Why your back hurts : Finding an efficient way to measure and evaluate sitting posture using a combination of body sensors placed on the body and machine learning

The goal of this paper is to test the effectiveness of various wearable sensors, wearable sensor positions, and machine learning for measuring posture quality. Various different seating postures were identified. These where divided into correct and incorrect postures. Measurements of people taking on these different postures where done using a sensor suit containing various stretch sensors, gyroscopes, and accelerometers in different positions. The data gathered during these measuring sessions was used to train the random forest machine learning algorithm. The variable importance measures of the random forest algorithm where used to identify the most important sensors and sensor positions. Using about 300 lines of data, the random forest algorithm can be trained to have an average accuracy of about 92 per cent. Adding random rotations to the sitting position (to simulate a real office setting) decreases said accuracy to 84 per cent. The variable importance estimations placed high importance on the shoulder and head placed gyroscopes. Using the data from only these sensors results in only a slight reduction in accuracy to 81 per cent.