Research On Sparse Multi-task Learning Algorithm Of Alzheimer's Disease Prediction

Alzheimer's disease(AD)is an important disease that threatens the health of the elderly,with a prevalence that is surpassed only by cardiovascular disease,cerebrovascular disease,and tumors.AD is a chronic neurodegenerative disease that usually starts slowly and worsens over time,presenting significant problems in our aging society.With the rapid expansion of the aging population,Alzheimer's disease has become a serious social and economic problem worldwide.At present,although significant advances have been made to understand the neurobiological mechanisms of Alzheimer's disease,there is still no effective method to cure or control the progress of the disease.However,studies have shown that early detection and intervention can delay progression of the disease.The detection and analysis of biomarkers are effective methods for the early diagnosis of Alzheimer's disease,which is of great significance for the prevention and control of the disease and the detection of its development.Therefore,the accurate prediction of Alzheimer's progress and the effective identification of biomarkers for detection in the earliest stage are key targets of current research efforts.This dissertation describes methods based on the multi-tasking learning framework to analyze clinical AD data quickly and efficiently.A clinical decision model was constructed by integrating ideas,such as how to incorporate domain knowledge into a structured sparse learning approach,task correlation analysis,time series tracking and nonlinear distribution,screening and mining in the course of the biomarkers from different levels and different stages for Alzheimer's disease patients with inherent law,and then enhance sensitivity to early diagnosis of Alzheimer's disease and the accuracy of the development trend.The dissertation consists of the following:(1)Neuroimaging characteristics of high dimension features and group structural information exist in the feature regression coefficients used to predict cognitive results will affect the computing performance,and could lead to the relative prediction error estimation and recognition.In this work,we simultaneously exploit the interrelated structures within the features and among the tasks,and present sparse group Lasso multi-task learning methods to effectively incorporate both the relatedness among multiple cognitive score prediction tasks and the useful inherent group structure in features.The proposed method can reduce the dimensionality and identify related biomarkers.A large number of experiments were conducted on Alzheimer's Disease Neuroimaging Initiative(ADNI)datasets.This allowed verification of prediction performance and biomarkers identification from different dimensions and indicators to demonstrate the effectiveness of our method.(2)Existing multi-task feature learning methods assume a uniform correlation among all the tasks,and task relatedness is modeled using a common subset of features by sparsity-inducing regularizations that neglect the inherent structure of tasks and image features.Instead,we assume that if some tasks are correlated,they should have small similar weight vector and similar selected brain regions.In this work,based on the structure of the graph,we proposed a fused group lasso regularized multi-task feature learning method.For features across tasks and the local task structure with respect to brain regions,the method models a common representation to capture the task-level and the feature-level underlying structures.This work included all the cognitive measures in the ADNI dataset(20 in total)to exploit the relationship.To the best of our knowledge,our approach is the first work on analysis and exploitation of all the cognitive measures in the ADNI dataset and their relationships.(3)Most existing models do not model correlation among multiple cognitive scores over time.Including neuroimaging measurements of cognitive scores from longitudinal data that tracks continuous time points of disease progression,a multi-task learning method with time smoothing property based on time series is proposed.Considering the tasks required to accurately predict a given same cognitive score over multiple time steps(longitudinal analysis),a more generalized weighted dependency graph was constructed.A large number of experiments demonstrate that the proposed algorithm not only showed the highest prediction performance,but also demonstrated the ability to accurately identify imaging biomarkers that are consistent with prior knowledge.(4)Most existing works formulate prediction tasks as linear regression problems,in which a linear relationship is assumed to exist between the features and the cognitive outcomes.However,this simplistic approach is unable to capture the complex relationships between brain images and the corresponding cognitive measures.To address these shortcomings,we developed two multi-kernel multi-task learning methods with a joint sparsity-inducing regularization.By exploiting and investigating the nonlinear relationship.between measures and cognitive scores,our methods allow the cognitive scores to be modeled as nonlinear functions of neuroimaging measures and the multiple kernel learning methods can learn the optimal combination of given base kernels.Experiments on ADNI dataset demonstrate that the nonlinear regularized multi-kernel learning methods not only achieved better prediction performance than the state-of-the-art competitive methods,but also allowed the effective fusion of the multi-modality data.