In-home neurogaming: Demonstrating the impact of valid gesture recognition method on high volume kinematic outcomes

The process of cleaning motion capture data of aberrant points has been described as "the bane of motion capture operators". Yet, managing the high volume kinematic data generated through in-home neurogames requires data quality control that, executed insufficiently, jeopardizes accuracy of outcomes. To begin to address this issue at the intersection of biomechanics and "big data", we performed a secondary analysis of a neurogame, evaluating gesture count as well as shoulder and elbow joint angle outcomes calculated from kinematic data in which valid gestures were identified through 3 methods: visual review of regions of interest by an expert (BP); manufacturer-recommended data smoothing (MS); and automated methods (AI). We hypothesized that upper extremity kinematic outcomes from BP would be matched by AI but not MS methods. From one person with post-stroke hemiparesis, upper-extremity kinematic data were collected for 6 days over 2 weeks using a Microsoft Kinect™-based neurogame. We calculated gesture count, shoulder angle, and elbow angle outcomes from data managed using BP, MS, and AI methods. BP identified 1929 valid gestures total over 6 days which was different than the other two methods (p = 0.0015). In contrast, the AI algorithm with best precision identified 4372 and MS identified 4459 valid gestures. Furthermore, angle outcomes calculated from AI and MS methods resulted in different values than BP (p < 0.001 for 5 of 6 variables). More research is needed to automate treatment of high volume, low quality motion data to support investigation of motion associated with in-home rehabilitation neurogames.