Hybrid Resampling and Xgboost Prediction Using Patient's Details as Features for Parkinson's Disease Detection

Abstract

The recognition/diagnosis of Parkinson's disease must be highly accurate to reduce the severity of the disorder with timely treatment. It is often seen that handwriting of the patient diminishes because it is tough to hold the pen/pencil due to muscle rigidity as the disease progresses. Men and women are neurologically different and so are the young and aged and thus respond differently to Parkinson's manifestation. Additionally, there is a significant link between the dominant hand of the person and the side of the body where the initial manifestation of the disease begins. This lays the foundation for research-based on gender, age and handedness (lateralization) to predict the disease. The HandPD dataset used here is inherently imbalanced. This gives rise to the issue of prediction model biasedness. The true nature of such a model is not quite revealed by the conventional accuracy alone. Thus, balanced accuracy is used to evaluate true efficiency. The technique proposed here alleviates model bias using hybrid resampling and extreme gradient boosting. It also explores the impact of features like age, gender and handedness on the mode efficiency. Experimental results of the technique proposed here yield the highest accuracy of 98.24%, a balanced accuracy of 98.14% with 100% sensitivity and 96.29% specificity when the age of the person is taken into account along with features extracted from the handwritten images.