Identifying Pathways to Disease Using Data Mining: Understanding the Combined Individual- and Neighborhood-Level Health Indicators of Diabetes Mellitus and Asthma among High-Risk Philadelphians

Abstract

Introduction: Disadvantaged urban neighborhoods suffer disproportionate chronic illness burden related to social determinants of health. Studies have shown that socioeconomic characteristics and factors related to poor neighborhood conditions, such as physical inactivity and neighborhood disorder, to be associated with an increased risk of asthma and diabetes. Objectives: The primary aim of this study was to determine the hierarchy of individual and combined neighborhood health indicators that are predictive of asthma and diabetes in a population of high-risk Philadelphians, in order to make actionable recommendations that promote disease prevention. The secondary aim was to illustrate the relevance of using decision trees (data mining) to understand multilevel relationships among the predictors of complex health outcomes. Methods: Secondary data on individual health measures and neighborhood characteristics (N = 450) and vacant lot data (N = 676) was obtained from researchers at the University of Pennsylvania. RapidMiner data science software was utilized to perform decision tree analyses, illustrating the levels of influence and patterns between individual and neighborhood characteristics predicting asthma and diabetes. Results: Individual- and neighborhood-level factors were intricately related and demonstrated significant trends of influence on the outcomes of asthma and diabetes. The decision trees created in this study had high specificity, accurately classifying the factors that are protective of each disease. Factors that emerged as influential across all decision trees were those involving non-demographic variables: hours outside, psychological distress, recreational walking, walk to work, social and physical disorder, and certain vacant lot characteristics (primarily lot trash). Understanding the complex relationships that exist between individual- and neighborhood-level factors are vital for creating disease prevention programs, particularly within low socioeconomic populations, who have limited access to other prevention resources Conclusion: Improved neighborhood-level conditions related to social and physical disorder were consistently found to be protective of both asthma and diabetes in this urban population. This study illustrates the practicality of applying machine learning techniques for understanding complex public health issues.