Neuroimaging Feature Extraction Meets Clinical Data Integration: A CNN-RF Hybrid for Parkinson’s Classification

Parkinson is a multifactorial chronic neurodegenerative disease that needs proper diagnosis and especially at the early stages of the disease. This work aims to develop a new Hybrid CNN-RF model that combines neuroimaging and clinical predictors in order to enhance the identification and categorization of PD. Specific biomarkers like cortical thickness and basal ganglia intensities which could be difficult for a CNN to pick up separately are extracted while patient demographics, UPDRS scores etc. are used by the RF for final classification. The proposed model was tested using a diverse dataset and produced high performance with an accuracy of 92.5 % and sensitivity and specificity of 91.2% and 93.8% respectively and an area under the curve of 0.95 only. The results of comparative analysis showed that application of the Hybrid CNN-RF model provides a higher accuracy (95.3%) compared to standalone CNN and RF models that demonstrated accuracies of 88.7% and 85.3% correspondingly. An ablation study also supported the above findings where if the neuroimaging features were removed, the accuracy was 84.7%; similarly, if clinical features were removed the accuracy was 85.3%. The results show the versatility of the model to incorporate additional information from neuroimaging and clinical data efficiently. Moreover, in the group of 20 epochs, the training accuracy and validations were also close and depicted a stable learning behavior of the model. In addition to providing evidence for the effectiveness of the Hybrid CNN-RF model in PD diagnosis, this research highlights the prospects of the multimodal paradigm for improving the diagnostic process. Future work will be based on the use of more comprehensive datasets and additional characteristics in an attempt to improve model portability.