Detect Cardio Respiratory Disorders Using Photoplethysmograph Biology Essay

Published:

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

In this project, Photoplethysmography(PPG) has been proposed as an alternative to investigate cardiac and respiratory condition instead of the commonly used ECG or respiratory signals. Photoplethysmography is a non-invasive technique that measures relative blood volume changes in the blood vessels close to the skin. Signal acquisition is carried out for healthy volunteers of a particular age group under normal conditions. Signal was acquired from reflection mode PPG sensor placed on the right index finger of the subjects. ECG was simultaneously acquired to analyze its variability with PPG. After initial pre-processing and data segmentation, several time and frequency domain features were extracted. ANalysis Of VAriance(ANOVA) was performed to select the best features for classifying normal from abnormal. Multi feature classification was carried out in MLP Neural network and yielded accuracy of about 95%. To study the performance of different classifiers, a comparison between Elman and MLP networks was performed for single feature classification. It was found that MLP has lesser time complexity and yields more accurate results. A comparative study between Elman and MLP was done using Receiver Operator Characteristics(ROC). SVM classifier was found to have the best classification accuracy for multiple feature classification. The significance of this work is that it can be converted to a real time system in future on integration with suitable PPG acquisition hardware to make it a complete automated decision making system for diagnosis of cardiac and respiratory disorders.

TABLE OF CONTENTS

CHAPTER NO. TITLE PAGE NO.

ABSTRACT

LIST OF TABLE

LIST OF FIGURES

1. INTRODUCTION

Introduction to PPG

Photoplethysmography Instrumentation

Feature Selection

Medical Expert System

Scope of the Project

Objectives

2. LITERATURE REVIEW

3. PROPOSED MEDICAL EXPERT SYSTEM FOR CARDIO-RESPIRATORY DISORDER DETECTION

3.1 Schematic Block Diagram

Description

Data Acquisition

Pre-Processing

3.2.2.1 Time Domain

3.2.2.2 Frequency Domain

3.3 Classification

3.3.1 Multi Layer Perceptron(MLP) Network

3.3.2 Support Vector Machine(SVM) Network

3.3.3 Elman Network

3.4 Receiver Operator Characteristics(ROC)

4. EXPERIMENTAL SETUP DETAILS

4.1 Subjects

4.2 Experimental Setup

4.3 Data Acquisition

4.3.1 Normal Condition

4.3.2 Respiratory Condition

4.3.3 Cardiac Condition

4.4 Statistical Analysis

5. PERFORMANCE OF CLASSIFIERS

5.1 MLP Neural Network

5.1.1 Single Feature Classification

5.1.1.1 Cardiac Condition

5.1.1.2 Respiratory Condition

5.1.2 Multi Feature Classification

5.1.2.1 Cardiac Condition

5.1.2.2 Respiratory Condition

5.2.1 Cardiac Condition

5.2.2 Respiratory Condition

5.2 SVM Neural Network

5.3 Elman Neural Network

5.3.1 Cardiac Condition

5.3.2 Respiratory Condition

5.4 Receiver Operator Characteristics Plots

5.4.1 Cardiac Condition

5.4.2 Respiratory Condition

6. CLASSIFICATION TECHNIQUES

6.1 Introduction

6.2 Multilayer Perceptron

6.3 Support Vector Machine

6.4 Conclusion

Appendix

Reference and Citations

LIST OF TABLES

4.1 Physical Characteristics of the subjects

4.2 Time domain features

4.3 Frequency domain features

4.4 ANOVA Analysis for frequency features

4.5 ANOVA Analysis for Time features

5.1 Description of MLP Network configurations for Single feature

5.2 Classification results for Testing data - Single Feature

5.3 Comparisons of the classifier performance for different features using MLP Network - Single Feature

5.4 Description of MLP Network configurations for Multi feature

5.5 Classification results for Testing data - Multi Feature

5.6 Comparisons of the classifier performance for different features using MLP Network - Multi Feature

5.7 Comparisons of the classifier performance for different features using SVM Network - Cardiac Domain

5.8 Comparisons of the classifier performance for different features using SVM Network - Respiratory Domain

5.9 Description of MLP Network configurations for Single feature

5.10 Elman Network Performance chart

5.11 Area Under the Curves (AUC)

LIST OF FIGURES

1.1 Instrumentation Block Diagram

3.1 Block Diagram

3.2 Filter Response

3.3 Typical PPG Signal

3.4 Typical frequency characteristics

3.5 MLP Architecture

3.6 Elman Architecture

4.1 Timing Diagram for Data Acquistion

4.2 Acquired PPG Signal

4.3 Peak Detection

4.4 Waveform under Different Conditions

4.5 Bar Graph

4.6 Bar Graph representation of Power Ratio(PR)

5.1 Cardiac Condition Single Feature Classifier

5.2 Cardiac Condition Single Feature Classifier

5.3 Multi-feature classification

5.4 Classification in the Cardiac Domain

5.5 Classification in the Respiratory Domain

5.6 Performance Evaluation(Cardiac)

5.7 Performance Evaluation(Respiratory)

5.8 Performance Comparison between Elman and MLP Network

5.9 ROC Curves

CHAPTER 1: INTRODUCTION

INTRODUCTION

A plethysmograph is an instrument or method used to measure the variations in blood volume within the body. Photoplethysmography(PPG) denotes the use of light to measure these changes in volume. This technique was introduced in 1937 by Hertzman. Although the human body is generally assumed to be opaque to light, most soft tissue will transmit and reflect both visible and near-infrared radiation. If light is projected onto an area of skin and the emergent light is detected after its interaction with the skin, blood and other tissue, time varying changes of light intensity having a relation with blood volume can be observed. This time varying light intensity signal will depend on a number of factors including the optical properties of the tissues and blood at the measurement site, and the wavelength of the light source. The signal results because blood absorbs light and the amount of light absorbed, and hence the intensity of remaining light detected, varies in relation with the volume of blood illuminated. Variation in the signal is caused by the variation in blood volume flowing in the tissue.

Changes in the blood volume can be caused by cardiovascular regulation, blood pressure regulation, thermoregulation and respiration. The plethysmogram is used to determine information on such parameters as pulse rate, breathing rate, blood pressure, perfusion, cardiac stroke volume and respiratory tidal volume. These can be observed as periodic and non-periodic changes in the amplitude of AC and DC components in the plethysmogram. It is also used to determine blood constituents. One such technique is pulse oximetry which determines the oxygen saturation in the blood.

The pulsatile component of the PPG waveform is the 'AC' component and usually has its fundamental frequency typically around 1 Hz. This AC component is superimposed onto a large quasi-DC component that relates to the tissues and to the average blood volume. This DC component varies slowly due to respiration, vasomotor activity, Traube Hering Mayer (THM) waves and also thermoregulation[1][2][3]. These characteristics are body site dependent[4]. With suitable electronic filtering and amplification both the AC and DC can be extracted for subsequent pulse wave analysis.

Photoplethysmography Instrumentation

Modern PPG sensors often utilize low cost semiconductor technology with LED and matched photodetector devices working at the red or infra-red wavelength[5]. LEDs convert electrical energy into light energy and have a narrow single bandwidth. They have a very long operating life, operate over a wide temperature range with small shifts in the peak-emitted wavelength, and are mechanically robust and reliable. The choice of photo-detector is also important [6][7]. Its spectral characteristics are chosen to match that of the light source. A photo-detector converts light energy into an electrical current. They are compact, low-cost, sensitive, and have fast response times. A high pass filter reduces the size of the dominant DC component and enables the pulsatile AC component to be boosted to a nominal 1 V peak-to-peak level. Filtering circuitry is also needed to remove the unwanted higher frequency noise such as electrical pick up from (50 Hz) mains electricity frequency interference. Excessive filtering can distort the pulse shape but too little filtering can result in the quasi-DC component dominating over the AC pulse [8].

There are two main PPG operational configurations: transmission mode operation where the tissue sample is placed between the source and detector, and reflection mode operation where the LED and detector are placed side-by-side. The PPG probe should be held securely in place to minimize probe-tissue movement artefact.

Fig. 1.1: Instrumentation Block Diagram

Feature Selection:

Many features have been investigated, including beat-to-beat PPG rise time, PTT, amplitude, shape, and the variability in each of these. The pulse shape can also be described after normalization in pulse width and height[9]. HRV exhibits frequency components from 0-0.5 Hz which are associated with the autonomic nervous system branches. Frequency components in the 0.15 -0.4Hz represent vagal tone and these frequencies are known as High Frequency (HF) components. Frequencies from 0.04-0.15Hz manifest the activation of parasympathetic and sympathetic nerves and are labelled as Low Frequency (LF) components. The ratio between LF and HF is defined as the Sympatho-vagal balance [10]. Pulse wave analysis: MATLAB (MathWorks Inc.) is a digital signal processing environment that is well suited to pulse wave analysis algorithm prototyping. It is known that PPG measurements are quite sensitive to patient tissue movement artifact. The automatic detection of such motion artefact, and its separation from good quality is a non-trivial exercise in computer signal processing. Computer-based filtering, feature extraction and waveform averaging have also been employed in PPG pulse wave analysis, including the analysis of frequency [11][12][13].

MEDICAL EXPERT SYSTEM:

Expert systems are most valuable for know-how experience and expertise. They are designed to carry the intelligence and information found in the intellect of experts and provide this knowledge for problem-solving purposes. The problems to be solved would normally be tackled by a medical or other professional. Real experts in the problem domain are asked to provide "rules of thumb" on how to evaluate the problems. Expert systems are used for problems for which there is no single "correct" solution which can be encoded in a conventional algorithm. Simple systems use simple true/false logic to evaluate data. More sophisticated systems are capable of performing at least some evaluation, taking into account real-world uncertainties, using such methods as fuzzy logic.

Advantages:

Provides consistent answers for repetitive decisions, processes and tasks.

Holds and maintains significant levels of information.

Encourages organizations to clarify the logic of their decision-making.

Can work continuously.

Can be used by the user more frequently.

SCOPE OF THE PROJECT

The use of Photoplethysmograph for clinical evaluation of the cardio-respiratory system provides information on:

Physiological parameters: blood oxygen saturation, blood pressure, cardiac output and respiration.

Vascular parameters: arterial compliance, arterial disease and aging.

The heart rate is an important physiological parameter to measure for a wide range of clinical settings, including hospital-based and ambulatory patient monitoring. The AC component of the PPG pulse is synchronous with the beating heart and therefore can be a source of heart rate information. Automatic assessment of the reliability of reference heart rates from patient vital-signs monitors using PPG based pulse measurements has been proposed [14]. Changes in pulse timing characteristics with breathing have also been studied. This includes using the PTT to track arousals during obstructive sleep apnea, and leading to a clinically useful non-invasive measure of inspiratory effort in patients with sleep-related breathing disorders[15][16]. Objective assessment of vascular ageing is very important since arterial stiffness is associated with hypertension, a risk factor for stroke and for heart disease.

OBJECTIVES

The main objectives of the work:

To create a database from subjects of a particular age group with no history of any cardiac or respiratory disorders.

To perform time domain analysis.

To perform frequency domain analysis.

To classify the acquired signal as normal or abnormal by using Artificial Neural Networks.

To design and develop an expert system to diagnose cardio-respiratory disorders using Photoplethysmograph.

CHAPTER 2 : LITERATURE SURVEY

S.NO.

TITLE

AUTHORS

ADVANTAGES

LIMITATIONS

1.

Analysis of Photo-plethysmographic Signals of Cardiovascular Patients

V.S. Moorthy, Sripad Ramamoorthy, Narayanan Srinivasan, Sriram Rajagopal, M. Mukunda Rao

Uses Power Spectral Density to evaluate Power Ratio.

Low and High frequency components are classified within the frequencybands 0.04-0.15Hz and 0.15-0.4Hz.

PSD is computed using Welch method. This evaluates atrial flutter and post MI patient respectively.

Heart Rate Variability(HRV) is used instead of Pulse Rate Variability(PRV).

Only single feature has been proposed for classifying normal and abnormal.

There is no specific diagnoses of what type of ailment patient is suffering from.

2.

Arterial Pulse Rate Variability analysis for Diagnosis

Aniruddha J. Joshi, Sharat Chandran

Proposes PRV as an alternative to HRV.

A correlation coefficient between several time and frequency domain features is found.

Power spectrum for various disorders such as stress, diabetes, BP etc have been found.

Further study is required to determine sensitivity, specificity and predict values of abnormal patients as compared to HRV.

Uses limited features for diagnoses.

3.

Contour Analysis of Photo-plethysmograph pulse

Sandrine C. Millasseau, James M. Ritter, Kenji Takazawa and

Philip J. Chowienczy

Pulse wave analysis is used to determine arterial stiffness.

Contour analysis is performed to determine arterial stiffness and vascular distension.

Time domain features such as Stiffness Index and Reflection Index have been analyzed.

Uses only time domain features.

Study doesn't relate to respiratory conditions.

4.

Non-invasive Studies on Age Related Parameters Using a Blood

Volume Pulse Sensor

Jayasree V.K, Sandhya T.V, Radhakrishnan P

Reflection Index was derived from parameters such as area under systolic peak and time difference between systolic and diastolic peaks

A relationship is obtained between time domain feature(Reflection Index) and age of the subject.

Range of values for Reflection Index has not been specified.

5.

Artificial Neural Network approach to PPG signal classification.

Jayasree V.K, Sandhya T.V, Radhakrishnan P

Multi Layer Perceptron model employ back propagation network to distinguish between normal and abnormal.

Linear time series analysis(ARMA) has been used for data preprocessing and dimensionality reduction.

94.7% classification accuracy has been found using MLP single feature classification.

Use of multi feature classification yields better accuracy.

Use of multi features allows us to use the best feature for classification.

6.

Support Vector Machine based Arrhythmia Classification using reduced features.

Mi Hye Song, Jeon Lee, Sung Pil Cho, Kyoung Joung Lee, and Sun Kook Yoo

The SVM was found to be superior when compared to other classifiers.

Purpose of SVM is to device an efficient way of learning good separating hyperplane in high dimensional feature space.

Computational time for SVM was found to be 100.094 seconds.

MLP can be used to classify multiple features but computational time is measured at 673.734 seconds.

SVM takes only a few fundamental features for classification.

CHAPTER 3 : PROPOSED MEDICAL EXPERT SYSTEM FOR CARDIO-RESPIRATORY DISORDER DETECTION

3.1 SCHEMATIC BLOCK DIAGRAM

FEATURE EXTRACTION

DATA SEGMENTATION

PRE-PROCESSING

ACQUISITION

TIME DOMAIN

FREQUENCY DOMAIN

CLASSIFICATION

Fig. 3.1: Block Diagram

3.2 DESCRIPTION

3.2.1 Data Acquisition

The PPG signal was acquired using BIOKIT Physiograph (Version 4.1 Build 3), TekSys Electronics. The sensor used is LED and LDR, optical transmit receive type finger sensor (wavelength 940nm) with input impedance of 1MΩ and a gain of Ã-5K maximum. Frequency response was recorded at 2-40Hz. Casing includes PCB mounted transmitter and receiver in a Velcro belt. The PPG signal was acquired from the right index finger. Subject is made to sit in an upright position with the forearm placed in a relaxed position on the thigh. Care is taken to reduce motion artifact due to respiration. After a short resting period for stabilization, the PPG data was acquired for 5 minutes. It condition about 30 minutes after food. During the procedure, the subjects breathed spontaneously at more than 12 cycles/min and the signals were recorded at 1000Hz sampling frequency. Room temperature was regulated at 28 degree Centigrade with humidity at 50%.

3.2.2 Pre-Processing

PPG has a lesser sophisticated morphology than other physiological signals and this also means peak detection of PPG relatively easy because there are few specific points. However, PPG could have an enormous baseline drift and wondering followed by physiological condition and movement, moreover it frequently happens. It was demonstrated that PPG contains fluctuation caused by respiratory and sympathetic activity, even arousal changes such as drowsiness causes PPG baseline wandering or drift [17]. These artifacts could be explained with the three major interferences of PPG, motion artifact, respiration effect and low perfusion. Baseline wandering can be removed using linear de-trending. Noise is removed by applying a digital FIR Band pass filter of the 8th order. Cut-off frequency ranges between 0.01-40Hz. This pre-processing helps increase the Signal to Noise (SNR) ratio. The data is then segmented in 15 second intervals prior to feature extraction. All these signal processing stages are implemented using MATLAB (The Math Works Co. MATLAB® version 7.0).

Fig. 3.2: Filter Response

3.2.3 Feature Extraction

3.2.3.1 Time Domain

Several features were extracted in the time domain for analysis of the normal PPG signal. The typical PPG signal is shown in the Fig.2 [18].

Fig. 3.3: Typical PPG Signal

Stiffness Index (SI) is a measure of the arterial dispensability and is used to find out age related problems such as arterial stiffness. The Stiffness Index is calculated from the body height divided by the time delay between the pulse systolic peak and the inflection point of the reflection wave (units m/s) [19]. The equation is given as follows:

The second time domain feature extracted is the Reflection Index (RI) [20]. The RI is found to fluctuate less under apnea condition. The equation for RI is given by:

3.2.3.2 Frequency Domain

Several frequency domain features were extracted for analysis of the PPG signal. Power Ratio(PR) is a bench mark parameter in evaluating the power distribution in the acquired signal[21]. The equation for PR is given as follows:

LF power is both a quantitative marker for sympathetic modulations and sympatovagal activity. HF power is considered to be a cardiac parasympathetic vagal nervous activity[22].

Fig. 3.4: Typical frequency characteristics

Some other features used for analysis are Mean Peak-to-Peak Interval (PPI), width in the LF and HF bands, Power Cent Ratio (PCR) in the LF and HF bands have been obtained.

3.3 CLASSIFICATION

The classification is done using several ANN's such as:

3.3.1 Multi Layer Perceptron(MLP) Neural Network:

A Multilayer perceptron is a network architecture in which has been formulated with two adaptive parameters, the scaling and translation of the post synaptic function at each node[23]. Artificial neural networks contain great number of processing elements which are connected to each other, the strengths of the connections are called weights. An MLP neural network is generally used for modeling the physical systems. It consists of a layer of input neurons, a layer of output neurons and one or more hidden layers. The optimal method of finding the number of hidden layers is by trail and error[24]. The training process is an important characteristic of the ANN methodology. There are number of training algorithms used to train a MLP and the most frequency used one is the back-propagation training algorithm.

Fig. 3.5: MLP Architecture

3.3.3 Support Vector Machines(SVM):

SVM is a supervised learning tool widely recognized for its ability to discriminator. The separation of data begins with a hyper plane and followed by extension of this s procedure to non-linear decision boundaries using the kernel trick [25]. Separating data into training and testing sets is the prime classification task.. Each instance in the training set contains one target value" (i.e. the class labels) and several attributes" (i.e. the features or observed variables). The ultimate aim of SVM is to produce a model (based on the training data) that can predict the target values of the test data using only the test data attributes [26].

3.3.4 Elman Recurrent Neural Network:

In a Recurrent Neural Network, connections between units form a directed cycle. An internal state of the network is thus created which enable it to exhibit dynamic temporal behavior. A recurrent network has feedback from each hidden node to all hidden nodes. Back propagation through- time is used for training the Elman network rather than the truncated version used by Elman.

An overview of Elman Network can be seen in the Fig 1.[27]

Fig. 3.6: Elman Architecture

3.4 RECEIVER OPERATOR CHARACTERISTICS(ROC):

A receiver operating characteristic (ROC) curve gives the relationship between the sensitivity and specificity of a clinical test for a variety of different cut points determination of an optimal cut point is thus made possible [28]. A particular point on the ROC curve corresponds to a specific threshold value which helps in enhancement of ROC curve plots in different ways. The 45-degree line is the most common enhancement. Along with reference lines inclusion of a few selected thresholds on the curve indicate the corresponding sensitivity and specificity [29].

CHAPTER 4 : EXPERIMENTAL SETUP DETAILS

4.1 SUBJECTS

The study was performed in 45 healthy, non-smoking and non-athletic volunteers (25 Male and 20 Female subjects) without symptoms of cardiac or respiratory diseases. Data was acquired at resting state for all subjects. 25 subjects were asked to perform breathe hold exercise for acquisition during respiratory condition. 20 subjects were asked to undergo physical exercise for 5 minutes prior to data acquisition. Hence respiratory and cardiac stress exercises were induced for evaluation of the respective conditions. Consent was obtained from all volunteers before data was acquired.

TABLE 4.1: Physical Characteristics of the subjects

PARAMETER

AGE(years)

BODY MASS INDEX(m/kg2)

MEAN ± S.D.

20.3839 ± 0.6978

21.0548 ± 1.724

VARIANCE

0.4869

2.9722

4.2 EXPERIMENTAL SETUP

The recording is carried out under postprandial conditions, about 30 minutes after food. The PPG electrode was mounted on the Velcro belt which was then attached to the right index finger. A resting period of about 5 minutes was allowed for the subject to stabilize. After stabilization, the recording was carried out for a period of 2 minutes. For recording under respiratory stress condition, the subjects were asked to undergo breath hold exercise. This induces apnea condition and gives an approximate measure of the changes in the time and frequency features for such a condition. Recording was carried out for a period of 30 seconds. This procedure was repeated under similar room conditions for accurate evaluation of the data acquired. For evaluation of cardiac stress condition, the subjects were asked to undergo rigorous physical exercise for a period of 5 minutes to induce increased cardiac activity. The same recording procedure was carried out and the accurate evaluation of time and frequency parameters is done.

SESSION 1:

16 Data segments

2 Trails

Task split considered for First recording

120 seconds normal PPG recording

120 seconds resting period

120 seconds normal PPG recording

1

Break for 300 seconds

SESSION 2:

4 Data segments

4 Trails

Task split considered for Second recording

30 seconds breath hold condition PPG recording

120 seconds resting period

30 seconds breath hold condition PPG recording

2

Break for 300 seconds

SESSION 3:

16 Data segments

2 Trails

Task split considered for Second recording

300 seconds cardiac stress exercise

120 seconds cardiac stress condition PPG recording

120 seconds resting period

300 seconds cardiac stress exercise

120 seconds cardiac stress condition PPG recording

3

Break for 300 seconds

Fig. 4.1: Timing Diagram for Data Acquisition

4.3 DATA DESCRIPTION

The acquired PPG signal supposedly contains cardiac and respiratory artifacts. Base-line wandering and motion artifacts also contribute a great deal to the noise in the raw PPG signal. The noise signal is removed using a digital band pass filter of order 8(Hanning window) with cut-off frequency ranging from 0.01-40Hz. This removes the artificats and gives an accurate measure of the frequency components present in the acquired signal.

(a)

(b)

Fig. 4.2: Acquired PPG Signal, (a) Raw PPG Signal, (b) Filtered PPG Signal

Peak detection is of utmost importance in calculating the time domain features such as peak to peak time(PPT) interval, the Stiffness Index(SI) and the Reflection Index(RI). The algorithm used in peak detection is shown below:

PSEUDOCODE

For PPG Peak Detection:

Initialize the Counter

FOR (counter values less than number of samples)

% For Peak 1 Detection

IF signal [current sample value] < set threshold

Peak not counted

END IF Loop

ELSE IF signal [current sample value] > set threshold

Peak is counted

END IF Loop

% For Peak 2 Detection

IF signal [current sample value] < set minimum threshold && > set maximum threshold

Peak not counted

END IF Loop

ELSE IF signal [current sample value] > set minimum threshold && < set maximum threshold

Peak is counted

END IF Loop

END FOR Loop

Using this algorithm, the anacrotic and catacrotic peaks are detected.

Fig. 4.3: Peak Detection

4.1.1 Normal Condition:

The visual inspection of the PPG waveform gives a clear understanding of how the Pulse Wave Analysis is significant for classifying clinical conditions. The frequency domain analysis using the Power Spectral Density(PSD) yields the distribution of power in the Lower and Higher frequency bands. The ratio between the power in these bands is an estimate of the type of clinical condition.

4.1.2 Respiratory Condition:

The PPG waveform under the respiratory conditions yields a reduce in the ratio between amplitudes of the two peaks. This is usually caused in apnea condition when the amount of oxygen saturation in blood is reduced. The change in the waveform also shows a change in the PSD of the PPG signal under the respiratory condition. The PSD reveals the presence of high power in the higher frequency bands, which ultimately results in the reduction of the Power Ratio under the Respiratory condition.

4.1.3 Cardiac Condition:

The PPG waveform under the cardiac condition shows a reduced time interval between the anacrotic and catacrotic peaks. This is caused due to vascular ditension during high blood pressure or heart rates. The PSD is studied to evaluate the Power ratio under the cardiac condition. The graph shows number of frequency components in the higher frequency band.

(b)

(c) (d)

(e) (f)

Fig. 4.4: Waveform under different Conditions (a) Normal PPG waveform and (b) Normal PSD

(c) PPG waveform under Respiratory Condition and (d) PSD of Respiratory Signal

(e) PPG waveform under Cardiac Condition and (f) PSD of Cardiac Signal

Tabulations:

TABLE 4.2 : Time domain features

PARA-METERS

NORMAL

RESPIRATORY CONDITION

CARDIAC CONDITION

Stiffness Index(SI)

6.8939±0.88

6.3653±1.351

7.9582±0.858

Reflection Index(RI)

0.5392±0.11

0.6883±0.173

0.6269±0.171

(b)

Fig. 4.5: Bar Graph (a) Stiffness Index(SI) and (b) Reflection Index(RI)

TABLE 4.3: Frequency Domain Features

PARA-METERS

NORMAL

RESPIRATORY CONDITIONS

CARDIAC CONDITIONS

Power Ratio(PR)

2.1070±0.317

1.01±0.083

0.6086±0.074

Width in LF Band

0.0651±0.101

0.1923±0.111

0.1781±0.122

Width in HF Band

0.0264±0.036

0.0755±0.003

0.0625±0.032

PCR in LF Band

0.1317±0.124

0.0175±0.010

0.0582±0.145

PCR in HF Band

0.0591±0.049

0.0217±0.013

0.0141±0.012

Fig. 4.6: Bar Graph representation of Power Ratio(PR)

4.3 STATISTICAL ANALYSIS RESULTS:

ANalysis Of VAriance (ANOVA) is performed to eliminate null hypothesis and non significant values. The comparison between normal vs cardiac and normal vs respiratory reveal the time and frequency features that are significant for the classification of normal and abnormal signals, the following table indicates the results of ANOVA analysis.

TABLE 4.4: ANOVA Analysis for frequency features

PARAMETER

NORMAL vs CARDIAC

NORMAL vs RESPIRATORY

Power Ratio(PR)

9.51 e-12

4.8 e-9

Heart Rate(HR)

6.37 e-7

0.1912

Peak Amplitude in HF Band

0.0019

0.0259

Peak Frequency in HF Band

0.4442

0.0483

Peak Amplitude in LF Band

0.3057

0.8187

Peak Frequency in LF Band

0.5585

0.3552

Width in HF Band

0.0016

6.49 e-5

Width in LF Band

0.0304

1.75 e-6

TABLE 4.5: ANOVA Analysis for Time features

PARAMETER

NORMAL vs CARDIAC

NORMAL vs RESPIRATORY

Peak to Peak Time (PPT)

0.04

0.034

Stiffness Index(SI)

0.034

2.89 e-5

Reflection Index(RI)

0.0003

0.0153

CHAPTER 5: PERFORMANCE OF CLASSIFIERS

Performance Evaluation:

5.1 Multi Layer Perceptron(MLP) Neural Network:

Classification results from single feature and multi-feature networks are as follows:

5.1.1 Single feature classification:

TABLE 5.1: Description of MLP Network configurations for Single feature

DOMAIN

HIDDEN

NEURONS

ACTIVATION FUNCTION

LEARNING

ALGORITHM

TRAINING

/

TESTING

EPOCHS

INPUT

OUTPUT

C

A

R

D

I

A

C

Power Ratio

22

Tan sigmoid

Pure Linear

Levenberg Marquardt

108/72

1000

Stiffness

Index

26

Tan sigmoid

Pure Linear

Levenberg Marquardt

108/72

1000

Reflection

Index

20

Tan sigmoid

Pure Linear

Levenberg Marquardt

108/72

1000

R

E

S

P

I

R

A

T

O

R

Y

Power Ratio

36

Tan sigmoid

Pure Linear

Levenberg Marquardt

108/72

1000

Stiffness

Index

26

Tan sigmoid

Pure Linear

Levenberg Marquardt

108/72

1000

Reflection

Index

34

Tan sigmoid

Pure Linear

Levenberg Marquardt

108/72

1000

5.1.1.1 Cardiac Condition:

Power Ratio (PR)

(b)

Stiffness Index (SI)

(c) (d)

Reflection Index (RI)

(e) (f)

Fig. 5.1: Cardiac Condition Single Feature Classifier

(a), (c) and (e) : Training States

(b), (d) and (f) : Performance Validation

5.1.1.2 Respiratory Condition:

Power Ratio (PR)

(b)

Reflection Index (RI)

(c) (d)

Stiffness Index (SI)

(e) (f)

Fig. 5.2: Respiratory Condition Single Feature Classifier

(a), (c) and (e) : Training States

(b), (d) and (f) : Performance Validation

TABLE 5.2: Classification results for Testing data - Single Feature

PARAMETER

POWER RATIO

STIFFNESS INDEX

REFLECTION INDEX

Cardiac

Respiratory

Cardiac

Respiratory

Cardiac

Respiratory

True Positive(TP)

72

69

64

63

50

60

True Negative(TN)

61

65

72

64

66

70

False Positive(FP)

11

5

0

6

6

0

False Negative(TN)

0

1

8

7

22

10

TABLE 5.3: Comparisons of the classifier performance for different features using MLP Network - Single Feature

FEATURE

SENSITIVITY (%)

SPECIFICITY (%)

ACCURACY (%)

C

A

R

D

I

A

C

Power

Ratio(PR)

100

84.7

92.36

Stiffness

Index(SI)

87.5

100

93.75

Reflection

Index(RI)

90.98

91.66

80.55

R

E

S

P

I

R

A

T

I

O

N

Power

Ratio(PR)

98.57

92.85

95.71

Stiffness

Index(SI)

90

91.4

90.71

Reflection

Index(RI)

85.7

100

92.85

5.1.2. Multi feature classification:

TABLE. 5.4: Description of MLP Network configurations for Multi feature

DOMAIN

HIDDEN

NEURONS

ACTIVATION FUNCTION

LEARNING

ALGORITHM

TRAINING

/

TESTING

EPOCHS

INPUT

OUTPUT

Cardiac

42

Tan sigmoid

Pure Linear

Levenberg marquardt

756/576

1000

Respiratory

46

Tan sigmoid

Pure Linear

Levenberg Marquardt

756/576

1000

5.1.2.1 Cardiac Condition:

(b)

5.1.2.2 Respiratory Condition:

(c) (d)

Fig. 5.3: Multi-feature classification (a) and (c) : Training States

(b) and (d) : Performance Validation

TABLE 5.5: Classification results for Testing data - Multi Feature

PARAMETER

CARDIAC

RESPIRATORY

True Positive(TP)

68

67

True Negative(TN)

69

67

False Positive(FP)

3

3

False Negative(FN)

4

3

TABLE 5.6: Comparisons of the classifier performance for different features using MLP Network - Multi Feature

SENSITIVITY (%)

SPECIFICITY (%)

ACCURACY (%)

Cardiac

94.5

95.8

95.13

Respiratory

95.7

95.7

95.71

5.2 SUPPORT VECTOR MACHINE (SVM) NEURAL NETWORK:

5.2.1 Cardiac Condition

(a) (b)

(c)

Fig. 5.4: Classification in the Cardiac Domain

Power Ratio vs Stiffness Index

Power Ratio vs Reflection Index

Stiffness Index vs Reflection Index

TABLE 5.7: Comparisons of the classifier performance for different features using SVM Network - Cardiac Domain

PARAMETER

SENSITIVITY (%)

SPECIFICITY (%)

ACCURACY (%)

PR vs SI

97.84

100

99.44

PR vs RI

96.72

100

98.33

SI vs RI

90.32

100

95

5.2.2 Respiratory Condition

(b)

(c)

Fig. 5.5: Classification in the Respiratory Domain

Power Ratio vs Stiffness Index

Power Ratio vs Reflection Index

Stiffness Index vs Reflection Index

TABLE 5.8: Comparisons of the classifier performance for different features using SVM Network - Respiratory Domain

SENSITIVITY (%)

SPECIFICITY (%)

ACCURACY (%)

PR vs SI

94.62

98.85

96.66

PR vs RI

94.62

100

97.22

SI vs RI

95.69

97.70

96.66

5.3 ELMAN NEURAL NETWORK

TABLE 5.9: Description of MLP Network configurations for Single feature

DOMAIN

HIDDEN

NEURONS

ACTIVATION FUNCTION

LEARNING

ALGORITHM

TRAINING

/

TESTING

EPOCHS

INPUT

OUTPUT

Cardiac

5

Tan sigmoid

Log

Sigmoid

Back

Propagation

108/72

5000

Respiratory

5

Tan sigmoid

Log Linear

Back

Propagation

108/72

5000

5.3.1 Cardiac Condition:

(b)

Fig. 5.6: Performance Evaluation(Cardiac) : (a) Power Ratio(PR) and (b) Stiffness Index(SI)

5.3.2 Respiratory Condition:

(b)

Fig. 5.7: Performance Evaluation(Respiratory) : (a) Power Ratio(PR) and (b) Reflection Index(RI)

TABLE 5.10: Elman Network Performance chart

CARDIAC

RESPIRATORY

POWER RATIO

STIFFNESS INDEX

REFLECTION

INDEX

POWER RATIO

STIFFNESS INDEX

REFLECTION

INDEX

ACCURACY (%)

92.85

94.28

72.91

95.71

89.58

82.75

(b)

Fig. 5.8: Performance Comparison between Elman and MLP Network

Cardiac Condition

Respiratory Condition

5.4 RECEIVER OPERATOR CHARACTERISTICS PLOTS

5.4.1 Cardiac:

(a) (b)

5.4.1 Respiratory:

(c) (d)

Fig. 5.9: ROC Curves

TABLE 5.11: Area Under the Curves (AUC)

TYPE OF NETWORK

CARDIAC

RESPIRATORY

Power Ratio(PR)

Stiffness Index(SI)

Power Ratio(PR)

Reflection Index(RI)

MLP

0.9412

0.9510

0.9701

0.9676

ELMAN

0.9210

0.9190

0.8863

0.9196

Writing Services

Essay Writing
Service

Find out how the very best essay writing service can help you accomplish more and achieve higher marks today.

Assignment Writing Service

From complicated assignments to tricky tasks, our experts can tackle virtually any question thrown at them.

Dissertation Writing Service

A dissertation (also known as a thesis or research project) is probably the most important piece of work for any student! From full dissertations to individual chapters, we’re on hand to support you.

Coursework Writing Service

Our expert qualified writers can help you get your coursework right first time, every time.

Dissertation Proposal Service

The first step to completing a dissertation is to create a proposal that talks about what you wish to do. Our experts can design suitable methodologies - perfect to help you get started with a dissertation.

Report Writing
Service

Reports for any audience. Perfectly structured, professionally written, and tailored to suit your exact requirements.

Essay Skeleton Answer Service

If you’re just looking for some help to get started on an essay, our outline service provides you with a perfect essay plan.

Marking & Proofreading Service

Not sure if your work is hitting the mark? Struggling to get feedback from your lecturer? Our premium marking service was created just for you - get the feedback you deserve now.

Exam Revision
Service

Exams can be one of the most stressful experiences you’ll ever have! Revision is key, and we’re here to help. With custom created revision notes and exam answers, you’ll never feel underprepared again.