Medical Signal Intelligent Classifications Based On Semi-Supervised Learning Algorithm

In the era of data,computing capability becomes a new productivity,and intelligent information processing technology based on machine learning algorithms shows high values in medical applications.This paper focused on the design and construction of automatic classification decision model in the field of physiological medical signal detection.From the perspective of data availability,machine learning algorithms can be divided into supervised learning,semi-supervised learning,and unsupervised learning methods.At present,the manual annotation program of medical data requires a lot of computing consumptions,and has serious noise interference.It is difficult to meet the application requirements of medical signal detection by fully supervised learning which needs a large number of labeled data or unsupervised learning which does not need labels but has poor classification performance.Therefore,combining unsupervised clustering idea with supervised learning algorithm,this paper proposed a novel type of semi-supervised learning algorithm to achieve higher medical signal classification performance under the condition of effectively using a small number of labeled medical data.This paper proposed a novel semi-supervised learning algorithm based on data density competitive optimization mechanism,and applied it to the field of medical signal pattern classification.The main process of the semi-supervised learning algorithm includes:(1)First,the sample points whose density is higher than the predefined threshold parameter are selected from the data set as candidate seed points,and the sample points with the maximum density are selected from the local region of the candidate seed points by competitive learning method as candidate seed points;(2)Then,taking the candidate seed point as the initial centroid,the K-means unsupervised clustering is carried out to obtain the cluster centroid after clustering;(3)Finally,the cluster centroid is used as the data sample points,and some labeled data sets are used as the training set for supervised learning classification.All the sample points in the cluster have the same label with the cluster centroid,and the semi-supervised learning result is obtained.In order to verify the classification effect of the proposed semisupervised learning algorithm,we applied it to three different medical signal pattern recognition fields.In the application of speech signal classification in Parkinson’s disease,we made correlation analysis on the features of speech signals,calculated the Pearson linear correlation coefficient in each parameter family,then used PCA to reduce feature dimension,and selected the acoustic features with significant difference for pattern analysis by p-value of Mann Whitney Wilcoxon hypothesis test.The results showed that the semi-supervised learning algorithm proposed in this paper outperformed several conventional machine learning tools such as k-nearest neighbor and support vector machine,in the overall classification accuracy(0.838),recall(0.825),specificity(0.85)and accuracy(0.846)on the speech signal data set.In the detection of knee joint VAG signals,we first used the cascaded sliding filter and the integrated empirical mode decomposition method to eliminate baseline drift and random noise carried by the original signal;then,we extracted the time-domain fluctuation characteristics of VAG signals,such as envelope amplitude mean,envelope amplitude standard deviation,envelope amplitude root mean square,turn counts,and calculated the symbol entropy and permutation entropy.The result was that the proposed semi-supervised learning algorithm achieved the overall classification accuracy of 0.904,the sensitivity of 0.889,the specificity of 0.909,and the area under the ROC curve of 0.945,all of which were better than the results of traditional classifiers like Fisher linear classification,quadratic discriminant classification and generalized logistic regression.In the sleep phase detection applications,we first normalized the EEG signals,then used t-SNE method to reduce the dimension of energy proportion features in Alpha,Beta,Theta and Delta frequency range,and finally carried out pattern classification.Experimental results showed that the proposed semi-supervised learning algorithm based on density modeling competitive learning can still achieve 86.2%classification accuracy under the condition of 21%labeled training sample set,and the results were better than strong learning algorithms such as SVM and back propagation neural network.The semi-supervised learning algorithm based on seed density competitive learning mechanism proposed in this paper constructed intelligent decision-making systems for medical signals.As a supplement to the research on medical pathological detection technology and physiological pathological mechanism,it can provide a new technical path for speech signal screening of Parkinson’s disease,pathological detection of VAG signal of knee-joint cartilage wear,and detection of sleep stage.