Effective Integration of Geotagged, Ancilliary Longitudinal Survey Datasets to Improve Adulthood Obesity Predictive Models

Obesity is a critical health issue world-wide and has been identified as a leading cause of chronic diseases such as cardiovascular disease, type-2 diabetes, stroke and certain types of cancer. In 2014, 1.9 billion adults were overweight and 600 million were obese. In this study, to facilitate early detection of childhood obesity, we present our methodology for effective data integration that allows modelers to import new attributes from auxiliary datasets using geospatial proximity, alongside the associated data uncertainty for each data point that is caused by the data aggregation process while estimating that attribute. We have used the data uncertainty estimate as input to various machine learning algorithms, to improve on the obesity prediction. As a case study, we have integrated the National Longitudinal Survey of Youth 1997 dataset with the US Census 2000 dataset and the 2000 CDC Growth Charts dataset to augment our prediction model with behavioral and environmental features. Compared to models with only biometric attributes, our empirical experiments show accuracy improvements when we incrementally consider behavioral aspects (8.9~10.2%), environmental aspects (12.1~12.3%) and data uncertainty estimates (18.3~25.6%).