest bushes have been set to 1000 with 5 most depth bushes, and the leaf node minimal quantity was set to 1, whereas the most quantity of samples to separate the inner nodes was set to 2.
◦
The SVM regularization parameter was set to 10, the RFB kernel was set to 0.001, and the bias error management issue was set to 1.
Based on the parameters optimization outcomes, the optimized algorithms (ANN, SVM, RF) have been used in figuring out and predicting food insecurity.Data evaluationData cleansing, transformation, and normalization processes have been performed previous to constructing the ML knowledge evaluation. The last dataset consisted of 13,446 individuals. The seven ML fashions have been constructed and carried out utilizing the python orange knowledge mining software program, which was then used for testing and validating the ML fashions [42].The research has made use of a various vary of ML fashions that cowl a broad spectrum of ML methodologies, from linear fashions (LR) to ensemble strategies (GB and RF), and from instance-based strategies (k-NN) to neural networks (ANN).However, there are various machine learning (ML) algorithms and approaches that might probably be explored to foretell food insecurity. For instance, the Decision Tree fashions, akin to the Classification and Regression Trees (CART), which supply simplicity and ease of interpretation. Additionally, if the dataset consists of pictures the deep learning techniques like Convolutional Neural Networks (CNN) or Recurrent Neural Networks (RNN) is also thought of. Our evaluation didn’t embody these algorithms as a result of of the particular nature of the research drawback and dataset. The utilized fashions have been chosen as a result of of their simplicity, accuracy, and interpretability, which is essential when addressing advanced and delicate issues like food insecurity. Deep learning fashions, though highly effective, typically operate as enigmatic entities, offering restricted perception into the factors contributing to their predictions. Conversely, fashions akin to Decision Trees have the potential to overfit the coaching knowledge and should not generalize properly to new, unseen knowledge. The fashions chosen for this research intention to attain a sensible stability between predictive functionality and interpretability, making them a prudent alternative.Nonetheless, it’s all the time useful to discover and validate the software of different ML algorithms in predicting food insecurity. As future work, researchers might contemplate conducting systematic mannequin comparability research to establish the finest performing fashions for this particular activity. This couldn’t solely improve the prediction accuracy but in addition deepen our understanding of the advanced nature of food insecurity.Model options significance evaluationThe Shapley Additive exPlanations (SHAP) is an important methodology utilized to find out the significance of numerous options in deciphering the predictions made by any ML mannequin. Drawing from cooperative recreation principle, SHAP equitably assigns the affect of every characteristic in figuring out the mannequin’s final result [21, 24]. It assists in understanding the distinct contribution of every characteristic to the prediction of food insecurity. This allows us to quantify the significance of every characteristic in the decision-making of the mannequin.Different efficiency measures have been used to judge whether or not the ML fashions can predict food insecurity ranges and the associated risk factors, akin to accuracy, specificity, precision, recall, and F-measure.
https://bmcpublichealth.biomedcentral.com/articles/10.1186/s12889-023-16694-5
