Classifying smartphone-based accelerometer data to obtain validated measures of subject activity status, step count, and gait speed

Background The ubiquitous spread of smartphone technology throughout global societies offers an unprecedented opportunity to ethically obtain long-term, highly accurate measurements of individual physical activity. For example, the smartphone intrinsic 3-D accelerometer can be queried during normal phone operation to save time series of acceleration magnitudes (in each of the component directions) for near-real time or post processing. Objective We describe simple, straightforward algorithms (based on windowed Fourier analysis) for accelerometer data quality control and behavioral classification. Methods To maximize the clinical utility of our classifications, we focused on differentiating the following conditions: forgotten phone, subject resting, low physical activity, high physical activity. We further differentiated high physical activity into epochs of walking and climbing stairs, and further quantified walking to infer step count and gait speed. Results We validated these algorithms in 75 individuals, in both laboratory (treadmill) and naturalistic settings. Our algorithm performance was quite satisfactory, with accuracies of 92-99% for all behavioral categories, and 87-90% for gait metrics in naturalistic settings. Conclusions We conclude that smartphones are valid and accurate platforms for measuring day-to-day physical activity in ambulatory, community dwelling individuals.