Parkinson’s disease (PD) is a neurodegenerative disorder that leads to various motor impairments, including bradykinesia, muscular rigidity, and resting tremor. These symptoms affect fine motor skills including handwriting. One possible way to model this biosignal is with deep learning methods; however the lack of large enough annotated datasets imposes a major challenge in this topic. This paper focuses on using data augmentation (DA) and transfer learning (TL) techniques to improve the classification performance of PD patients vs. healthy control (HC) subjects across five handwriting tasks. Three methodological approaches were explored: (1) a baseline, (2) DA and TL using micrographia emulation with EMNIST characters, and (3) denoising diffusion probabilistic model (DDPM). Among these approaches, the second one achieved the highest classification accuracy of 67%, representing a 9.7% improvement over the baseline. The DDPM method provided moderate improvements, particularly in tasks that were shape-similar to the pretraining data (e.g., Numbers and ID), achieving accuracy gains of up to 8.7%. Overall, the findings suggest that incorporating domain-specific TL strategies and generative models like DDPMs help in addressing data scarcity and enhancing classification performance.