Parkinson's disease (PD) is a catastrophic neurodegenerative disorder and a major culprit of neurological disability worldwide. Accurate diagnosis of PD, especially in its early stages, is paramount for timely intervention and effective therapeutic management. However, contemporary clinical diagnostic methods are hindered by the complexities of procedures, inter-evaluator subjectivity, and challenges related to the accuracy and reproducibility of diagnoses. This study employs transcranial sonography (TCS)-a non-invasive, widely accessible, and highly affordable neuroimaging-and integrates radiomic analysis with deep learning via a stacked multimodal ensemble model for PD assessment. A retrospective cohort of 324 participants (103 with PD and 221 Non-PD controls) was recruited from Beijing Tiantan Hospital. The development of the proposed model involved systematically evaluating its performance across different combinations of clinical, radiomic, and deep learning features. The results demonstrated that: The integration of multimodal features-clinical, radiomic, and deep learning-consistently improved predictive performance across all ensemble methods (Bagging, Boosting, and Stacking classifiers). The Stacking ensemble (RF, XGB and HGBDT) model achieved the best performance using the combined feature set, with an AUC of 0.915 (0.002), PR-AUC of 0.868 (0.005), accuracy of 0.810 (0.019), and precision of 0.968 (0.040), with values reported as mean (standard deviation). Diagnostic efficacy including TCS grading significantly exceeded that without TCS, underscoring its clinical value in PD assessment. Furthermore, decision curve analyses showcased that the stacked multi-classifier (RF, XGB, and HGBDT) yielded the greatest net benefit across a wide range of threshold probabilities compared with other comparing methods. These findings suggest that the proposed Stacking model may serve as a promising TCS-based multimodal tool for PD assessment.