Can machine learning automatically detect the aligned trunk in sitting directly from raw video using a depth camera?

Background The Segmental Assessment of Trunk Control (SATCo) evaluates sitting control at seven separate trunk segments, making a judgement based on their position in space relative to a defined, aligned posture. SATCo is in regular clinical and research use and is a Recommended Instrument for Cerebral Palsy and Spinal Cord Injury-Paediatric by The National Institute of Neurological Disorders and Stroke (US). However, SATCo remains a subjective assessment. Research question This study tests the feasibility of providing an objective, automated identification of frames containing the aligned, reference trunk posture using deep convolutional neural network (DCNN) analysis of raw high definition and depth (HD+D) images. Methods A SATCo was conducted on sixteen healthy male adults and recorded using a Kinect V2. For each of seven segments tested, two different trials were collected (control and no-control) to simulate a range of alignment configurations. For all images, classification of alignment obtained from a trained and validated DCNN was compared to expert clinician’s labelling. Results Using leave-one-out testing, at the optimal operating threshold, the DCNN correctly classified individual images (alignment v misaligned) with average precision 92.7±16% (mean±SD). Significance These results show for the first time, automation of a key component of the SATCo test, namely identification of aligned trunk posture directly from raw images (HD+D). This demonstrates the potential of machine learning to provide a fully automated, objective SATCo test to enhance assessment of trunk control in children and adults for research and treatment of various conditions including neurodisability and stroke.