| Electrocardiogram(ECG)abnormality detection is a typical muti-label classification problem,which is usually realized by training a binary classifier for every abnormality.However,the performance of automatic detection is not acceptable because of the large number of abnormality labels and the complex correlations between abnormalities and features.The frequent co-occurence of abnormality labels reflects the label correlations,so we propose to study the automatic abnormality detection based on the frequent labelset patterns,thus taking advantage of the correlations of abnormality labels.We focus on the following two aspects.On one hand,we notice that the public features are the characteristics of a frequent labelset pattern,which can be used to distinguish different label groups while the private features are the ones whose values can uniquely identify a specific label.To find the the correlations among different ECG abnormality labels,we employ the Frequent-Pattern-Growth method to construct the frequent labelsets.This is achievd by the construction of the specific feature space for both frequent labelsets and individual labels.Specifically,we use Dirichlet Clustering to cluster the samples to build the supporting axis of the specific feature space.Then,we get the public feature of frequent-itemset and private feature of individual labels.On the other hand,the number of possible abnormality combinations grows exponentially with the number of labels,but most of them can be captured by the frequent labelsets.We propose a novel K-nearest-Neighbor detection method based on the label correlations.Sepcifically,we construct the scoring model and the threshold model for the frequent-itemset and single label,respectively.The scoring model is used to measure the correlation degree between a sample and a label or a frequent labelset.The threshold model is used to set the correlaton threshold corresponding to frequentlabelsets or individual labels.Based on the two types of scoring models and the corresponding thresholds,sample abnormality sets are predicted.Our main contributions are as follows.First,we propose an algorithm for extracting the public feature and private feature based on Dirichlet Clustering.Specifically,we use Dirichlet Clustering to cluster the samples to construct the supporting axis of feature space.Then,we get the public feature of frequent-itemset to distinguish different label combinations effectively.Second,we propose a novel K-Nearest-Neighbor detection method based on the label correlation.This method can exploit the different correlation degrees among labels to predict the relevant labelset of a sample.Consequently,it overcome the pitfalls of the improper use of label correlations for the multi-order classification model and the ignorance of label correlation for the single-order classfication model. |
