Early Diagnosis Of Parkinson's Disease Using Machine Learning

As a common neurological disease,the incidence of Parkinson's disease is relatively high.At the same time,from the actual situation,Parkinson's disease will lead to the loss of the body's motor performance,which is classified as a movement disorder by clinicians.At present,the diagnosis of Parkinson's disease mainly depends on clinical symptoms,which largely depends on the experience of clinicians.Therefore,an effective early diagnosis method is necessary.In view of the characteristics of high dimension data and small data size of brain images,the method of feature screening is often used to select the optimal feature subset from it to achieve the effect of dimensionality reduction and prevent overfitting.However,the traditional machine learning algorithm will filter out many features with strong internal correlation in the process of feature dimensionality reduction and classification,which will cause the performance of the classification model to decrease.In order to solve this problem and ensure the high correlation between features while reducing dimensionality to achieve the purpose of improving classification performance,this paper proposes a recurrent network classification model based on self-attention mechanism.The model uses a self-attention mechanism combined with a recurrent network to classify and train five brain region features(brain gray matter,white matter,cerebrospinal fluid density,etc.)extracted from MRI and DTI images.At the same time,it also comparatively analyzed the other five classic classification algorithms commonly used in brain diseases(support vector machine,multi-layer perceptron,extreme learning machine,etc.).This study took 10-fold cross validations,and conducted experiments and analysis on 403 subjects,and divided the subjects into three groups,namely the normal group,Parkinson's group and special group,of which the normal group was 154 Group,Parkinson Group is 165 groups,special group 84 groups.Category research and analysis based on multimodal data can more effectively ensure the accuracy of experimental results and classification results.The experimental results show that: multiple image features are more beneficial to improve classification accuracy;the effect of multi-layer perceptron and convolutional neural network is better,the effect of extreme learning machine is not ideal most of the time,and the best performance is based on self-attention Recurrent neural network classification model of mechanism;the classification results obtained in this experiment have better classification performance than other widely used methods at present.