Leveraging Machine Learning for Predicting Health Perceptions: A Behavioral Analytics Approach Using U.S. National Survey Data

Published in the 25th International Conference on Electronic Business (ICEB 2025), affiliated with the Association for Information Systems (AIS). Access at - https://aisel.aisnet.org/iceb2025/1/

Project overview

This repository showcases an interpretable machine learning pipeline for predicting self-rated health (good vs poor) among U.S. adults using the 2023 Behavioral Risk Factor Surveillance System (BRFSS) national survey dataset. The project accompanies the ICEB 2025 paper “Leveraging Machine Learning for Predicting Health Perceptions: A Behavioral Analytics Approach Using U.S. National Survey Data.

The work emphasizes the importance of behavioral and socio-economic factors in shaping perceived health and contributes to the intersection of data science, public health, and behavioral analytics.

Key contributions

• End-to-end pipeline: data cleaning, feature engineering, class balancing, model training, and evaluation on BRFSS 2023.
• Focus on health perceptions: modeling self-rated health (SRH) using socio-demographic, behavioral, psychosocial, and clinical factors.
• Imbalanced learning: comparison of SMOTE, ENN, SMOTE-ENN, and a repeated under-sampling strategy with detailed performance reporting.
• Interpretable modeling: logistic regression and a stacked ensemble achieve AUC ≈ 0.81–0.82 with balanced F1 across both health classes.
• Research-ready code: implemented in Python 3.11 with scikit-learn, XGBoost, imbalanced-learn, and SHAP for explainability.

Repository structure

ML_model_train.ipynb – Main notebook containing:

• BRFSS 2023 loading and preprocessing (selection of 19 features, cleaning, encoding).​
• Multiple balancing strategies: SMOTE, ENN, SMOTE-ENN, and repeated under-sampling.​
• Model training: Logistic Regression, Random Forest, XGBoost, Decision Tree, and Stacked Ensemble.​
• Evaluation: confusion matrices, classification reports, cross-validated AUC/F1, and ROC curves.​
• SHAP-based interpretation for feature importance and directional effects.

Methods and models

Data & target

• Source: 2023 BRFSS (Behavioral Risk Factor Surveillance System), 433,324 records and 350+ attributes; after cleaning, 117,386 records are used.
• Target: Self-rated general health recoded to binary – “Good” (excellent/very good/good) vs “Poor” (fair/poor).
• Predictors (19 features): age group, sex, race/ethnicity, education, income, employment, exercise, smoking, heavy drinking, diabetes, heart attack, stroke, cancer, BMI category, depression, days of poor mental health, healthcare provider, and checkup regularity.

Handling class imbalance

Class distribution is skewed (≈76% good vs 24% poor health), so multiple strategies are evaluated. Implemented in the notebook:

• SMOTE oversampling.
• ENN (Edited Nearest Neighbours).
• SMOTE-ENN hybrid.
• Repeated under-sampling: create 5 balanced subsets, train models on each, and aggregate performance.

Repeated under-sampling yields the most balanced F1 for both classes and is used in the final models.

Models

All models are implemented with scikit-learn / XGBoost APIs in Python 3.11.

• Logistic Regression (baseline and cross-validated).
• Random Forest (with class weights and CV).
• XGBoost classifier.
• Decision Tree (depth-constrained with CV).
• Stacked Ensemble (LR, RF, XGB, DT, KNN as base learners with LR meta-learner).

Model performance:

Model	AUC	Precision	Recall	F1-score
Logistic Regression	0.81	0.73	0.73	0.73
Random Forest	0.72	0.71	0.71	0.71
XGBoost	0.73	0.73	0.73	0.73
Decision Tree	0.72	0.72	0.72	0.72
Stacked Ensemble	0.82	0.74	0.74	0.74

Interpretation & key findings

• Top predictors: poor mental health days, exercise, income, employment, and diabetes show the strongest influence on self-rated health in SHAP analysis.
• Behavioral and socioeconomic factors (mental health, physical activity, income, employment) dominate over traditional clinical risk behaviours like smoking or heavy drinking for perceived health.
• Under Social Determinants of Health (SDOH) and Health Belief Model (HBM), results highlight how lived social context and beliefs shape perceived health beyond purely clinical status.

Academic & portfolio use

• Applied health analytics on large-scale national survey data (BRFSS 2023).
• Handling severe class imbalance with multiple resampling techniques.
• Building interpretable models with logistic regression and SHAP, alongside stronger ensembles.
• Connecting ML results to public health theory (SDOH and HBM) in a peer-reviewed conference paper.

If you use this project in academic work, please cite the journal paper. You can access it online at - https://aisel.aisnet.org/iceb2025/1/

---