Parkinson’s disease is a common neurodegenerative disorder that affects about 1% of people aged 60 and above in developed countries. The number of Parkinson’s patients is expected to surpass 8 million in Europe alone within the next decade due to an aging population. More than 80% of patients experience motor symptoms, which can be alleviated with personalized medication plans. However, precise and continuous measurement of motor symptoms during daily activities is crucial for developing these plans. A wrist-worn smartwatch with 3D motion sensors has been proposed to estimate motor fluctuation severity in real-life settings. Using a semi-supervised deep learning approach, this method provides accurate motor symptom estimation with a nine-level granularity, benefiting medical professionals. The approach utilizes Co-Teaching, allowing self-learning through pseudo-labels generated from a large volume of unlabeled data. In summary, this approach offers a practical solution for measuring motor symptoms accurately and creating personalized medication schedules to enhance patients’ quality of life.