A Neural Network Based Algorithm for Dynamically Adjusting Activity Targets to Sustain Exercise Engagement Among People Using Activity Trackers

Abstract

It is well established that lack of physical activity is detrimental to overall health of an individual. Modern day activity trackers enable individuals to monitor their daily activity to meet and maintain targets and to promote activity encouraging behavior.However, the benefits of activity trackers are attenuated over time due to waning adherence. One of the key methods to improve adherence to goals is to motivate individuals to improve on their historic performance metrics. In this work we developed a machine learning model to dynamically adjust the activity target for the forthcoming week that can be realistically achieved by the activity-tracker users. This model prescribes activity target for the forthcoming week. We considered individual userspecific personal, social, and environmental factors, daily step count through the current week (7 days). In addition, we computed an entropy measure that characterizes the pattern of daily step count for the current week. Data for training the machine learning model was collected from 30 participants over a duration of 9 weeks. The model predicted target daily count with mean absolute error of 1545 steps. The proposed work can be used to set personalized goals in accordance with the individual's level of activity and thereby improving adherence to fitness tracker.

UBIWEAR: An end-to-end, data-driven framework for intelligent physical activity prediction to empower mHealth interventions

“…While they explored a variety of significant contextual factors, their small-scale dataset might not be representative of the general population, while it prohibited them from bechmarking data-hungry SotA deep learning approaches. Mohammadi et al [11] experimented both with traditional machine learning and a neural network architecture to predict a dynamically adjusted daily number of steps based on personal, environmental, and social factors. However, their experimentation was solely based on a small-scale dataset and they did not provide any information concerning the deep learning architectures and hyperparameters used, limiting the reproducibility and reusability of their approach.…”

“…C3 -Physical Activity Prediction Benchmarking & Evaluation: We experiment with six different learning paradigms for physical activity prediction, from machine learning to advanced deep learning architectures, and benchmark their performance for this complex learning task (L3). Through the experimentation with more advanced architectures, UBI-WEAR achieves a MAE (Mean absolute error) of 1087 steps, 65% lower in terms of absolute error than that of the SotA model [11], proving the feasibility of intelligent physical activity prediction.…”