Parkinson’s Disease (PD) manifests through a variety of motor impairments that significantly impact patients’ quality of life. Accurate motion data analysis enables earlier diagnosis and treatment planning. Recent advancements leverage Inertial Measurement Unit (IMU) sensors to non-invasively capture time-series data during activity, supporting motion deficit assessment in PD. This study presents QUaternion + AI-based Deficit Detection (QUAID), a deep learning-based framework designed to classify motion deficits in PD patients using IMU data. QUAID introduces a novel integration of quaternion kinematics and physical constraints using a physics-guided quaternion estimator for orientation estimation, followed by motion classification through a Bayesian-inspired classification network that quantifies predictive uncertainty. Comparative analysis shows that QUAID outperforms the Robust IMU-based Attitude Neural Network in orientation estimation accuracy, reducing the quaternion angular loss in orientation estimation from 0.0373 to 0.0083. On the classification task, QUAID achieves 97% accuracy while providing well-calibrated, uncertainty-aware predictions that enhance clinical interpretability.