Abstract
Introduction: Asthma and COPD patients are advised to stay moderate physically active as much as possible. The most common active means of transport are walking and cycling [1]. On average, people in The Netherlands, use a car 0.83 times/day and a bicycle 0.76 times/day for distances under the 10 km. People above the age of 50 also tend to cycle 2.2 km/day and walk 0.8 km/day. Activity monitors, worn on the lower back, are increasingly used to measure physical activity. Distinguishing walking from cycling has not been successful with a lower back activity monitor. This abstract describes a method for detection of cycling.
Methods: A classifier was developed on 46 days of data from a convenience sample of 17 subjects, aged 24 to 81 years, both males and females. The classifier received frequency content of 5-s windows of 3D low-back accelerations (DynaPort MoveMonitor) as input. Cycling was reported in a diary to train the classifier. A learning algorithm was developed to reduce the propensity for overlearning. This method optimizes both sensitivity and positive predictive value and classified every second of data as cycling or non-cycling. Subsequently, continuous bouts shorter than 62 s were considered non-cycling.
Results: The classifier detected 88.8% of all true cycling periods (sensitivity - SE). From all data classified as cycling, 88.6% was true cycling (positive predictive value - PPV). This resulted in a F1-score, (2*SE*PPV) / (SE+PPV), of 88.7%.
Conclusions: The classifier developed detected cycling with good accuracy. The method was developed and validated on the same data, hence validation on independent data is required.
Projectteam Mobility in NL, Government of The Netherlands, November 2010.
- Copyright ©ERS 2015
