A novel hybrid CNN-KNN ensemble voting classifier for Parkinson’s disease prediction from hand sketching images
Authors: Saleh, S., Ouhmida, A., Cherradi, B., Al-Sarem, M., Hamida, S., Alblwi, A., Mahyoob, M. and Bouattane, O.
Journal: Multimedia Tools and Applications
eISSN: 1573-7721
ISSN: 1380-7501
DOI: 10.1007/s11042-024-19314-5
Abstract:Parkinson's disease is a progressive neurodegenerative disorder that causes significant physical disabilities and reduces the quality of life. This disease is caused by the loss of dopamine-producing cells in the brain. Its symptoms are Speech disorders, muscle rigidity, bradykinesia and tremors that cause involuntary shaking or trembling, typically starting in the hands, fingers, or limbs at rest. In this work, we focused on predicting this disease via the hand tremor, which appears in the speed and pen pressure that vary between healthy and affected people during sketching spiral and waves. To enhance the medical services, improve lifestyle and for early detection of people with Parkinson’s disease, we proposed an ensemble voting classifier that combines the Convolutional Neural Network and K-Nearest Neighbours (KNN) to decide whether or not a person has Parkinson's disease based on predicting spiral and wave sketching separately. Contrary to the traditional Convolutional Neural Network, the proposed architecture offers better flexibility in scenarios where data may be small and imbalanced to avoid overfitting or when capturing the nuanced relationships between data points (by considering their neighbours) can be beneficial. Moreover, the proposed system has been built to automate the extraction of features from images and perform classification. This work represents several approaches, such as image processing, developing Convolutional Neural Networks models, hyper-parameters tuning, transfer learning, feature extraction and developing a hybrid classifier that combines deep learning and machine learning to enhance the performance of the prediction. We have developed six models to predict Parkinson's disease using a Spiral-Wave dataset and provided a detailed explanation and comparison between their performances. Based on these models, we built the hybrid ensemble voting CNN-KNN classifier that reached 96.67% accuracy and 93.33% and 100% sensitivity and precision, respectively. This system demonstrates better performance compared to existing systems in the literature that predict Parkinson's disease based on hand tremors. Graphical Abstract: (Figure presented.)
Source: Scopus