Stage classification and report generation of Parkinson’s disease using multimodal data and machine learning

Parkinson’s Disease (PD) is a progressive neurodegenerative disorder characterized by dopaminergic neuron loss, resulting in motor deficits such as gait freezing, tremors, and speech impairment. Since these abnormalities appear across multiple diagnostic dimensions, such as gait patterns, vocal characteristics, and structural brain changes, they offer several measurable signals that can aid in the detection and staging of Parkinson’s disease. This study explores an explorat3ory multimodal machine learning framework designed to integrate gait, voice, and MRI data with graph-based knowledge retrieval to investigate its utility in automated PD detection and clinical reporting. In the gait module, biomechanical features were extracted from planar-pressure recordings and refined through feature selection. Results using XGBoost suggest that these specific gait parameters may be highly sensitive indicators for PD classification. Similarly, the speech module utilized Wavelet Scattering Transforms and weighted majority voting, indicating the potential of speech modality for Parkinson’s disease classification. The MRI module focused on identifying 30 key brain region volumes through LASSO based selection. While results were more modest for 4-class staging, these findings highlight the complexities of structural imaging in granular disease progression. When these three modalities were combined via an ensemble approach, the system showed strong performance across healthy and disease states, pointing toward the complementary nature of multimodal data in improving diagnostic stability. Finally, a graph-based Retrieval-Augmented Generation system was implemented as a proof-of-concept to synthesize these findings with existing medical literature. This integration demonstrates a possible pathway for generating interpretable, patient specific reports. The gait pipeline obtained a cross-validated classification rate of 90.15% when subject-wisely partitioned, the most rigorous methodologically result of this study. The speech pipeline exhibited an exploratory segment level accuracy of 95.5% that is influenced by speaker-level leakage and is only given as a measure of feasibility. The subject-wise analysis produced a low estimate of 85.26%. Multimodal aggregation stage, which is built on pseudo-patient profiles (which combine predictions on independent datasets) attained about 92% under majority voting scheme and is reported as a proof-of-concept architecture implementation, not a validated multimodal performance metric. Overall, these results suggest that the fusion of diverse biomechanical data with knowledge graphs may provide a viable framework for enhancing clinical decision support and remote monitoring.