Research And Application Of Multi-Task Feature Selection Algorithm For Alzheimer's Disease Diagnosis

Alzheimer's disease is a recessive neurodegenerative disease,which occurs in people over 65 years old.The process of onset is slow and irreversible.The main clinical symptoms are memory degradation and cognitive decline and may be accompanied by other physiological and psychological disorders.So far,there is no effective treatment to prevent further deterioration of the disease in the late stage.However,studies have shown that effective treatment in the early stage of Alzheimer's disease(mild cognitive impairment)can slow down the development of the disease and prevent its further deterioration.In recent years,the application of machine learning methods for recognizing and diagnosing neurodegenerative diseases has become a hot research direction.In particular,multi-task learning is widely applied in Alzheimer's disease study for it can simultaneously assess the disease severity(classification)and predict corresponding clinical scores(regression).Motivated by this idea,this paper discusses the adaptive multi-task dual-structure learning and sub-class personalized feature selection,which can effectively select the disease-development related discriminative features,namely biomarkers.The main contributions of this dissertation are summarized as follows:(1)A new adaptive multi-task dual-structure feature selection algorithm is proposed,which explores the manifold structure of the label space and regression space of disease data,and learns the adaptive similarity matrix and the corresponding feature mapping space between the two tasks at the same time.By coding the reconstructed label space and clinical scoring space,the similarity between disease data can be measured adaptively.The experimental results show that the proposed method can achieve promising results in both tasks.(2)The algorithm of subclass personalized feature selection is proposed.In clinical,different patients often show a high degree of personalization in the characteristic,although the disease severity of patients is classified at a similar stage.Therefore,how to capture the heterogeneity of patients for carrying out personalized predictive modeling has become an urgent problem.Aiming at this problem,we propose a novel personalized feature selection method,which models all instances and specific instance sub-class and determine shared and personalized features respectively.The experimental results show that this method has achieved good results for Alzheimer'sdisease dataset.