<p>Diagnosing Parkinson’s Disease (PD) is challenging due to its progressive nature and the heterogeneity of motor and non-motor symptoms. Early detection—particularly of subtle non-motor indicators such as olfactory loss—can substantially improve disease management. Multi-modal analysis has emerged as a promising strategy for enhancing diagnostic performance by integrating complementary information from diverse data sources. However, acquiring complete multi-modal data for every patient is often impractical due to financial and logistical constraints. To address this limitation, we propose a novel multi-modal learning framework for PD classification using three distinct and unpaired datasets. The proposed approach follows a two-stage training strategy: first, modality-specific classifiers are trained independently; second, their learned representations are leveraged to train a unified multi-modal classifier using a Siamese network with a Triplet loss function. To enable inference on unpaired test samples, we introduce a “Mix-and-Match” pairing strategy, in which an unlabeled test sample from one modality is paired with pre-selected positive and negative samples from another modality. Predictions from multiple such pairings are aggregated using max-voting to produce a robust final decision. The effectiveness of the proposed method is validated using gait sensor data (PhysioNet), clinical data (PPMI), and speech data (UCI). Experimental results demonstrate that the proposed framework achieves over 99% classification accuracy in most scenarios and improves the AUC by 5 to 12% compared to single-modality baselines, consistently outperforming existing unimodal approaches.</p>