Automatic Diagnosis Of Parkinson's Disease Based On Voice Signal And Hand Drawing

Parkinson's disease(PD)is mainly diagnosed by the neurologists' subjective evaluation of the patient's symptoms based on the Unified PD Rating Scale,which intends to causing a high misdiagnosis rate.Some wearable devices have been developed to assist in providing accurate diagnostic results,but this process requires the patient and the doctor to cooperate,which is extremely inconvenient for PD patients with limited mobility.Based on previous reports,the PD patients will show certain dysphonia and dyslexia early.Therefore,to provide a non-invasive and convenient diagnostic solution,this study is devoted to develop the diagnosis methods for PD based on the speech and hand-drawing data,as follows:Firstly,we developed a PD diagnosis scheme based on public speech data.The original voice dataset is with 752 sets of feature data,which is a huge amount of data for the classification model.However,not all features show a good correlation with PD,so to extract the optimal feature subset,we designed a feature extraction method based on AdaBoost.Noteworthily,too many weak classifiers can retain more feature information,but also contain more low-correlation features,and the corresponding program processing time is longer.Nevertheless,too less weak classifiers can extract the features with higher correlation with PD,but missing information is also a major concern.Therefore,we designed an optimal scheme for weak classifiers to balance the feature information retention and model processing time cost.Additionally,to achieve the PD diagnosis with superior generalization performance,we developed XGBoost based on the regularization loss function to achieve the final diagnosis.We compared the designed scheme with some advanced models.The proposed scheme achieves optimal performance(Accuracy = 97.28%)and is superior to previous reports.Then,we designed a PD diagnosis scheme based on a hand-drawing dataset,which contains spiral and meander hand-drawings.We provided the diagnostic solutions based on Convolutional neural networks(CNN)and Capsule neural networks(CapsNet).Differentfrom traditional hand-drawing preprocessing schemes only with trajectories features,we retained the color information for the hand-drawing images,which provides the possibility for the subsequent models to learn information such as drawing pressure,speed,and time based on color information.Compared to CNN,due to dynamic routing and vectorization for the feature processing methods,the CapsNet can capture the relative position difference of images.Based on the analysis of the results,the CNN is a more reasonable solution for diagnosing PD based on the meander hand-drawing,while the CapsNet is a better solution for the spiral hand-drawing.It is also worth noting that the CapsNet based on the spiral hand-drawing provides the best performance for diagnosing PD based on hand-drawing,and the diagnostic accuracy of CNN is better than the corresponding reports of the previous binary preprocessed images.Overall,three conclusions will be reported: 1)Hand-drawings with color information are more conducive to high-precision diagnosis of PD;2)the spiral hand-drawing covers more differences than the meander hand-drawing to distinguish PD patients from healthy people;3)the relative position difference of the spiral hand-drawing is significantly more than that of the meander hand-drawing.Overall,the two diagnostic schemes proposed in this study only require the patient to provide a hand-drawing or speech data to confirm whether the patient is with PD,which is of great significance for the convenient diagnosis of PD.