Identifying Types of Physical Activity With a Single Accelerometer: Evaluating Laboratory-trained Algorithms in Daily Life

Abstract

Accurate identification of physical activity types has been achieved in laboratory conditions using single-site accelerometers and classification algorithms. This methodology is then applied to free-living subjects to determine activity behavior. This study is aimed at analyzing the reproducibility of the accuracy of laboratory-trained classification algorithms in free-living subjects during daily life. A support vector machine (SVM), a feed-forward neural network (NN), and a decision tree (DT) were trained with data collected by a waist-mounted accelerometer during a laboratory trial. The reproducibility of the classification performance was tested on data collected in daily life using a multiple-site accelerometer augmented with an activity diary for 20 healthy subjects (age: 30 ± 9; BMI: 23.0 ± 2.6 kg/m²). Leave-one-subject-out cross validation of the training data showed accuracies of 95.1 ± 4.3%, 91.4 ± 6.7%, and 92.2 ± 6.6% for the SVM, NN, and DT, respectively. All algorithms showed a significantly decreased accuracy in daily life as compared to the reference truth represented by the IDEEA and diary classifications (75.6 ± 10.4%, 74.8 ± 9.7%, and 72.2 ± 10.3 %; p <; 0.05). In conclusion, cross validation of training data overestimates the accuracy of the classification algorithms in daily life.