A Research On Classification Of Late-life Depression Based On Machine Learning

As depression has become the focus of social attention,diagnosing depression correctly in the early curable stage has become a hot and difficult point in medical research.For its unique advantages in detecting early lesions,such as large amount of imaging parameters,fast scanning speed,high tissue resolution and clearer images,MRI has become a powerful tool for early screening of tumors,heart diseases and cerebrovascular diseases.It has also been widely used in depression research.How to improve classification performance is always a research hotspot in using machine learning algorithms to screen depression.This paper draws on existing research results,making some improvements to existing problems in the application of traditional machine learning algorithms to the binary classification of elderly depression and healthy people.The main research contents are as follows:1.This article uses four classifiers including SVM-L(support vector machine linear kernel),SVM-RBF(support vector machine Gaussian kernel),Bayes(naive Bayes),RF(random forest),three feature selection methods including F-score,Corr,two-sample ttest to test the classification performance respectively.It is found that although the accuracy rates measured under different experimental conditions are similar,experiment using the SVM-RBF classifier has a higher accuracy rate but its sensitivity is 0.In order to reflect correctly the problem,this paper proposes a new index g for sensitivity and specificity.Compared with the accuracy,this index reflects the performance of the SVMRBF classifier more comprehensively.Due to the limited experimental samples,the classification performance of the existing feature selection method before and after the leave-one method is quite different.To solve the problem,this paper proposes a new feature selection method MKR.The experimental results show that the absolute difference of the average correct rate of the existing feature selection methods before and after is more than 10%,while the absolute difference of the average accuracy of using MKR is less than 5%.In addition,this paper studied the different brain regions between elderly depression and healthy subjects.2.Existing research shows that multi-modality can obtain better classification performance compared with single-modality,and multi-index perform better than singleindex.In order to find the optimal modalities and indicators to improve the detection accuracy and efficiency,this paper defines the contribution of each indicator to the classification based on the leave-one-out method of the support vector machine.The experimental results show that the functional connection indicator FC is more important than the gray matter volume data,the Re Ho indicator is more important than the ALFF indicator and the functional connection indicator FC.Based on this result,data of a certain modal can be scanned specifically or a certain index can be calculated predictively in the future,so as to obtain better classification performance with less time cost.In addition,this paper introduces the core fusion technology of support vector machines to obtain the contribution of each index data of the optimal performance.The results verify the above conclusions,and the most optimal classification result is obtained using this method,the correct rate is 94.59 %.3.At present,most studies on the optimal performance of depression classification adopt hypothesis-driven.This method is simple to implement and easy to operate,but its drawback is that it may ignore more valuable modal data combinations.In order to find the global optimal solution and improve the efficiency of the solution,this paper introduces genetic algorithm to study the classification of late-life depression.In order to verify the impact of feature selection on classification performance,this paper designs two chromosome structures.The results show that feature selection can help improve classification performance.At the same time,it was discovered during the experiment that optimal individual loss and the limitation of the quality by the first generation genes.To solve the problem,this paper proposes a method adding the optimal individual retention and gene mutation inheritance,the highest accuracy rate measured increased from 83.78% to 89.19 %。...