Population Health-Based Machine Learning Reveals Associations Between Psychosocial Factors and Chronic Kidney Disease

Introduction

• BRFSS and NHIS surveys extend CKD risk modeling beyond treated hospital cohorts, but self-reports lack laboratory confirmation and can only flag high-risk people for screening.The introduction notes that mass laboratory screening also entails operational, infrastructural, and logistical challenges.Source: Introduction · pp. 2, 5, 6
• CKD affects an estimated 37 million Americans, more than one in seven, and 40% are unaware; it is projected to become the fifth leading cause of death by 2050 [1, 2].Traditional risk factors include diabetes, obesity, hypertension, and adverse life events [6–9].Source: Introduction · p. 2
• Imputation methods capture different missing-data patterns; MICE can be computationally challenging with high-dimensional data and may poorly represent nonlinear feature relationships.The introduction describes machine-learning and deep-learning approaches as alternatives that can better handle complex patterns without traditional distributional assumptions.Source: Introduction · p. 4
• Deleting incomplete survey records can reduce sample size, worsen class imbalance, and bias estimates when data are not missing completely at random.The introduction argues that missing records should be handled carefully because underrepresented classes may lose further representation.Source: Introduction · pp. 3–4
• Earlier studies rarely addressed mixed data types, substantial missingness, and severe class imbalance together; this study compares multiple imputers and resampling methods and adds explainability.The introduction also notes that BRFSS missing values had not previously been addressed in CKD classification research.Source: Introduction · pp. 4–5
• The study evaluates nine imputation methods, five resampling strategies, and eleven ML/DL approaches, then applies a stacked XGBoost, Random Forest, and Logistic Regression model with SHAP interpretation.The stacked model uses a Multi-Layer Perceptron as its meta-learner, and medical professionals reviewed the interpretation.Source: Introduction · pp. 5–6

Performance on NHIS datasets

• The NHIS 2020 Stacked Ensemble had the highest Balanced Accuracy and AUROC in the model ranking, while GAN, XGBoost, and Diffusion best preserved original feature properties.Source: Performance on NHIS datasets · p. 6
• RUS generally produced the best performance, whereas no sampling produced the worst; on NHIS 2020, RUS and Stacking Ensemble led CKD classification performance.Source: Performance on NHIS datasets · pp. 6–7
• 0.7537 Balanced Accuracy, 0.7397 Sensitivity, and 0.7700 Specificity were highest for NHIS 2021 ANN_2_Layer with GAN imputation and RUS sampling.Cross-validation Balanced Accuracy was 0.6730 ± 0.01.Source: Performance on NHIS datasets · p. 6
• 0.7516 Balanced Accuracy, 0.7977 AUROC, 0.7362 Sensitivity, and 0.7600 Specificity characterized the NHIS 2021 Stacked Ensemble with GAN imputation and RUS sampling.Cross-validation Balanced Accuracy was 0.7003 ± 0.03 and cross-validation AUROC was 0.7724 ± 0.02.Source: Performance on NHIS datasets · p. 6
• 0.8069 AUROC was highest for NHIS 2021 LR with XGBoost imputation and ROS sampling; cross-validation AUROC was 0.7776 ± 0.02.Source: Performance on NHIS datasets · p. 6
• 0.8229 AUROC and 0.7661 Specificity were highest for the NHIS 2020 Stacked Ensemble with GAN imputation and RUS sampling.Its cross-validation AUROC was 0.7724 ± 0.02.Source: Performance on NHIS datasets · p. 7
• 0.7648 Balanced Accuracy was highest for NHIS 2020 GradientBoosting with GAN and ROS, while ANN_2_Layer with XGBoost and SMOTE had the highest Sensitivity, 0.7498.Source: Performance on NHIS datasets · p. 7

Performance on BRFSS datasets

• On BRFSS 2021, GradientBoosting+GAN+RUS led AUROC (0.7977) and balanced accuracy (0.7289), while Stacked Ensemble+XGBoost+RUS led sensitivity (0.7547); in 2019, GradientBoosting+GAN+ROS led AUROC (0.7984) and balanced accuracy (0.7321).Cross-validation AUROC and balanced accuracy were 0.7924 ± 0.004 and 0.7221 ± 0.005 for the 2021 GradientBoosting pipeline, and 0.7936 ± 0.004 and 0.7239 ± 0.004 for the 2019 pipeline. Stacked Ensemble+GAN+RUS had the highest 2019 sensitivity, 0.7751.Source: Performance on BRFSS datasets · p. 7

Model performance analysis across all datasets

• At the approximately 3.9% baseline CKD prevalence, leading pipelines had PPVs of 8.5%–9.8% and NPVs of 98.5%–98.8%; the proposed framework reduced missed diagnoses versus unadjusted baselines.Stacked Ensemble false-negative rates were 8.78–9.60 cases per 1,000 screened for BRFSS, versus 39.51–121.14 missed cases per 1,000 for simple baselines. Brier scores of 0.167–0.193 and reliability diagrams suggested good calibration.Source: Model performance analysis across all… · pp. 7–8
• The Stacked Ensemble appeared among the top five models in all four test sets, and McNemar tests found significant superiority across BRFSS and NHIS comparisons.Ensemble Stacking (GAN+RUS) significantly outperformed all compared models on BRFSS 2019, while Ensemble Stacking (XGBoost+RUS) showed significant superiority on BRFSS 2021; significant advantages were also reported for NHIS comparisons.Source: Model performance analysis across all… · pp. 7–8

Feature importance analysis using SHAP values

• Across BRFSS cohorts, difficulty walking, blood pressure, age, employment status, and physical health ranked among leading CKD classification features.For BRFSS 2021, difficulty walking, blood pressure, and age had SHAP values of 0.19, 0.15, and 0.14; for 2019, these were 0.17, 0.21, and 0.14. Alcohol consumption, BMI, mental health, and parental separation also contributed, while verbal insult and health insurance had minimal or no impact in 2021.Source: Feature importance analysis using SHAP… · pp. 8–9
• In NHIS, difficulty walking, age, income, concentration difficulty, and cholesterol awareness were influential features; regular checkups also contributed to 2021 CKD classification.For NHIS 2021, SHAP values were 0.23, 0.17, 0.15, 0.16, and 0.12 for difficulty walking, age, income, concentration difficulty, and cholesterol awareness, respectively; regular checkups had a value of 0.08. In 2020, the corresponding leading values were 0.29, 0.16, 0.16, 0.14, and 0.13, with blood pressure, regular checkups, and BMI also contributing.Source: Feature importance analysis using SHAP… · p. 9

Discussion

• The study combined nine imputation techniques, five sampling methods, and eleven classifiers; parent separation retained similar SHAP impact without imputation, while dropping missing data lost information and reduced performance.The authors report that GANs better preserved minority-class patterns and that strict in-loop cross-validation avoided data leakage.Source: Discussion · pp. 9, 11
• SHAP ranked regular checkups highest in both NHIS years and identified age, high blood pressure, difficulty walking, employment, income, and parental separation as influential CKD-classification features.The authors hypothesize that chronic stress may affect renal health through neurohormonal pathways, while cautioning that SHAP importance does not establish biological causality.Source: Discussion · p. 10
• McNemar tests found significant improvements over Gradient Boosting (p < 0.001) and Logistic Regression (p ≤ 0.015) on both NHIS and BRFSS cohorts.The authors attribute the gains to combining three base models with an MLP meta-learner.Source: Discussion · pp. 10–11
• Because CKD status was self-reported in cross-sectional surveys, the model predicts likelihood of reported diagnosis, may miss undiagnosed early cases, and cannot establish causality.Highly ranked indicators such as difficulty walking and checkups may proxy comorbidities, frailty, healthcare use, or consequences of illness, limiting prospective early-warning use.Source: Discussion · pp. 12–13
• Unweighted survey data limit results to the analyzed cohort rather than nationally representative estimates, so the authors call for survey-weighted analyses and external-cohort validation.They also note that LLMs and other advanced AI methods were not evaluated and could be explored in future work.Source: Discussion · p. 13
• At 0.018 to 0.025 milliseconds per patient (>38,000 to >54,000 evaluations/second), the framework is compatible with typical EHR systems without dedicated GPU and offers a low-cost population survey-based triage tool for identifying high-risk individuals.The authors describe its intended role as concurrent case finding and opportunistic risk stratification.Source: Discussion · pp. 13–14

Methods

• The multi-stage workflow applied feature selection and mapping, advanced imputation with discrete-code constraints, resampling, model training and evaluation, and SHAP analysis.Imputed continuous values for categorical and ordinal survey variables were rounded to the nearest integer to match valid CDC codebook values.Source: Methods · p. 14

Dataset description

• The study used BRFSS 2021 (438,693 participants, 304 features), BRFSS 2019 (418,268, 342), NHIS 2021 (29,482, 622), and NHIS 2020 (31,568, 617).The public telephone-survey datasets cover health behaviors, conditions, and socioeconomic factors and contain substantial missing data and class imbalance.Source: Dataset description · p. 14

Feature selection

• The study selected 32 input features spanning health status, adverse childhood experiences, demographic characteristics, and disability information.BRFSS features included 28 categorical and four numeric variables; 11 NHIS features were excluded and common features were marked.Source: Feature selection · pp. 14–15

Data validation and splitting

• The 80/20 stratified split preceded imputation; algorithms were fitted only on training data, while the 83,333-sample test set remained isolated.The target was excluded during imputation, and the test set was transformed using parameters and synthetic distributions learned from training data only.Source: Data validation and splitting · p. 15
• Five-fold cross-validation stayed within the pre-imputed training set, with scaling, synthetic data generation, and ROS, RUS, SMOTE, and ADASYN restricted to training folds.This was intended to prevent validation-fold information leakage and optimism bias.Source: Data validation and splitting · p. 15
• The held-out test sets retained their naturally imbalanced distributions, at approximately 3.9% baseline CKD prevalence, for screening metrics and Brier scores.The study also used McNemar’s test against baseline models to assess the superiority of its best models.Source: Data validation and splitting · p. 15

Data preparation

• CKD labels were coded 1 for positive and 0 for negative responses in BRFSS and NHIS, with missing, unsure, or refused target responses dropped.BRFSS asked whether respondents had ever been told they had kidney disease; NHIS asked about weak or failing kidneys, excluding stones, infections, or incontinence.Source: Data preparation · p. 15
• Invalid and structurally missing survey responses were converted to NaN for imputation, while ordered features were coded sequentially and binary responses as 0 or 1.The NaN values were passed directly to imputation algorithms rather than filled using manually specified rules.Source: Data preparation · pp. 15–16

Missing value generation

• Imputation methods were selected with regard to whether data were Missing At Random or Missing Not At Random, because complex medical datasets may not fit traditional statistical assumptions.The authors state that advanced methods can handle complex patterns without imposing strict mathematical assumptions.Source: Missing value generation · p. 16
• Forward-backward filling served as a baseline, while KNN averaged nearest-neighbor values and processed large datasets in five chunks of at most 100,000 samples.Forward-backward filling considers only one variable; KNN was excluded from the final framework because of performance and time constraints.Source: Missing value generation · p. 16
• Autoencoders encoded and reconstructed data to impute missing values, while XGBoost handled missingness by separating incomplete rows and training on complete data.The passage characterizes autoencoders as scalable for large, complex datasets and notes XGBoost uses L1 and L2 regularization.Source: Missing value generation · p. 16
• The diffusion approach added Gaussian noise in a forward step and recovered data through reverse denoising, but sampling was slow because it required many steps.The described network used three fully connected layers, hidden size 128, and ReLU activations in its two hidden layers.Source: Missing value generation · p. 17
• The GAN generator imputed missing components while a discriminator distinguished observed from imputed components, aiming to preserve patterns in the full dataset.The two ReLU networks used Adam at learning rate 0.001 and binary cross-entropy for 400 epochs; the trained generator replaced only missing values.Source: Missing value generation · p. 17
• Other approaches included matrix factorization with SVD for balanced, dense matrices, mean imputation using observed-value averages, and dropping rows with missing values.The passage describes matrix factorization as reconstructing a matrix from reduced subparts and notes its sensitivity to the balance of observed and missing entries.Source: Missing value generation · p. 17

Resampling techniques

• The dataset contained only 3.9% positive cases, risking majority-class bias and poor minority-class sensitivity, which motivated use of sampling strategies.Source: Resampling techniques · p. 17
• ROS duplicates minority cases [127], RUS removes majority cases [128], and SMOTE interpolates new minority samples between neighbors [72,129].Source: Resampling techniques · p. 17
• AdaSyn targets sparse, hard-to-classify minority cases and generates additional synthetic samples by interpolation with their nearest neighbors [130, 74].Source: Resampling techniques · pp. 17–18
• ROS, RUS, and SMOTE yielded nearly balanced class distributions of about 50:50 for most missing-value methods; AdaSyn slightly improved minority-class representation.Without resampling, Class 0 comprised about 96% of the data and Class 1 about 3–4%.Source: Resampling techniques · p. 18

Predictive Modeling Architectures

• Eleven predictive approaches were trained on BRFSS 2019 and NHIS 2020 training sets, then evaluated on both the corresponding 2019/2021 and 2020/2021 survey datasets.Architectural details and hyperparameters are shown in Table 14.Source: Predictive Modeling Architectures · p. 18
• The study explored classical machine learning, deep learning, and a custom stacked hybrid architecture to address high dimensionality, class imbalance, and complex data patterns.Source: Predictive Modeling Architectures · p. 18

Classical machine learning models

• Classical models include LR, DT, XGBoost, AdaBoost, and RF, using logistic probabilities, tree splits, sequential error correction, weak-learner weighting, and tree voting, respectively.The passage notes limitations including LR’s difficulty with many additive coefficients, DT overfitting on small datasets, XGBoost complexity and memory use, AdaBoost sensitivity to noise, and RF computational cost.Source: Classical machine learning models · p. 18
• Gradient Boosting sequentially adds models to correct earlier errors and can handle complex patterns in structured medical data, but training may be slow on large datasets.The passage also cautions that it cannot achieve accurate results without missing-value handling.Source: Classical machine learning models · p. 19
• LightGBM builds leaf-wise decision trees with histogram-based processing for structured data, though it can overfit small datasets.Source: Classical machine learning models · p. 19

Deep learning architectures

• The study designed three ANN variants: a 64-neuron single-layer model, a 128/64-neuron two-layer model, and a 64/32-neuron model with input attention.The hidden layers use ReLU and the output layer uses sigmoid; attention is intended to prioritize important variables before the dense layers.Source: Deep learning architectures · p. 19

Hybrid model implementation

• The hybrid ensemble combines Logistic Regression, Random Forest, and XGBoost base estimators with an MLP meta-learner to learn from their out-of-fold predictions.The three base models were trained independently; the MLP was trained on out-of-fold predictions generated during 5-fold cross-validation.Source: Hybrid model implementation · p. 19
• Hyperparameters were selected using preliminary validation within training folds, while the 20% test set remained untouched until final evaluation.The text states that the held-out test set was used once after the ensemble was finalized.Source: Hybrid model implementation · pp. 19–20
• The proposed heterogeneous stack differs from commonly used homogeneous CKD ensembles and linear meta-learners by using an MLP to model nonlinear transformations and error patterns [36,38,42].Its base estimators capture linear and nonlinear patterns, while the MLP is intended to learn complex decision boundaries in public health survey data.Source: Hybrid model implementation · p. 20

• SHAP was used to interpret the ensemble’s predictions and analyze feature contributions.Source: Model interpretability using SHAP · p. 20
• The study selected SHAP over heuristic tree-based metrics and approximation methods such as LIME, whose performance may be inconsistent near complex decision boundaries145.Source: Model interpretability using SHAP · p. 20
• Grounded in cooperative game theory, SHAP provides local accuracy and consistency while distributing credit across correlated socioeconomic and clinical factors for individual- and population-level insights.Source: Model interpretability using SHAP · p. 20

Code availability

• The analyses used Python libraries current at submission in 2025, ran on Google Colab, and the code is publicly available on GitHub.https://github.com/AtikShams/CKDSource: Code availability · p. 20

• The supplementary document accompanies the submission; correspondence and requests for materials should be addressed to Tanmoy Sarkar Pias.Source: Additional Information · p. 21

Figures and tables

• NHIS pipelines show sensitivities of 0.73–0.75 and specificities of 0.74–0.77, while BRFSS pipelines show sensitivities of 0.72–0.77 and specificities of 0.68–0.71.Figure 2 displays normalized confusion matrices for the five leading pipelines on NHIS 2020/2021 and BRFSS 2019/2021.Source: Figure 2 · p. 34, Figures and tables · p. 35
• Figure 3 presents GAN + RUS calibration curves for NHIS 2020/2021 and BRFSS 2019/2021, with proximity to the diagonal indicating more trustworthy risk predictions.Source: Figure 3 · p. 36
• Across all four datasets, Ensemble Stacking sits on the top-tier sensitivity–specificity line with a high overall score, while its AUROC score range varies more with data preparation.Figure 4 also presents AUROC curves and radar charts comparing AUROC, balanced accuracy, sensitivity, and specificity.Source: Figures and tables · p. 35, Figure 4 · p. 37
• SHAP rankings identify blood pressure, difficulty walking, and age as the largest-impact BRFSS features, and difficulty walking, age, and income level as the most influential NHIS features.Source: Figure 5 · p. 38
• The proposed Stacked Ensemble combines XGBoost, Random Forest, and Logistic Regression as base estimators with MLP as the meta-learner.Figure 1 depicts the broader workflow from data collection toward helping patients develop a healthy life.Source: Figure 1 · p. 33, Figure 6 · p. 38
• Tables 1–5 cover NHIS model and sampling performance, imputation-related feature-distribution changes, and top-model comparisons for NHIS and BRFSS.Source: Table 1 · p. 39, Table 2 · p. 39, Table 3 · p. 40

NHIS 2020 Feature map

• Supplementary Tables 4–6 map original NHIS 2020 demographic, activity, and disability/health feature names to new names, including BMI, insurance, and mental-health variables.The mapped variables include computed BMI categories, health-insurance status, and computed mental-health status.Source: Table 4 · p. 50, Table 5 · p. 51, Table 6 · p. 51
• 32 dataset features span demographics, adverse childhood experiences, and behaviors such as alcohol use, smoking, and physical activity.The distribution plots distinguish positive and negative kidney-disease cases across feature categories.Source: NHIS 2020 Feature map · p. 58

Discussion:

• The discussion proposes that brain-regulated neurotransmitters and hormones can alter kidney function, citing dopamine’s effects on sodium reabsorption and the renin-angiotensin system.It also links stress-related norepinephrine and epinephrine release and altered serotonin to possible renal effects.Source: Discussion · p. 58
• CKD prevalence is reported as 34% among adults over 65 versus 12% among those aged 45–64, while the discussion cautions that aging has no direct linear relation to CKD.The discussion relates age-associated risk to declining glomerular filtration and other comorbidities.Source: Discussion · p. 58
• The discussion describes psychological stress as a possible indirect CKD contributor through effects on blood pressure and health, while stating employment status has no direct impact.It says employment may affect physical and mental health, and economic status can make treatment and dialysis expenses difficult.Source: Discussion · p. 59
• The discussion says high BMI is not directly related to CKD but may accompany comorbidities that indirectly affect renal function.It also associates low educational attainment with adverse outcomes and CKD etiology, citing lack of awareness and health advice.Source: Discussion · p. 59
• The discussion presents health-insurance coverage as a nonclinical factor that can affect CKD outcomes through coverage limitations.The feature map includes a renamed NHIS variable for whether a respondent has health insurance.Source: NHIS 2020 Feature map · p. 52, Discussion · p. 60
• Limited healthcare access may impede diagnosis and treatment, while affordability has a socioeconomic role in CKD.The discussion says patients unable to access healthcare facilities are more prone to advancing CKD and cites diagnosis, investigations, medications, and dialysis as needed care.Source: Discussion · p. 60
• Touched sexuality has no significant impact on CKD, but CKD can affect sexual life and cause sexual dysfunction.Source: Discussion · p. 60
• Sexual harassment is described as having no direct CKD impact, though related psychological stress or depression may impair kidney function and raise blood pressure.The discussion links these effects to CKD advancement.Source: Discussion · p. 60
• Housing instability can negatively affect CKD outcomes, and the discussion calls for screening, interventions, and advocacy to address housing needs.Source: Discussion · p. 60
• Black, Hispanic or Latino, Asian, and Native American communities are described as being at higher CKD risk due to genetic, socioeconomic, and healthcare-access barriers.Source: Discussion · p. 60