A fuzzy support vector machine approach for ECG analysis

Abstract

In this thesis, fuzzy approach is used for ECG analysis. The ECG dataset in the UCI database is used. This dataset consists of inputs from normal and abnormal ECGs. All the anomalies are used to construct one class and normal ECG data is used to construct another class. The main purpose of the system is to detect anomalies correctly. In order to achieve this goal, a fuzzy support vector machine is constructed. Five different fuzzy membership functions are tested to reach the best performance: OCW, DTCM, DTOCM, CAR and FCM. Output of the fuzzy support vector machine system is compared to other classification methods. Results show that the fuzzy support vector machine outperforms other methods. In order to interpret the classification model, rule base extraction methods are applied. C4.5, PART, RIPPER and ANFIS are the selected algorithms for ECG rule base generation. When accuracy is considered as performance metric, RIPPER method outperforms the other techniques. The advantage of using ANFIS is the membership function generation for the features in the dataset. The resulting membership functions are found to be consistent with medical knowledge.