A hybrid two-stage SqueezeNet and support vector machine system for Parkinson’s disease detection based on handwritten spiral patterns

• Published in: International journal of applied mathematics and computer science Vol. 31; no. 4; pp. 549 - 561
• Main Authors: Bernardo, Lucas Salvador, Damaševičius, Robertas, De Albuquerque, Victor Hugo C., Maskeliūnas, Rytis
• Format: Journal Article
• Language: English
• Published: Zielona Góra Sciendo 01.12.2021; De Gruyter Poland
• Subjects: Artificial neural networks; Classifiers; convolutional neural network; deep learning; Digital imaging; Feature extraction; Handwriting; Machine learning; Neurological diseases; Parkinson's disease; spirography; Support vector machines; Training
• ISSN: 1641-876X; 2083-8492
• DOI: 10.34768/amcs-2021-0037

Parkinson’s disease (PD) is the second most common neurological disorder in the world. Nowadays, it is estimated that it affects from 2% to 3% of the global population over 65 years old. In clinical environments, a spiral drawing task is performed to help to obtain the disease’s diagnosis. The spiral trajectory differs between people with PD and healthy ones. This paper aims to analyze differences between handmade drawings of PD patients and healthy subjects by applying the SqueezeNet convolutional neural network (CNN) model as a feature extractor, and a support vector machine (SVM) as a classifier. The dataset used for training and testing consists of 514 handwritten draws of Archimedes’ spiral images derived from heterogeneous sources (digital and paper-based), from which 296 correspond to PD patients and 218 to healthy subjects. To extract features using the proposed CNN, a model is trained and 20% of its data is used for testing. Feature extraction results in 512 features, which are used for SVM training and testing, while the performance is compared with that of other machine learning classifiers such as a Gaussian naive Bayes (GNB) classifier (82.61%) and a random forest (RF) (87.38%). The proposed method displays an accuracy of 91.26%, which represents an improvement when compared to pure CNN-based models such as SqueezeNet (85.29%), VGG11 (87.25%), and ResNet (89.22%).