• No results found

This chapter presents the various methods used for denoising, segmentation, and classifica- tion of PCG signal. Developing an intelligent CAD system for analysis of heart sound signals will be of immense advantage for the clinical diagnosis of CVD and to monitor the progress of undergoing treatment. PCG signal carries extensive information on the structural and functional activities of the heart. Any dysfunction of the heart is reflected in the PCG signal as an abnormal sound. Many of these anomalies are quantifiable using advanced signal processing and machine learning techniques. With the growing repository of information derived from the PCG signals, it is becoming more enriching and cognitive for developing an automated diagnostic system.

The reviews of the prevailing popular algorithms for the analysis of PCG signals are discussed in details in the following sections. In Section 2.1, the PCG database used for the evaluation of the thesis work is discussed. Section 2.2 presents the denoising process using wavelet decomposition and total variation filter. Both the denoising methods were considered to produce better noise reduction of PCG signals. Section 2.3 discusses some popular feature extraction methods for the analysis of heart sound signals. In Section 2.4, HSMM based heart sound segmentation algorithm is reviewed. The performance matrices that will be used for evaluating the thesis work is discussed in Section 2.5. The motivation of this thesis is presented in Section 2.6.

location of PCG files is [71]. They are available as the training set (a through f), in total 3,153 heart sound recordings from 764 subjects/patients [4].

An example dataset used in [5], which is downloadable from [72], available as training set-a by MIT [4, 18] is briefly discussed as follows. The Massachusetts Institute of Technology heart sound database (MITHSDB) is contributed by Syedet al [73]. The PCG signals are recorded simultaneously with an electrocardiogram (ECG) using a Meditron electronic stethoscope.

The sample rate of the recordings is 44.1 kHz with 16-bit quantization. There are a total of 409 recordings from 121 subjects. The length of the recordings ranges from 9 s to 37 s. The database comprises 117 recordings from 37 healthy subjects, 134 recordings from 37 patients with symptoms of mitral valve prolapse (MVP), 118 recordings from 34 patients showing innocents or benign murmurs (Benign), 17 recordings from patients with aortic diseases (AD), and 23 recordings from 7 patients that have other miscellaneous pathological conditions (MPC). The recordings are performed by either visiting the patients at their homes or in the hospital. As a result, many of the recordings are corrupted with various noise sources from the surrounding. Other noise sources include respiratory noise, intestinal sounds, and artifacts from stethoscope motions. This particular dataset provides corresponding ECG signals for validating the heart sound components. Therefore, it is used in this thesis to evaluate different proposed modules involved in developing the HSMM-based HSS algorithms. That includes the denoising process in Chapter 3, the envelope enhancement methods in Chapter 4 and the proposed duration models for the HSMM-based HSS algorithm in Chapter 5.

The above-discussed dataset lag details information to categorize heart sound murmur from other pathological PCG recordings. Therefore, to evaluate different diagnostic features for the classification of heart sound into categories: normal, noisy, and murmur, in Chapter 6, a separate dataset available in the ‘classifying heart sounds challenge’ sponsored by PASCAL [74] is used. The database by the PASCAL/CHSC2011 has been gathered from two sources: Dataset A is from the general public via the iStethoscope Pro iPhone app and Dataset B is from a clinical trial in hospitals using the digital stethoscope DigiScope [74]. But details of the subject or pathology associated with the signals are not specified. Dataset

Table 2.1: Profiles of PCG database.

Data Recording # Simultaneous Sampling

Database Set Category state Recordings signals rate Sensor

PhysioNet/ a (MITHSDB) Normal Uncontrolled 117 one PCG 44.1 kHz Meditron

CinC MVP environment 134 one ECG electronic

Challenge Benign (in-home 118 stethoscope

2016 AD visits or in 17

MPC hospital) 23

PASCAL/ A Normal Uncontrolled 31 one PCG 44.1 kHz iStethoscope

CHSC2011 Murmur environment 34 Pro iPhone

Extra heart- from general 19 app

sound public

Artifact 40

B Normal Clinical trial 200 one PCG 4000 Hz Digital

Murmur in hospitals 66 stethoscope

Extrasystole 46 DigiScope

A is categorised into four classes, namely: normal, murmur, extra heart sound and artifact.

Category wise, there are 31 recordings in the normal, 34 recordings in the murmur, 19 recordings in the extra heart sound and 40 recordings in the artifact category. Dataset B is subcategorised into normal, murmur and extrasystole. There are 200 recordings in the normal, 66 recordings in the murmur and 46 recordings in the extrasystole category. The recordings might have been collected from children and adults in both calm and excited states.

The heart rates may vary from 40 to 140 bpm [75, 76]. Some recordings are corrupted with background noise (from traffic to radio), physiological sound and artifactes from recording devices (movement of microphone, and connecting wires). In the murmur category, the types of murmurs are not specified. However, it contains a wide range of murmurs having temporal locations in systole or diastole intervals and with varying degrees of intensities. Some of the recordings are noisy due to background noise or distortion. The extra heart sound category of Dataset A is the collection of recordings containing additional sounds in the normal heart sound cycle. It may include split sounds or gallop sounds. The artifact category (of Dataset A) is a collection of a wide range of noise including feedback squeals and echoes, speech, music and background noise. There are no discernible heart sounds and purely meant to indicate possible noise signals. The extrasystole category (Dataset B) is an occasional occurrence

in which the heart sound is out of rhythm. There may be extra or skipped heart beats. The detailed profiles of the PCG databases discussed above are tabulated in Table 2.1.

2.1.2 Noise database

Various ambient noise is added into the original signal to simulate noisy PCG recordings from a noisy hospital-like environment. The noise signals are downloaded from the open web database Freesound (https://freesound.org). They are broadly categorized as (a) ventilation/air conditioner (AC) noise, (b) ambulance noise, and (c) hospital ambient noise.

There are nine recordings of AC noise type which may be affected by faulty leakage, crickets, broken fan, angered sound, and additional street and footstep noise. For ambulance noise, there are 13 recordings consisting of vehicle sounds, sirens, people talking, rain, and traffic sounds. There are 12 hospital ambient noise recordings consisting of crowd (nurse, doctor, patient) noise, coughing, door opening or closing, footsteps, humming sound of various machines, the beeping of heart machine recorded at hospital corridor, emergency room, and ICU noise in it. There is no discernible heart sounds in the noise signals. In the following chapters, these noises will be known as real-time ambient noise (RTAN).