Assessing the severity of gait impairment in Parkinson’s disease (PD) using the Movement Disorder Society’s Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) is typically performed by clinical experts, but this process is time-consuming, subjective, and costly. To address these challenges, we propose a Guided Diffusion Model with an encoder-only transformer that automatically predicts gait severity by learning the underlying distribution of PD gait and leveraging domain knowledge critical for clinical evaluations. Our diffusion model enables us to generate synthetic PD gait video frames conditioned on clinical features determined by experts to assess disease severity. These synthetic samples contain novel movement patterns not present in the observed data; systems trained on this information have better prediction performance. In addition, we propose a novel classification algorithm that can learn a predictive model, from both observed training data and synthetic samples, to accurately assess PD severity. We evaluate the effectiveness of the proposed method using two human motion datasets across two tasks: PD severity prediction and action classification. Our approach in predicting PD, and our action classification is sufficiently accurate that it can be applied to general applications with healthy subjects performing similar tasks. The full codebase is available on GitHub: https://github.com/arshakRz/DiffuseGaitNet.