A novel algorithm to detect non-wear time from raw accelerometer data using convolutional neural networks

Current non-wear detection algorithms frequently employ a 30- to 90-minute interval in which recorded acceleration needs to be below a threshold value. Such intervals need to be long enough to prevent false positives (type I errors), while short enough to prevent false negatives (type II errors), limiting their ability to achieve a high F1 score. In this paper, we propose a novel non-wear detection algorithm that eliminates the need for an interval. Rather than inspecting acceleration within intervals, we explore acceleration patterns right before and right after an episode of non-wear time. By drawing on insights from the field of activity type recognition, we propose an algorithm that uses a convolutional neural network to detect the preceding activity 9taking off the accelerometer9 and the following activity 9placing it back on9. We evaluate our algorithm against several baseline and existing non-wear algorithms for raw accelerometer data, and our algorithm achieves a perfect precision, a recall of 0.9962, and an F1 score of 0.9981, outperforming all evaluated baseline and non-wear algorithms. Although our algorithm was developed using patterns learned from a hip-worn accelerometer, we propose algorithmic steps that can easily be applied to a wrist-worn accelerometer and a retrained classification model.