Abstract:
|
Diabetes is a significant public health problem which significantly affects vulnerable populations. Social determinants of health are critical to understanding how disadvantaged groups face barriers. While the predictors of diabetes are well established, the relative importance of each sociodemographic risk is unclear. We proposed a machine learning (ML)-based system for predicting diabetes disease using a nationally representative sample. We used NHANES 2009-10 data to study cross-sectional associations between social (age, income level, education, marital status, race, employment status) and health (smoking status and depression) risk factors and diabetes. We applied several machine learning algorithms, in which the classifiers had a training-to-test split of 80% to 20%. In comparison to naïve Bayes (NB) and decision tree (DT), we found random forest plots to have superior accuracy to predict diabetes. Age had the strongest association with diabetes; income and education, had the second strongest association with diabetes. While age was expected to be a strong predictor, income level had a high relative importance. Socioeconomic status should be applied to disease prediction.
|