Parkinson’s Disease (PD) is a progressive neurodegenerative disorder affecting millions globally, characterized by motor symptoms such as rigidity, tremor, bradykinesia and postural instability. While deep learning has shown promise in detecting motion deficit in PD, its black-box nature and reliance on large, homogeneous datasets limit clinical interpretability and generalizability. To address this, we propose a Neuro-Symbolic Motion Deficit Detection (NSMDD) framework that combines data-driven learning with symbolic reasoning to detect motion deficit in PD patients. The study uses real-world IMU data from nine structured motor tasks, extracting 810 time-series features and seven motion-based indicators such as standard deviation, range of motion, angular velocity magnitude, root mean square, dominant frequency, jerk, and sway magnitude to represent key motor symptoms. Symbolic knowledge encoded in first-order logic (FOL) generated interpretable motion deficit labels, integrated with neural features in the NSMDD model. The framework achieved 92% accuracy on a hold-out patient set, showing strong generalization. In contrast, CNN and SVM baselines, though effective on validation data, dropped significantly on the hold-out set, with 34% and 31% performance gaps, respectively. These results highlight the advantage of embedding clinically grounded symbolic knowledge for robust and interpretable motion deficit detection in PD.