Obesity is a serious epidemic in the United States. Prevalence of the disease has more than tripled since 1980 among children aged 2 to 19, while 27 percent of adults in the U.S. are obese. This number climbs to approximately 31 percent in Detroit and the state of Michigan as a whole.
Identifying the underlying obesity risk factors (ORFs) of this disease is a crucial step to treatment and prevention, and while analysis of these factors is often done using a single-task learning (STL) approach, this method may not tell the whole story.
“Obesity is a multi-faced health outcome,” said Ming Dong, professor of computer science. “Some ORFs are highly specific to a certain subpopulation and others are universal to the entire population.”
Recognizing the complexity of the issue and the disadvantages of STL approaches, Dong and his collaborators sought solutions using various multi-task learning (MTL). This method makes it possible to capture the data heterogeneity in the population while also utilizing shared information among subpopulations.
ORFs can be classified in three groups: health conditions (e.g. diabetes), social behaviors (e.g. smoking or alcohol use), and demographic characteristics (e.g. age or family size).
“Obesity may influence some subpopulations more than others,” said Dong. “Since people in various regions, ages and races are vastly different from each other, the subpopulations can be immensely distinguished in terms of ORFs.”
The research team’s analysis on a public dataset, using both multi-task feature learning (MTFL) and clustered MTL models, outperformed STL to demonstrate multi-level risk factor prioritization, which would enable a more precise intervention systems.
“Identifying the salient risks for obesity and variance among subpopulations is imperative to optimize prevention efforts and treatment,” said Dong.