An Integrated System for Unbiased Parkinson's Disease Detection from Handwritten Drawings

Current Parkinson's disease (PD) diagnosis relies on a series of hospitalbased clinical examinations. To enable early PD detection at home, recognition from hand-written drawings is one way for automated PD detection system. However, existing methods have two main problems i.e., biasedness and lack of generalization in independent testing. The biasedness problem is due to two factors. The first factor is subject overlap between training and testing datasets caused by conventional validation methods. The second factor is imbalanced classes. In this paper, to avoid biasedness in the constructed modelswe utilize a balanced handwritten images. To avoid biasedness due to subject overlap, we use a more robust cross validation scheme i.e., leave one subject drawings out. In order to develop a decision support system to generalize to unseen data, we use several feature driven systems, and integrate F-score based feature selectionmodel with those systems. Experimental results show that integration of F-score based model with Gaussian Naive Bayes model is a good candidate for PD detection based on hand-written drawings. It yields PD detection accuracy of 71.21% on main dataset and 63.04% on another dataset during independent testing.