Optimal Neuro-Fuzzy Classification for Arrhythmia Data Driven System

This paper presents a data driven system used for cardiac arrhythmia classification. It applies the Neuro-Fuzzy Inference System (ANFIS) to classify MIT-BIH arrhythmia database electrocardiogram (ECG) recordings into five (5) heartbeat types. In fact, in order to obtain the input feature vector from recordings, a time scale method based on a Discrete Wavelet Transform (DWT) was investigated. Then, the time scale features are selected by applying the Principal Component Analysis (PCA). Therefore, the selected input feature vectors are classified by the Neuro-Fuzzy method. However, the ANFIS configuration needs mainly the choice of an initial Fuzzy Inference System (FIS) and the training algorithm. Indeed, two clustering algorithms which are the fuzzy c-means (FCM) and the subtractive ( SUBCLUST) algorithms, are applied to generate the initial FIS. Besides, for tuning the ANFIS membership function and rule base parameters, Gradient descent and evolutionary training algorithms are also evaluated. Gradient descent consists of the backpropagation (BP) method and its hybridization with the least square algorithm (Hybrid). However, the evolutionary training methods involve the Particle Swarm Optimization (PSO) and the Genetic Algorithm (GA). Therefore, eight (8) ANFIS are configured and assessed. Accordingly, a comparison study between their obtained Root Mean Square Error (RMSE) is analyzed. At the end, we have selected an optimal ANFIS which uses the SUBTRUCT algorithm to generate the initial FIS and the GA to tune its parameters. Moreover, to guarantee the effectiveness of this work, a comparison study with related works is done.