Patient Identification based on Physical Rehabilitation Movements using Skeleton Data

Human activity recognition (HAR) is used to observe human movement in healthcare. Physical disability evaluation requires the help of experts, and with the development of artificial intelligence and sensors, we want to observe diseases and conditions without experts in everyday life. In this paper, we propose an algorithm for identifying patients through physical rehabilitation movements. Algorithm evaluation was performed using a dataset of physical rehabilitation movements acquired with a Kinect sensor. The dataset includes skeleton data of 15 patients and 14 normal subjects for 9 gestures. The proposed algorithm obtains heatmaps from skeleton joints and detects features using the ResNet backbone. HAR was pre-trained for 9 gestures because the dataset for patient identification was insufficient to learn ResNet. Pre-trained ResNet with 50% of layers frozen was additionally trained for the patient and normal subjects. As a result, an accuracy of 98.59% was obtained for the shoulder flexion left gesture.