PARKINSON’S DISEASE DETECTION BASED ON DEEP TRANSFER LEARNING USING OPTIMIZED FEATURE SELECTION

Abstract

Parkinson's Disease (PD) diagnosis remains challenging due to the absence of definitive clinical tests, 
particularly in its early stages. This study addresses the critical need for an effective and non-invasive 
methodology for early PD detection by leveraging deep learning, specifically Convolutional Neural Networks 
(CNNs), to analyze handwriting patterns. Various models including ResNet50, VGG19, InceptionV3, and 
Xception are employed for feature extraction, with K-Nearest Neighbors (KNN), Random Forest (RF), Support 
Vector Machine (SVM), and Decision Tree used for classification. The proposed ensemble method combines 
predictions from multiple models, enhancing accuracy. In the base model, ResNet50 + VGG19 + InceptionV3 
with KNN achieved 95% accuracy. As an extension, further exploration of ensemble techniques, including 
Voting Classifier, is conducted, aiming for 98% accuracy or higher. Additionally, a front end using Flask 
framework is developed for user testing, incorporating user authentication. This research contributes to 
advancing early PD detection, crucial for prescribing timely treatment and improving patients' quality of life. 
INDEX TERMS Parkinsons disease, neurological disorder, handwritten records, transfer learning, deep 
learning.