Parkinson’s Disease (PD) functions as a continuous neurodegenerative medical condition that develops motor complications of tremors, rigidity, and slowness of movement, as well as cognitive and autonomic conditions. The correct early identification of symptoms remains essential to slow disease worsening and enhance treatment results for patients. DL and ML innovations have enabled MRI diagnosis of PD to reach new levels of improvement. This research creates a diagnostic model that unites the Hierarchical Variational Transformers (H-ViT) network for extracting features and the XGBoost classifier by combining gradient boosting and attention mechanisms to boost PD diagnostic performance. The researchers applied the framework through three neuroimaging databases, including PPMI, OpenNeuro and ADNI. When evaluated on the PPMI dataset, the combination of H-ViT and XGBoost delivered 94.5% accuracy and 92.8% precision, 93.4% recall and 93.1% F1-scores and a measurement of 96% AUC. Model results on the OpenNeuro dataset matched the first dataset by producing 93.7% accuracy and an AUC of 95%, indicating successful input generalisation indicating successful input generalisation with multiple modalities. This model demonstrated 90.8% accuracy and 92% AUC on the ADNI dataset while dealing with various neurodegenerative conditions. H-ViT model outperformed traditional CNN-based approaches in extracting complete information from hierarchical latent variables to enhance diagnosis reliability. These findings show that transformative models and XGBoost ensemble methods present an effective solution for real-life PD diagnosis, which displays scalability. Researchers will continue developing multi-modal integration systems to enhance AI model capabilities for medical use and explainable AI features