Skeleton-Based Continuous Scale Parkinsonian Gait Score Estimation Using Omni-Dimensional Self-Attention Convolution Networks

Abstract

Gait abnormalities constitute a primary motor symptom of Parkinson’s disease (PD). Clinically, the Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) is widely recognized as the standard to evaluate gait impairment in PD. Recent skeleton-based methods have sought to estimate MDS-UPDRS gait scores, but most approaches treat score prediction as a coarse classification task, limiting their ability to capture subtle, progressive gait changes over time. These methods also often neglect clinical prior knowledge and fail to model localized gait features, leading to unsatisfactory performance. In addition, a continuous scale evaluation of gait impairment could result in a better formulation and adjustment of the treatment plan. In this paper, we introduce a novel framework for providing continuous-scale gait impairment estimation from the discrete annotation of MDS-UPDRS using skeleton data. First, we convert non-Euclidean skeleton information into two Euclidean spatiotemporal feature maps, ensuring a rigid spatial-temporal structure around the central joint. Next, we employ an omni-dimensional attention convolutional network to extract local spatiotemporal gait features within these normalized feature maps. We then integrate the features from both maps using an adaptive channel feature fusion module, capturing comprehensive gait information. Finally, we propose a numerical score prediction strategy that leverages MDS-UPDRS scores as anchors to predict gait impairment on a continuous scale without requiring continuous-scale annotations from clinicians. The effectiveness of the proposed approach is validated using a substantial clinical PD gait skeleton dataset.