Diagnosis And Prediction Of Parkinson’s Disease Based On Longitudinal Neuroimaging Data

Parkinson’s disease(PD)is one of the most common neurodegenerative diseases in the world.In general,patients with PD develop several types of non-motor symptoms(depression,sleep,olfactory,and cognitive impairment)and some motor symptoms.These symptoms are caused by degenerative death of midbrain substantia nigra dopaminergic neurons.Although this pathological change of PD is irreversible and there is still no cure for PD so far,longitudinal analysis with early diagnosis and clinical score predictions(depression,sleep,olfactory,and cognitive impairment)can make patients confirm their condition early and get timely treatment,which is vital to improve the quality of life of patients and delay the progress of the disease.At present,the computer-aided diagnosis researches of PD mostly exploit feature selection methods.These studies usually only screen the key features from single time point neuroimaging data and do not carry out further feature learning,nor do they carry out a unified longitudinal study on the early diagnosis and clinical score predictions.To solve these problems,this paper constructs a diagnosis and prediction model of Parkinson’s disease based on longitudinal neuroimaging data,which mainly includes the following three research points:First,in view of the fact that most of the existing PD researches exploit feature selection method and lacks feature learning,this paper constructs an optimized stacked sparse autoencoder model to perform feature learning on longitudinal neuroimaging data to complete early diagnosis tasks.Specifically,the number of samples of medical PD neuroimaging data is usually small,and the neuroimaging data of multiple modals contains high-dimensional features,which is easy to cause the problem of overfitting in deep learning models.To solve this problem,we optimize the autoencoder model of deep learning and combine it with sparse learning and constraints of coding weights.Experiments show that the deep features learned by the stack sparse autoencoder can improve the accuracy of longitudinal diagnosis of Parkinson’s disease.Second,there are no unified models for early diagnosis and clinical scoring prediction in existing longitudinal studies of PD.To address this problem,this paper constructs a unified model that combines longitudinal diagnosis and longitudinal prediction.Specifically,we continue to optimize the stacked sparse autoencoder through combining it with sparse learning and nonnegative matrix factorization idea.The optimized sparse nonnegative autoencoder can carry out sparsity on both features and coding weights,which can enhance the ability of feature learning and improve the generalization of the model.Experiments of longitudinal diagnosis and clinical score predictions(depression,sleep,olfactory,and cognitive impairment)show that this feature learning method is effective.Thirdly,both feature selection and feature learning have their own limitations.In order to fully combine the advantages of feature selection and feature learning,this paper optimizes the feature selection method based on relational regularization,and combines it with sparse nonnegative autoencoder to construct a unified model for longitudinal diagnosis and predictions of PD.Based on the relationship among samples and features,the model selects the key features related to the disease and then complete feature learning through sparse nonnegative autoencoder.Finally,this model uses the key information after feature learning to complete the longitudinal diagnosis and clinical score predictions of PD.In this paper,a preliminary study for longitudinal diagnosis and predictions of PD based on neuroimaging data is carried out.The experimental results demonstrate that our research can deal well with some key problems in current researches on PD.Extensive experiments on the longitudinal neuroimaging data also show that the unified model we finally construct outperforms state-of-art methods on PD.