Comparison of Several Algorithms to Estimate Activity Counts with Smartphones as an Indication of Physical Activity Level

Background: Dedicated devices like GT3X+, Actical or ActivPal have been widely used to measure physical activity (PA) levels by using cut-points on activity counts. However, the calculation of activity counts relies on proprietary software. Since smartphones incorporate accelerometers they are suitable candidates to determine PA levels in a wider population. Objective: Our aim was to compare several algorithms so that smartphones can reproduce the results obtained with GT3X+. The influence of smartphone location was also investigated. Methods: Volunteers participated in the experiment performing several activities carrying two smartphones (hip and pocket) and one GT3X+ (hip). Four algorithms (A1-A4) were considered to obtain GT3X+ counts from smartphone accelerometer signals. A1 was based on a traditional filtering on temporal domain and a posterior calculation of the area under the curve. A2 was based on computing histograms of acceleration values, which were used as independent variables in a standard linear regression procedure. A3 also used a linear regression, but in this case the independent variables were power spectrum bands, leading to a kind of filtering in the frequency domain. A4 was based on a direct measure of area under the rectified curve of the raw accelerometer signal. Performance was measured in terms of raw activity counts or the corresponding PA level classification. The influence of the algorithm was tested with a Quade test. Multiple comparisons were performed with Wilcoxon test with Bonferroni's correction. Besides, battery consumption was also measured as a secondary parameter. The output of the selected algorithm was compared with GT3X+ counts using correlation (pearson and spearman) and agreement (Intra-Class Coefficient, ICC and Bland-Altmann plots for raw counts, and weighted kappa for activity levels). Several experimental conditions regarding smartphone location were compared with Wilcoxon tests. Results: Thirty-two volunteers participated in the experiment. More refined algorithms based on filtering techniques did not prove to achieve better performance than A2 or A4. In terms of classification of PA level, A4 got the lowest error rate, although in some cases the differences with other algorithms were not statistically significant (p-value > 0.05). A4 is also the simplest and the one that implies less battery depletion. The comparison of A4 with GT3X+ gave good agreement (ICC = 0.937) and correlation (spearman = 0.927) for raw counts and good agreement when classifying four or two PA levels (weighted kappa = 0.874 or 0.923 respectively). Besides, in real situations, activity classification into four levels was significantly improved (p-value < 0.05) if data from several body locations were used to find model parameters. Conclusions: Simple algorithms can reproduce the results of GT3X+. Thus, smartphones could be used to control the fulfillment of PA recommendations previously validated with cut-points. However, it must be acknowledged that accelerometers are not the gold standard to measure PA. (C) 2018 AGBM. Published by Elsevier Masson SAS. All rights reserved.