Parkinson’s Disease (PD) is a prevalent neurodegenerative disorder that affects a big portion of the global elderly population. One of its frequent symptoms is hypomimia, a reduction in facial expressiveness that impacts patients’ social interaction and quality of life. As clinical assessment of this motor symptom is often subjective, this study develops and compares classical and DL computational approaches for the automatic classification of PD patients vs. healthy control (HC) subjects by analyzing hypomimia in facial videos. A database of 53 videos (30 PD patients and 23 HC subjects) recorded during a text-reading task was employed. Videos were pre-processed starting with facial landmark detection using Facemesh, followed by head orientation correction to ensure an adequate pose, facial region isolation, and size normalization. The classical methodology consisted of features extracted using Histogram of Oriented Gradients and Local Binary Patterns, which then fed Support Vector Machine classifiers. Strategies based on statistical features and all-frame analysis were evaluated, as well as early and late fusion techniques. In the DL approach we explored custom-designed CNNs with varying depths and regularization techniques. Alongside a pre-trained model for emotion recognition that was adapted via fine-tuning. All models were validated using a five-fold, subject-independent and cross-validation strategy. The results show that DL-based approaches outperform classical approaches, achieving an UAR of \(82.6\%\) .