Unobtrusive machine learning based leg position detection during seated office work

Musculoskeletal disorders (MSD) affect more than 1.63 billion people worldwide. Office workers are at high risk of developing MSD due to prolonged sitting during office work. Technical aids that provide users with insights into their sitting behaviour are useful for the prevention of MSD. Despite the fact that leg position is known to influence spine curvature, most position detection devices focus on upper body posture. In comparison, we recorded contact force with the chair using pressure sensors with the additional information from videos that showed the sitting position in 30 participants. Each participant was recorded during 5 working days resulting in a median recording time of 22 hours and 53 minutes. The active number of sensors, mean value and area under the curve of sensors in specific areas, feature extraction with convolution filters, and center of pressure were extracted from the force sensor values. An XGboost classification algorithm was trained on these features to discriminate between four different leg positions and the absence of a user. This algorithm obtained an overall accuracy of 73% on the test set consisting of 6 participants. The f1-scores for the classes 'away', 'no cross', 'knee cross', and 'ankle cross' were 0.96, 0.54, 0.76, and 0.44 respectively. The class legs on the chair, which rarely occurred in the monitored population, were mistaken for 'no cross' and could not be identified correctly in the test set. A second XGboost classifier was able to differentiate between symmetric and asymmetric sitting leg positions and away with a weighted accuracy of 85%. Overall pressure mats are a promising technology for observing common leg postures in office environments.