Integrating Markov Model, Bivariate Gaussian Distribution and GPU Based Parallelization for Accurate Real-Time Diagnosis of Arrhythmia Subclasses

In this paper, we present the integration of SIMT (Single Instruction Multiple Threads), Markov model and bivariate Gaussian distribution as a general-purpose technique for real-time accurate diagnosis of subclasses of arrhythmia. The model improves the accuracy by integrating both morphological and temporal features of ECG. GPU based implementation exploits concurrent execution of multiple threads at the heart-beat level to improve the execution efficiency. The approach builds a bivariate Gaussian Markov model (BGMM) for each subclass of arrhythmia where each state includes bivariate distribution of temporal and morphological features of each waveform and ISO-lines using ECG records for each subclass from standard databases, and the edge-weights represent the transition probabilities between states. Limited 30-second subsequences of a patient's beats are used to develop bivariate Gaussian transition graphs (BGTG). BGTGs are matched with each of the BGMMs to derive the exact classification of BGTGs. Our approach exploits data-parallelism at the beat level for ECG preprocessing, building BGTGs and matching multiple BGTG-BGMM pairs. SIMT (Single Instruction Multiple Thread) available on CUDA resources in GPU has been utilized to exploit data-parallelism. Algorithms have been presented. The system has been implemented on a machine with NVIDIA CUDA based GPU. Test results on standard MIT- BIH database show that GPU based SIMT improves execution time further by 78% with an overall speedup of 4.5 while retaining the accuracy achieved by the sequential execution of the approach around 98%.