Parkinson's Disease (PD) is a neurodegenerative disorder with four cardinal motor symptoms: bradykinesia, tremor, rigidity, and postural instability. Adaptive Deep Brain Stimulation (aDBS) is a promising treatment for PD that provides stimulation based on the expression of PD symptoms, improving effectiveness and reducing side effects compared to continuous DBS. Smartwatches can facilitate aDBS by enabling continuous detection of tremor and bradykinesia. However, for bradykinesia detection, existing studies using smartwatch data from PD patients' natural environment were limited in study duration and sample size. To address this, the current study collected smartwatch data for up to seven months in PD patients' natural environments. From 22 PD patients, the smartwatch data was pre-processed, features were extracted and analysed, and a Variational Autoencoder (VAE) was trained to develop a bradykinesia detection model. However, the VAE could only learn from frequency-domain inputs and not from time-domain input data after current pre-processing methods. Limitations were identified in data quantity, distribution, and quality, including low-frequency artefacts and noise. Despite these limitations, feature analysis indicated that the data set contains valuable information about PD motor symptoms. The results of the feature analysis and VAE training on frequency-domain inputs suggest that, after addressing limitations in data quantity and quality, it could be possible to train the VAE using smartwatch data from PD patients' natural environments. In conclusion, although a bradykinesia model was not successfully developed, this study demonstrated the required steps for training the VAE and laid the groundwork for future studies to develop a bradykinesia detection model.