The Social Side of Health: Using Machine Learning to Predict Mortality

Abstract

Why do some people live longer than others? Age and disease obviously play a major role, but factors such as smoking, income, BMI, and diabetes may also be connected to mortality. For this project, I used data from the 2013–2014 National Health and Nutrition Examination Survey (NHANES) linked with CDC mortality records. I focused on six variables: age, gender, income, smoking status, BMI, and diabetes.
After merging and cleaning the datasets, I had 6,100 participants with known mortality outcomes. I trained two machine learning models, Random Forest and Logistic Regression, using an 80/20 training and testing split. Random Forest achieved 92.9% accuracy and an AUC of 0.889, while Logistic Regression achieved 92.2% accuracy and an AUC of 0.856.
Age was the strongest predictor in the Random Forest model, followed by BMI and income. Logistic Regression also showed relationships involving smoking, diabetes, income, and age. Overall, the project showed me that mortality is not connected to just one factor. Both health and social conditions can provide useful information when predicting mortality.