Time-frequency based automatic classification of normal and pathologic heart saound recordings

Cardiovascular disorders or heart diseases that identify general status of heart are main reason of increasing death in the world. Easy, cheap and early diagnosis of heart diseases has potential to help to improve life expectancy of people. For 50 years, many experts have tried to develop diagnostic algorithms that can diagnose heart diseases by using heart sounds. Heart sounds recordings that are called as phonocardiogram (PCG) are recorded by an electronic stethoscope and have important information about mechanical activity of heart. However, algorithm developed so far are tested on data sets that contain few clean recordings obtained from only one body location. Therefore, it is necessary to develop a heart disease diagnostic algorithm with high performance and to test on a data set which contains both clean and noisy recordings obtained from different body locations. In this thesis, it is tried to develop a machine learning and signal processing based pattern recognition algorithm that can diagnose heart diseases of people. Developed algorithm is tested on a database that contains 3240 PCG recordings. In this work, an algorithm is proposed for heart disease diagnosis. Proposed algorithm consists of seven main phases which are preprocessing, feature extraction-1, data set classification, segmentation, feature extraction-2, feature selection and final classification. In preprocessing phase, PCG recordings were denoised from spikes and normalized. In next phase, features that are used to train and test classifier segregating recordings with respect to data sets which they are collected from are extracted. Classifier ensembles which created with AdaBoostM2 and decision trees are used to categorize recordings according to their data sets. Thus, it is enabled to use different features and classifiers for data from different databases. In segmentation phase of algorithm, Springer's segmentation algorithm is used to divide PCGs into fundamental heart sounds. Various time, frequency and time-frequency features are extracted from segmented and non-segmented signals so two feature vectors are created per signal in feature extraction-2 phase. After number of features is reduced by feature selection algorithm, extracted features are handled for training and test of classifiers in the last phase of algorithm. Since there are two classifiers per signal, five different voting rules are used for final classification in order to find rule achieving best performance results. After best rule is determined, best results (sensitivity: %87.22, specificity: %97.28 and accuracy: %95.21) are achieved when this rule is used for voting. Then, the proposed algorithm is compared with similar works with respect to performance results and proposed algorithm appears to be successful.