Deep Learning Meets Traditional Machine Learning: CNN-SVM Hybrid Models for Parkinson's Diagnosis

• Published in: 2025 International Conference on Automation and Computation (AUTOCOM) pp. 1402 - 1406
• Main Authors: Mehta, Shiva, Bhardwaj, Rishabh
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.03.2025
• Subjects: Accuracy; Action; Convergence; Convolutional neural networks; Data models; Feature extraction; Health diseases Workout; Hybrid power systems; Parkinson's disease; Patient Action Identification; Sensitivity; Strokes; Support vector machines; SVM; Training; Vectors
• DOI: 10.1109/AUTOCOM64127.2025.10956635

Parkinson's disease (PD) is a neurodegenerative disorder that needs early and accurate diagnosis for good management. Our approach in this study is to have a hybrid diagnostic framework which uses Convolutional Neural Networks (CNNs) for the feature extraction, and Support Vector Machines (SVMs) for the classification. By leveraging on the complementary strengths of the above-mentioned approaches, the model reports improved performance in detection of PD by using medical imaging data. We evaluated proposed CNN-SVM hybrid model on a dataset consist of 5,000 MRI images (2,500 with PD cases and 2,500 with healthy control) with achieved accuracy of 94.8%, sensitivity of 93.7%, and specificity of 95.5%. These results demonstrate a significant improvement over baseline methods, including standalone CNNs (accuracy: 91. The performance of the proposed methods is experimentally demonstrated to be better than SVMs with handcrafted features (accuracy: 84.5%, sensitivity: 83.2%, specificity: 85.0%) and even SVMs using Fisher vectors (accuracy: 81.4%, sensitivity: 86.1%, specificity: 81.1%) for certain classes (e.g. 2%, sensitivity: 89.5%, specificity: 92.3%). The class separability of the hybrid model was excellent, with an Area Under the Receiver Operating Characteristic Curve (AUC−ROC) of 0.96. These findings were further confirmed under confusion matrix analysis with 2,343 true positives and 2,388 true negatives and with only 157 false negatives (sensitivity: 0.017%) and 112 false positives (specificity: 0.970%). The model demonstrated generalizability by the convergence of training and validation curves throughout 50 epochs with a final loss of 0.06 and validation error stabilizing at 0.04. The CNN-SVM hybrid model also can demonstrate the ability to become a reliable diagnostic tool for PD detection, with robust classification performance and few errors. Multi-modal data integration and scalability for use over a broader range of clinical applications is the subject of future work.