A machine learning approach to measure and monitor physical activity in children

详细信息    查看全文

• 作者：Paul Fergus^aAuthor Vitae ; Abir J. Hussain^a ; Author Vitae ; John Hearty^bAuthor Vitae ; Stuart Fairclough^cAuthor Vitae ; Lynne Boddy^aAuthor Vitae ; Kelly Mackintosh^dAuthor Vitae ; Gareth Stratton^dAuthor Vitae ; Nicky Ridgers^eAuthor Vitae ; Dhiya Al-Jumeily^aAuthor Vitae ; Ahmed J. Aljaaf^aAuthor Vitae ; Jenet Lunn^aAuthor Vitae
• 关键词：Physical activity ; Overweight ; Obesity ; Machine learning ; Classification ; Neural networks ; Sensors
• 刊名：Neurocomputing
• 出版年：2017
• 出版时间：8 March 2017
• 年：2017
• 卷：228
• 期：Complete
• 页码：220-230
• 全文大小：646 K
• 卷排序：228

文摘

The growing trend of obesity and overweight worldwide has reached epidemic proportions with one third of the global population now considered obese. This is having a significant medical impact on children and adults who are at risk of developing osteoarthritis, coronary heart disease and stroke, type 2 diabetes, cancers, respiratory problems, and non-alcoholic fatty liver disease. In an attempt to redress the issue, physical activity is being promoted as a fundamental component for maintaining a healthy lifestyle. Recommendations for physical activity levels are issued by most governments as part of their public health measures. However, current techniques and protocols, including those used in laboratory settings, have been criticised. The main concern is that it is not feasible to use multiple pieces of measurement hardware, such as VO2 masks and heart rate monitors, to monitor children in free-living environments due to weight and encumbrance constraints. This has prompted research in the use of wearable sensing and machine learning technology to produce classifications for specific physical activity events. This paper builds on this approach and presents a supervised machine learning method that utilises data obtained from accelerometer sensors worn by children in free-living environments. Our results show that when using an artificial neural network algorithm it is possible to obtain an overall accuracy of 96% using four features from the initial dataset, with sensitivity and specificity values equal to 95% and 99% respectively. Expanding the dataset with interpolated cases, it was possible to improve on these results with 98.8% for accuracy, and 99% for sensitivity and specificity when four features were used.