Parkinson’s disease (PD) is a progressive neurodegenerative disorder affecting millions worldwide, characterized by motor and non-motor symptoms that significantly impact daily life. Clinically, the Movement Disorder Society – Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) is the standard tool for evaluating PD severity. However, its reliance on subjective clinical assessment introduces variability, making it difficult to achieve standardized and continuous monitoring. Additionally, MDS-UPDRS scores are influenced by clinician bias, inter-rater inconsistencies, and variations in patient condition across different sessions, leading to a lack of real-time, objective, and repeatable assessment. Furthermore, the scale does not inherently support longitudinal tracking or real-world movement analysis, limiting its effectiveness for early detection and personalized treatment adjustments. To address these challenges, we propose a neuro-symbolic learning approach that combines formal logical reasoning with deep learning model trained by collected inertial measurement unit (IMU) data. The symbolic knowledge is extracted based on MDS-UPDRS and represented by first-order logic format. Then the architecture of the neural network is designed based on the extracted symbolic knowledge. The data is extracted from IMU sensors that are placed on chest, upper and lower arms. The collected IMU data, including triaxial accelerometer and gyroscope readings, is calibrated for accuracy, with feature extraction focusing on identifying kinematic patterns indicative of PD symptoms. Unlike traditional deep learning models that operate as black-box systems, our hybrid approach improves both interpretability and performance by incorporating prior medical knowledge. By enabling automated severity grading and personalized treatment adjustments, our model provides a scalable solution for real-world PD monitoring, reducing the burden on healthcare systems and supporting early intervention strategies.