Automatic Classification of Parkinson's disease patients vs Healthy controls using a vision-based finger-tapping test

In recent years, there have been remarkable advancements in machine learning. The rise in computational power and the abundance of data have made these systems indispensable in various fields, such as disease follow-up. Parkinson's disease (PD) is the second most prevalent neurodegenerative disorder worldwide. The most notable symptoms of PD are bradykinesia, rest tremor, postural instability, and rigidity. In clinical practice, a variety of diagnostic techniques, including the finger-tapping test, the walk test, the gait analysis, and the evaluation of speech impairment, are used to examine these symptoms. According to the Movement Disorder Society's published guidelines, the finger-tapping test (FTT) is one of the most frequently administered evaluations of bradykinesia, but manual visual evaluations can result in score discrepancy due to clinicians' subjectivity. Moreover, the application of wearable sensors necessitates physical contact and may inhibit the natural movement patterns of PD patients. The information related to these patterns was provided in the vision-based 3D Parkinson's disease (PD) hand dataset, consisting of 133 finger-tapping video samples, including recordings of 74 PD patients and 59 healthy cases, also completely annotated by qualified clinicians in clinical settings. This endeavor has considered the movement of the index and thumb digits toward and away from one another. Then, Python libraries were used to acquire the distance and velocity signals of these movements. Utilizing the tsfresh library for feature extraction was the next step. Identifying and deriving meaningful features from time series is a time-consuming process because scientists and engineers must take into account the numerous algorithms of signal processing and time series analysis. Then to remove redundant features and improve the classification algorithm's accuracy, feature selection based on extracted features has been implemented. Methods of Boruta and Principal component analysis (PCA) were applied to select meaningful features. Furthermore, several supervised machine learning classification algorithms have been evaluated to determine whether or not a person has PD. These algorithms include k-nearest neighbors, random forests, eXtreme Gradient Boosting, and support vector machine. These algorithms were selected based on their ability to correspond to medical evaluation criteria, their visualization capabilities, and the data size and computation constraints of real-world applications. A k-fold cross-validation method was exploited to verify the final classification accuracy in order to quantitatively compare the performance of distinct classifiers. In fact, two feature selection methods and four machine learning classifiers were combined to work out accuracy, precision, recall, and f1-score. These results would lead to the evaluation of the optimal combination method for diagnosing PD.