Medical time series classification using global and local feature extraction strategies

Abstract

Objective: Present a method to improve the accuracy of the time series classification task, as well as to enable the interpretation of its generated model. Method: Features were extracted from time series combining two strategies: the global strategy, which uses statistical and complexity descriptors; and the local strategy, which uses the motif representation. In the next step, the data was submitted to three different learning algorithms in order to create classification models. The performances of the models were evaluated in terms of mean error rate using five medical datasets. Results: fFr all datasets, the best classification accuracy was obtained combining both local and global strategies. The approach improved the performance of the J48 algorithm, which generates a more interpretative model. The comparison among 1-NN, MLP, and J48 shows no significant statistically difference. Conclusion: The method aims at an enhanced descriptive power for time series data and increasing the performance of the models.