A Hybrid Model to Classify Physical Activity Profiles

Abstract

Diabetes is a chronic disease characterized by high blood glucose levels. This condition has a strong impact on the heart, eyes, and even kidneys, leading to several long-term health problems. It is estimated that about 422 million people live with this condition and over 1.5 million deaths per year are related to diabetes. Although there is no cure for diabetes, it can still be prevented or in the worst case managed, by implementing a healthy lifestyle, where exercising is a priority. One of the most basic ways to exercise is by walking. Although simple, it can be helpful to reduce blood sugar levels. The first step toward the right lifestyle for the diabetic patient is to maintain an active routine and improve it every day. Therefore, it is important to create an environment where the person can be motivated to be healthier and at the same time be supported to do so. Additionally, it is needed to consider that every person is different and therefore the support provided for each diabetic patient must be personalized according to his/her capabilities and necessities. In this paper, using a dataset of user activity, more specifically the daily walking data of different users, the focus was to define a machine learning model, capable of identifying distinct groups of users, to find their favorite routines related to physical activity data. To reach the proposed goal, a classification model with 95,6% prediction accuracy was produced. The resulting hybrid model, using temporal predictors, such as period of day and weekday, could identify 13 clusters that describe 13 different profiles of users according to 31 generated rules.