Machine Learning Identifies Parity as Key Predictor of Under-Five Mortality

A national epidemiological investigation utilizing data from the Bangladesh Demographic and Health Survey has revealed that cumulative fertility, measured as the total number of children ever born, acts as a primary predictor of under-five child mortality across the country. Published in the journal Discover Public Health, the collaborative study evaluated reproductive, demographic, and socioeconomic indicators using both survey-weighted multivariable logistic regression and ten supervised machine learning algorithms. The analytical findings demonstrate that elevated parity is strongly associated with an increased likelihood of pediatric death before the fifth birthday, with each additional birth conferring higher adjusted odds of mortality. By integrating modern algorithmic classification with traditional epidemiological regression, the researchers uncovered subtle non-linear trajectories within child survival patterns that standard linear modeling structures frequently overlook. The computational framework simultaneously established that maternal breastfeeding status and household residential density serve as pivotal survival factors, offering fresh empirical insights into the complex demographic dynamics influencing pediatric health across contemporary Bangladeshi communities.

Under-five mortality represents a fundamental global public health benchmark that reflects healthcare access, nutritional adequacy, and broader household socioeconomic stability. Over the past three decades, Bangladesh has achieved remarkable progress in reducing pediatric fatalities, recording an estimated 75 percent reduction in under-five mortality between the 1993–1994 survey period and 2022. Despite this substantial national decline, child mortality persists as an urgent challenge in low- and middle-income nations striving to fulfill Sustainable Development Goal 3.2, which targets a reduction of under-five deaths to at least 25 per 1,000 live births by 2030. Public health scholars have long theorized that reproductive patterns influence offspring survival through biological maternal depletion, wherein closely spaced or numerous pregnancies diminish maternal physiological reserves, and socioeconomic resource dilution, wherein limited household provisions and parental care are divided among multiple dependents. However, conventional statistical investigations have largely relied on linear modeling assumptions, underscoring the necessity of deploying machine learning algorithms capable of detecting complex interaction structures and potential threshold responses without requiring subjective parametric pre-specifications.

To investigate these reproductive determinants with rigorous demographic representation, the research team analyzed the publicly available Bangladesh Demographic and Health Survey conducted in 2022. The nationwide survey utilized a two-stage stratified random sampling framework designed to facilitate consistent comparative analyses across administrative divisions and urban-rural stratifications. In the initial sampling stage, primary enumeration areas were extracted systematically from national census sampling frames across all administrative divisions, followed by the randomized selection of a standardized quota of households within each enumerated cluster. Structured face-to-face interviews administered by trained personnel gathered extensive demographic, reproductive, and health records from eligible women following informed consent. For the specific empirical investigation, the analytical dataset was restricted to female respondents possessing comprehensive records across all designated study variables. The investigators excluded de facto temporary residents and any records exhibiting missing entries, resulting in a finalized, well-characterized cohort of 8,784 reproductive-aged female participants whose complete reproductive histories formed the empirical foundation of the subsequent statistical and machine learning analyses.

The analytical architecture designated under-five mortality as a binary dependent variable, coded affirmatively for any child death occurring before age five and negatively for surviving offspring. The primary independent exposure was cumulative fertility, operationalized as total children ever born up to the survey date, while potential confounders encompassed parental educational attainment, parental occupational classifications, place of residence, maternal age at first marriage and first birth, preceding birth intervals, household size, media exposure, and religious affiliation. Traditional multivariable analysis utilized complex-sample logistic regression executed in SPSS Version 25, incorporating specialized survey modules that explicitly integrated sampling weights calculated as V005 divided by one million, geographic stratification through variable V023, and cluster primary sampling units via variable V021. This design-based estimation established that each additional child ever born significantly increased the adjusted odds of under-five mortality by 73.4 percent, yielding an adjusted odds ratio of 1.734 with a 95 percent confidence interval spanning 1.495 to 2.012. The complex-sample regression yielded pseudo-R-squared values of 0.0317 for Cox and Snell, 0.1313 for Nagelkerke, and 0.1165 for McFadden.

The complex-sample logistic regression also identified household membership as a profound determinant of pediatric survival, uncovering an inverse association whereby smaller living arrangements carried substantially elevated risks compared to large domestic environments. Relative to households containing more than six members, children residing in homes with only one to three individuals exhibited an adjusted odds ratio of 7.022 for under-five mortality, with a 95 percent confidence interval between 3.584 and 13.758. Children in four-member households similarly experienced elevated odds of 2.096, whereas five- and six-member households did not show statistically significant differences, demonstrating a non-monotonic protective relationship that likely reflects the caregiving buffers and resource-pooling benefits provided by extended familial networks in Bangladesh. Maternal and paternal education demonstrated significant associations across specific categories, whereas preceding birth interval, maternal age at first birth, age at first marriage, residential location, media access, parental occupation, and religious affiliation did not reach statistical significance. Although current breastfeeding exhibited a strong inverse statistical relationship with mortality, the authors emphasized that cross-sectional survey structures introduce potential temporal ambiguity, precluding causal inference regarding protective timing.

To expand beyond traditional parametric regression, the investigators constructed ten supervised machine learning algorithms in Python, encompassing Logistic Regression, Support Vector Machine, K-Nearest Neighbors, Decision Tree, Random Forest, Gradient Boosting, AdaBoost, Multi-Layer Perceptron Neural Networks, Extreme Gradient Boosting, and Light Gradient Boosting Machine. Because under-five deaths represented an acute minority class comprising approximately 4 percent of the sample against 96 percent surviving children, standard predictive algorithms risked developing severe bias toward majority-class accuracy. The researchers counteracted this disparity by applying the Synthetic Minority Oversampling Technique within an isolated imbalanced-learn cross-validation pipeline, ensuring artificial minority synthesis occurred strictly within training folds to prevent data leakage into validation subsets. Stratified partitioning divided the full dataset into an 80 percent training corpus and a preserved 20 percent test partition that retained original demographic proportions. Hyperparameter optimization proceeded via RandomizedSearchCV across five stratified folds, explicitly prioritizing the minority-class F1-score as the governing objective, followed by post-hoc five-fold cross-validated sigmoid calibration and out-of-fold optimal classification threshold determination to optimize prospective clinical discrimination.

Evaluation across the independent testing partition demonstrated that overall classification accuracy constitutes an unreliable performance criterion when analyzing highly imbalanced public health outcomes. Notably, the Neural Network and Light Gradient Boosting Machine achieved misleadingly high overall accuracies of 93.84 percent and 93.18 percent, respectively, yet both exhibited an acute failure to detect actual mortality events. The Neural Network identified only 4 of the 34 validated mortality instances in the test data, yielding a substandard recall of 11.76 percent and an F1-score of 0.1081. Similarly, LightGBM isolated merely 6 mortality cases, producing a recall of 17.65 percent and an F1-score of 0.1412, confirming that their elevated general accuracy was driven almost entirely by the trivial classification of majority survival cases. Consequently, the research protocol prioritized balanced discriminant metrics, judging model efficacy through mortality-specific recall, precision, F1-scores, Matthews Correlation Coefficients, Cohen’s Kappa, and the Area Under the Receiver Operating Characteristic curve to determine which algorithmic architectures successfully balanced false negatives against false positive alarms.

Across all multi-metric evaluations, the Decision Tree emerged as the superior overall classification architecture for identifying under-five mortality cases within the test population. The Decision Tree attained a mortality precision of 15.25 percent, an F1-score of 0.2368, a Matthews Correlation Coefficient of 0.2424, a Cohen’s Kappa of 0.1973, and the highest discriminatory capability with an AUROC of 0.8045. It accurately identified 18 of the 34 observed test deaths, achieving a mortality recall of 52.94 percent while successfully categorizing 937 non-mortality cases. Gradient Boosting demonstrated the highest overall sensitivity by detecting 19 of the 34 fatalities for a peak recall of 55.88 percent, but achieved an AUROC of 0.7614 and an F1-score of 0.2275. Meanwhile, XGBoost attained a recall of 52.94 percent, an F1-score of 0.2182, and an AUROC of 0.7728, while Random Forest achieved an AUROC of 0.7854 and an F1-score of 0.2051. Baseline scikit-learn Logistic Regression matched the Decision Tree’s 52.94 percent recall but yielded a lower precision of 10.98 percent and an AUROC of 0.7271.

Model interpretability techniques offered crucial physiological and structural explanations regarding how individual features steered predictive determinations within the optimized Decision Tree. Gini-based feature importance calculations ranked the total number of household members as the single most influential predictive attribute with an importance weight of 0.294, followed closely by cumulative children ever born at 0.177 and breastfeeding status at 0.172. Paternal educational attainment ranked fourth at 0.122, while preceding birth interval scored 0.058 and paternal occupation registered 0.044. Explainable artificial intelligence frameworks utilizing SHAP value distributions and partial dependence plots elucidated significant non-linear dynamics governing parity. Specifically, the partial dependence trajectory demonstrated that predicted under-five mortality escalated progressively as parity increased from two to five children, after which the mortality response curve flattened into a plateau. This empirical stabilization indicates a clear parity threshold effect, verifying that the biological and resource strain of additional births operates in a distinct non-linear manner that traditional linear regression equations are inherently unequipped to capture without complex manual transformations.

While demonstrating the analytical power of machine learning in demography, the investigators highlighted essential methodological limitations that frame the appropriate interpretation of their findings. The fundamental cross-sectional design of the Bangladesh Demographic and Health Survey prevents causal attributions, restricting all conclusions to observed statistical associations. Furthermore, standard machine learning libraries in Python cannot natively accommodate complex multistage survey weighting, clustering, or geographic stratification, meaning algorithmic findings represent predictive patterns within the observed study sample rather than design-weighted population parameters. The authors also cautioned that self-reported survey records remain vulnerable to recall inaccuracies and social desirability influences. Nonetheless, the identification of a non-linear parity threshold beyond four births provides valuable empirical guidance for maternal and child health initiatives in Bangladesh. Aligning targeted family planning services, enhancing educational outreach, and reinforcing community breastfeeding support can assist public health agencies in mitigating high-parity vulnerabilities, providing substantive support toward fulfilling national commitments under Sustainable Development Goal 3.2.

Subject of Research: Machine learning and epidemiological analysis of fertility factors associated with under-five child mortality in Bangladesh

Article Title: Machine learning analysis of the association between fertility and under five mortality in Bangladesh using BDHS 2022 data

Article References: Naznin, S., Uddin, M. J., Ahmad, I., & Kabir, A. (2026). Machine learning analysis of the association between fertility and under five mortality in Bangladesh using BDHS 2022 data. Discover Public Health, 23(1), Article 1697. https://doi.org/10.1186/s12982-026-03102-7

Image Credits: AI Generated

DOI: 10.1186/s12982-026-03102-7

Keywords: under-five mortality, Bangladesh Demographic and Health Survey, fertility-related factors, children ever born, machine learning in public health, decision tree classifier, complex-sample logistic regression, SHAP analysis, partial dependence plots, class imbalance, synthetic minority oversampling technique, household size, breastfeeding status, child survival, maternal depletion, predictive modeling, Sustainable Development Goal 3.2