Food insecurity and its related indicators in an Iranian population: An application of machine learning methods in data from Golestan Province

Background: Household food insecurity (FI) is a major public health concern associated with socioeconomic vulnerability. Unfortunately, the prevalence of food insecurity in Iran has increased substantially in recent years due to the economic challenges associated with prolonged international sanctions. In this context, this study aimed to estimate the prevalence of household FI in northern Iran and compare the predictive performance of traditional and machine learning algorithms for classifying households according to presence of FI. Methods: This cross-sectional study included 4,948 households from Golestan Province in north of Iran. The study sample was selected using the multistage cluster sampling between December 2021 and September 2022. FI was assessed using the Household Food Insecurity Access Scale. The data was analyzed using the logistic regression, random forest, support vector machine, decision tree, naïve Bayes, and artificial neural network models. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, balanced accuracy, Brier score, calibration plots, and DeLong’s test for pairwise AUC comparisons. The bootstrap resampling was applied to assess the stability and uncertainty of predictor importance. In addition, the estimated rankings derived from the logistic regression and random forest models were compared using the Spearman’s correlation coefficient. Results: The prevalence of household FI was 61.7%, including 32.3% mild, 20.3% moderate, and 9.1% severe FI. The multivariable logistic regression analysis showed that lower income, lower educational attainment, unemployment, lack of insurance coverage, absence of supplementary insurance, lack of car or house ownership, fewer employed household members, larger household size, being under coverage of supportive organizations, having incurable disease, and Turkmen ethnicity remained associated with FI. The pairwise DeLong’s test showed modest differences between the fitted models, with no statistically significant difference between the logistic regression and either neural network or SVM radial models after Bonferroni correction. Household income was ranked as the most important predictor in both the logistic regression and random forest models. Conclusions: In general, our estimates revealed that household FI was highly prevalent in the sample under study and strongly associated with socioeconomic disadvantage. Logistic regression, neural network, and SVM models demonstrated relatively similar predictive performance, while more complex ML approaches did not show substantial practical improvement over conventional methods in this dataset. Given the serious consequences of household food insecurity, health policymakers should pay more attention to improve socioeconomic conditions of households residing in the studied areas.