A Machine Learning Approach to Measure and Monitor Physical Activity in Children to Help Fight Overweight and Obesity

Abstract

Abstract. Physical Activity is important for maintaining healthy lifestyles. Recommendations for physical activity levels are issued by most governments as part of public health measures. As such, reliable measurement of physical activity for regulatory purposes is vital. This has lead research to explore standards for achieving this using wearable technology and artificial neural networks that produce classifications for specific physical activity events. Applied from a very early age, the ubiquitous capture of physical activity data using mobile and wearable technology may help us to understand how we can combat childhood obesity and the impact that this has in later life. A supervised machine learning approach is adopted in this paper that utilizes data obtained from accelerometer sensors worn by children in free-living environments. The paper presents a set of activities and features suitable for measuring physical activity and evaluates the use of a Multilayer Perceptron neural network to classify physical activities by activity type. A rigorous reproducible data science methodology is presented for subsequent use in physical activity research. Our results show that it was possible to obtain an overall accuracy of 96% with 95% for sensitivity, 99% for specificity and a kappa value of 94% when three and four feature combinations were used.

The Key Factors in Physical Activity Type Detection Using Real-Life Data: A Systematic Review

“…Different sensors can obtain different movement information, informing the classifiers used for PA type recognition. Our results indicate that in addition to accelerometers, there were various additional portable sensors, termed “additional sensors,” such as heart rate sensors (Kwak and Lee, 2012; Fergus et al, 2015), barometer and foot pressure sensors (Skotte et al, 2014; el Achkar et al, 2016), imaging sensors (Ruch et al, 2011; van Hees et al, 2013; Adaskevicius, 2014), and GPS (Troped et al, 2008; Nguyen et al, 2013). For example, in el Achkar et al (2016) a gyroscope helped to differentiate between motion and posture states.…”

“…The sampling rate used for PATD varied between 10 and 100 Hz. Some studies also applied the terms “activity count” (Troped et al, 2008; De Vries et al, 2011; Ruch et al, 2011; Fergus et al, 2015) or “activity steps” (Troped et al, 2008; Nguyen et al, 2013; el Achkar et al, 2016) to report their sampling granularity (Table 1). Activity counts are the sum of the accelerations measured over a selected period (epoch time) (Ruch et al, 2011).…”

“…Children may perform activities more quickly (Ruch et al, 2011), whereas older adults would be expected to be slower. Eligible studies in this review cover different age ranges, from children aged 10–12 (Ruch et al, 2011; Fergus et al, 2015) to older adults with a mean age of 83 (Godfrey et al, 2011). Table 3 illustrates that most studies focused on young people, with only a limited number of studies of older people or children.…”

“…However, unlike with DTs, it is more complex and difficult for a user to comprehend how exactly an NN algorithm classifies its inputs. The NN classifiers are powerful approaches that have the potential to be used for detecting different posture and motion activities with high classification performance (Gyllensten and Bonomi, 2011; Ruch et al, 2011; Barshan and Yuksek, 2014; Bayat et al, 2014; Fergus et al, 2015; Spinsante et al, 2016). One of the restrictions of this classifier is the high processing time for the model development, which makes it not a very optimal classifier for a real-time application.…”

“…When the number of activities increased, the performance of ANN decreased by 14% using two 1D accelerometers (De Vries et al, 2011), whereas in the study of Barshan and Yuksek (2014), NN achieved more than 99% performance in detecting 19 activity types using five 3D IMU and half a day of real-life data. MLP, as a particular type of NN, also gained a very high performance in detecting four activities using unrestricted training data (Fergus et al, 2015).…”