Application of neural networks for filtering non-impact transients recorded from biomechanical sensors

A number of biomechanical sensor applications are available on the market today for sports, as well as military applications. These sensors are capable of measuring accelerations and velocities of impact that may be useful for understanding head and neck biomechanics; although the proper use and accuracy of these analysis techniques is still a topic of research. A challenge with these sensors is filtering the raw data to include authentic impacts and exclude false events that are sometimes registered due to the sensitivity of the sensors. In this study we have developed an algorithm based on artificial neural networks and discrete fourier transform based filtering that will be used to distinguish between a biomechanically relevant event and a non-impact transient which is of no concern. Linear accelerations and angular velocities are used as inputs to the pattern recognition algorithm and the output vector contains the probability of each impact belonging to a certain class. Using this approach we report a specificity of 47% and a sensitivity of 88% in the ability to distinguish between real impacts and non-impact transients.