Detection of Parkinson disease by using keystroke data

Parkinson's Disease (PD) is a neurological movement disorder that occurs in the hands and feet with tremor, rigidity, slowing of movements and difficulty walking and postural instability. Generally; the measures from force sensors, accelerometers and inertia measurement units used to gain informations about gait, posture and disorderly movements have been studied for analyzing the PD's characteristics. In the last few years, due to the superior aspects of the current diagnostic methods, based on human-computer interaction; evaluation of data obtained from keystroke dynamics of keyboard use of Parkinson's patients has gained importance. In this study total 14 features, including asymmetry, entropies, high degree momentums and statistical quantities were calculated from datas and have been studied to determine the PD. All these significant features, statistical tests and Random Forest algorithms were used and applied to the inputs of two-class Support Vector Machines (SVM) and k Nearest Neighbor (kNN). The classifier accuracies were both found for training and testing in terms of 50-50 & 30-70 respectively. These results are listed for both 646 and 515 records. The obtained features were evaluated in four different cases. In all cases, the highest test accuracy is 80,15% (training: 83,78%) for 646 records and 82,4% (training: 86,64%) for 515 records, found by kNN classifier. These results shown that keystroke datas are able to used for PD diagnosing instead of other sensor measures. KEYWORDS: Parkinson's disease, keystroke dynamics, feature selection, human – computer interraction, classification Advisor: Dr. Derya YILMAZ, Baskent University, Department of Electrical and Electronics Engineering.