We aim to develop a reliable and robust algorithm that accurately analyses a single from a single precordial location to determine the presence of heart abnormality for the Physionet/ Computing-in-Cardiology 2016
IntroductionWe aim to develop a reliable algorithm that accurately analyses a single short PCG recording (10-60s) from a single precordial location to determine the presence of heart abnormality.We have considered features based on the statistical parameters derived from the analysis of the PCG recordings in the time and time-frequency domains. We segmented each PCG sequence into four states, namely, first heart sound (S1) and second heart sound (S2), Systole and Diastole using Hidden Markov Model based Springer's improved version of Schmidt's method [8]. Different features like standard deviation of S2 intervals, mean value of ratio between wavelet coefficient energy of Diastole period and RR in each heart beat are derived and overall 54 features are selected that represent a broad spectrum of energy, frequency, and temporal signature of typical PCG signal. Further, using minimum Redundancy Maximum Relevance (mRMR) algorithm, top 5 features are identified from the initial pool of 54 features. Then, non-linear radial basis function based Support Vector Machine (SVM) classifier is trained with the reduced set of optimal feature set, where balanced training set of 630 normal and 630 abnormal PCG datasets are considered. We also used ensemble based supervised learning with bagging and boosting techniques to handle the skewed class distributions. The trained SVM classifier and ensemble based classifier have been tested with the validation datasets as well as uploaded to test with hidden PCG datasets [1].It is also noted that detection of pathological conditions based on machine learning techniques is flawed by a) variety of datasets b) presence of outside noise in the datasets and poses a greater challenge to detection of abnormality. We also detect signal quality to handle this problem.
State-of-the-artSeveral articles are available in the literature on the analysis of the PCG signal. The segmentation of the heart sound is the first step to analyze the signal and for the identification of systolic and diastolic regions allowing for the detection of pathological events. Various existing literature are available for the segmentation of the PCG signal. Authors in [7], have applied an ensemble empirical mode decomposition (EEMD) based method combined with kurtosis features to detect the location of first (s1) and second (s2) heart sound. In [8] authors have proposed a