Internship

The effects of electrocardiogram noise on ultra-short-term heart rate

variability indices.

First, we will aim to simulate different types of noise that interfere with
PPG signals (which measure volume changes in tissues).

1er script: We take a defined time range initially, which we then vary to
obtain an ECG signal that we can use for further analysis.

2nd script: This part allows us to extract the interval between two
heartbeats. We do not consider the first peak for subjects 2, 3, 4, 5, 6, 7,
8, 9, and 10.

Interest time: 2 for subjects 6 and 9

Pan –Tompkins : To simulate a band-pass filter, high-pass and low-pass

filters are applied in cascade. Next, differentiation is used to highlight the
R-peaks using a digital differential filter. The squaring function is
necessary to amplify the output signal of the derivative filter and to make
it positive. Finally, the signal is integrated to more easily determine its
amplitude.

3 script: White noise (random noise) : Used to assess the robustness of

the program and to develop better corrections under real-world
conditions.
Which peak is removed :

dB -3 1 5 10 20

Suject
1 1 1 1 1 White
1

1/163 1 1/166 1/166 1/166 Hz

dB -3 1 5 10 20

Suject
X 140 1/140 X X White
2

1/142 1/140 1/139 1/139 1/139 Hz

dB -3 1 5 10 20

Suject
1/153 X 1/160 1/163 1/161 White
3

156 1/157 1/160 1/161 1/161 Hz

dB -3 1 5 10 20

Suject 4 1/157 1/160 1/163 1/164 1/164 White

1 1/158 1/161 1 164 Hz

dB -3 1 5 10 20

Suject
1/141 1/138 X 1 1/134 White
5

1/139 1/137 1/133 1/133 1/133 Hz

dB -3 1 5 10 20

Suject
1/148 X 1 1/148 1/148 White
6

1/148 1/148 1/148 1/148 148 Hz

 dB -3 1 5 10 20

Suject
1/131 1/128 1/127 1/121 1 White
7

1/135 1/128 1/124 1/124 1/120 Hz

dB -3 1 5 10 20

Suject
1/142 1/141 1/142 1/142 1/142 White
8

1/143 1/142 1/142 142 1/141 Hz

dB -3 1 5 10 20

Suject
1/127 1/127 1/123 1/123 1/123 White
9

1/133 X 1/123 X 1/123 Hz

dB -3 1 5 10 20

Suject 10 1/134 1/130 1/130 1/130 1/130 White

1/131 1/131 1/130 1/130 X Hz

4 Script : Frequency noise, (constant noise) at different levels

HRV = Heart Rate Variability (Variat de Hz cardiaque) to know : Temporal

Variability of Hz heart rate

- Activity of SNA

- Study the heal and stress

Ultra-Short Term HRV

- Analyzing variability in heart rate on short period (5 min)

- Sensibility at noise and measures errors

ECG = Electrocradiographic Noise

-Sources : EMG muscular noise, electric interference, electrodes

movement

- Can fake the R peak ( peaks les + élevés du complexe QRS ) -> Calculs
of HRV

Time-Domain HRV Metrics

- SDNN écart-type

- RMSSD moyenne quadratique des intervalles RR

- PNN50 pourcentage d’intervalles RR diff > 50 ms

Frequency-Domain HRV Metrics

- Low Frequency : < 0.15Hz SNA

- High Frequency : 0.15 Hz - 0.4 Hz respiration

Effect of Noise on HRV Metrics

-Random noise : increase instability and fake the tempo measures

-Constant noise : fake the spectral analysis and HF/LF

-Wrong peak R

Noise Reduction Techniques :

- Digital Filtering
- R-peaks detection
- Correction of abnormal RR interval errors

Study of the P wave: This represents atrial depolarization, which is often

hidden by the ventricular wave.

Study of the QRS complex: This corresponds to ventricular depolarization

on the PR segment.

Study of the T wave: This reflects ventricular repolarization on the ST

segment.
Study of the U wave: Following the T wave, it is inconsistent and not
always present.

To reduce muscle noise interference, signal averaging and segmentation

methods are used.

The use of Pan & Tompkins is crucial for extracting an ECG signal, as it
allows for filtering the signal to retain only the relevant data. The first four
steps involve preprocessing before reaching the decision stage, which
determines the error rate and the detection rate of peaks within the
transformed signal.

Context and Importance: Wearable devices, such as ECG and PPG

recorders, are used to analyze physiological regulation but data analysis is
complex due to challenges related to noise.

Heart Rate Variability (HRV): HRV is a key indicator of cardiovascular

health, measuring the balance between the sympathetic nervous system
(SNS) and parasympathetic nervous system (PNS). High HRV is generally
associated with better cardiac health, while low HRV is linked to adverse
cardiovascular outcomes.

Analysis Methods: The study uses 2-minute ECG windows to evaluate HRV.
Algorithms : Pan-Tompkins to detect R-peaks + extract R-R intervals.

Noise Analysis: Different types of noise (respiratory, atrial fibrillation,

power line interference, muscular noise) are simulated at diff signal-to-
noise ratio (SNR) lvls to assess their impact on HRV measurements.

Statistical Analysis: With Wilcoxon test ( Non-parametric statistical tests )

are performed to evaluate how noises affect HRV measurements by
comparing noise-free and noise-contaminated signals.

Frequency Indicators: Contributions of different frequencies to the signal

power are analyzed to extract measures of power in the low and high-
frequency bands.

Information-Domain Analysis: Indexes quantifying the information content

and complexity of the time series are also evaluated to provide additional
insights into the effects of noise.
Time series : A sequence of data points collected or recorded at regular
time intervals. Each point represents the state of variable at a specific
moment.

Composition of cardiovascular signal :

Onde P Complexe QRS Onde T Intervalle RR, PR Segment ST

Experience : ECG were collected from 10 volunteers

Gaussian white noise and colored noise at specific frequencies ( 0.01- 300
Hz) at various SNR levels ( -3, 1, 5, 10, 20 dB)

Pan-Tompkins Algorithm : identify R peaks et extract R-R intervals time

series

- Time-Domain HRV with noise-free and noise-corrupted ECG signals

to assess the variability of interbeat intervals
- Frequency-Domain analysis involved computing the power spectral
density, using the weighted covariance method
 - Infromation-Domain was performed to compute indexes quantifying
the information content and complexity of the time series.

Impact of noises :

- White Noise : Random noise with constant power across all

frequencies Affects high frequencies, more harder to detect R-
peaks accurately, Using low-pass filters to reduce the effect.
- Pink Noise : 1/f noise and non-uniform frequency distribution, in low
frequency Can cause baseline wander, affecting accurate
measurement of segments.
- Brown noise : In very low frequencies
Causes large baseline shifts, mimicking exaggerated respiration,
disrupts interval measurements
- High-Frequency noise : from muscle activity or electrical
interference Distorts the shape of fine waves, can be
filtered but at high intensity, it compromises waveform analysis.
Research

Different techniques for recovering RR peaks.

Different types of signals interfere with the ECG :

- Power line interference (first filtered with a low-pass filter, then a

Notch filter)
- Baseline distortion interference
- Use of a differential amplifier to immunize the signal against noise
and interference.

Pan & Tompkins:

- Low-pass filter (cutoff at 6 Hz defined at 3 dB, centered at 8 Hz)

- Derivative: First-order derivative calculation to extract the slope
- Squaring the derivative: Rectification and emphasizing large
amplitudes
- Sliding integrator: Detects the energy of the derivative by properly
adjusting the sample for analysis
- Adaptive threshold comparator: Determines the detection rate and
error rate of the algorithm

This highlights the QRS waves.

P&T and Hilbert algorithms = reduces the error rate over time.

Wavelet transform: Allows separating the components of a signal into

multiple frequency bands to obtain various waves for analysis.
Empirical Mode Decomposition Algorithm:
A method for analyzing nonlinear and non-stationary signals. EMD
decomposes a complex signal into a set of elementary components called
IMFs.

It decomposes a signal into low-frequency and high-frequency

components.

Algorithm for detecting heartbeats in noisy electrical signals:

- Use of a band-pass filter, squaring stage, followed by a low-pass

filter, and finally a decision stage to detect pulses.
- Clipping to eliminate fluctuations caused by noise (removing the
signal below the moving average).
- Decision block: Determines if the pulse is a heartbeat by examining
the slope width.
- Tachogram: Graph showing the evolution of the heart rate.
- Use of a tachogram to estimate errors in detection accuracy

Adding white noise to the ECG signal to define the signal-to-noise

ratio (SNR). A band-pass filter is also applied.

Use of heart rate = a complex dynamic system subject to external

variations.

Procedure:

- Retrieve empirical data from an electrocardiogram.

- Extract the R-R intervals presented as a time series.

- Calculate the recurrence matrix containing information about recurrence

in the biological system.

- Develop a recurrent program that allows for analyzing the nature of the
process in the system: the presence and influence of noise, as well as
repetitive states and stagnation.

- Quantitatively calculate recurrence measures to determine the number

of measurements based on the compilation of recurrent density points.

-When taking measurements : Select the appropriate point for measuring.

Time Series : A finite sequence of quantitative data indexed by time.

Time Series Analysis :

 - Use of Fourier transform to obtain a sine wave at its maximum size.
- Use of the ARIMA model to analyze cycles, which is a more efficient
way to model a wide variety of time series.

ECG is primarily discrete and often periodic. It consists of trends,

seasonality, and noise.

Seasonality = Repetition of cardiac cycles.

Preprocessing = Low-pass filtering followed by band-pass filtering.

Anomaly detection = Simple analysis of R-peaks to distinguish them from

false peaks.

Analysis in time domain adverts to high correlation between statistic and

geometric parameters and similarly with signal HRV.

Results of frequency domain analysis shows similarly power spectral

density, which was calculated by two different ways.

Implement some signals processing theorems to analyze the data and use
MatLab to verify the theories

Noise has higher frequency than the true signal, use low filter and a
pseudo-Gaussian

To find the legitimate peaks, eliminate false peaks with the theory of
detection and estimation. To determinate the power spectral density

Using wavelet transform to denoising white Gaussian noise and power

noise removal

Multi – Resolution Analysis ( MRA ) filterbank by wavelet transform and

threshold-denoising will be used to remove noise from the ECG signal.

Wavelet transform : Wavelet function to extract information in both

frequency and time domain. Like Fourier transform but uses different
basis functions.

MRA : observes different time and frequency resolution via scaling and
wavelet functions. Noise can be removed by wavelet thresholding or
wavelet skrinkage.
Denoising white noise: Coiflets 5 wavelets for the filter bank. Synthetic
ECG signals + random noise and baseline wandering. Threshold levels are
set to remove noise white preserving ECG components

For human ECG + muscle noise : Threshold levels are adjusted to remove
muscle noise while preserving ECG signals

Raw ECG data is used for power noise reduction due to high sampling rate
requirements

MRA filter bank = remove power noise and soft thresholding techniques
are compared for removing harmonics of power noise.

All cardiovascular noise during ECG : white noise, muscle contractions,

baseline wandering and power noise.
Acquisition de signaux avec capteur piézoélectrique + préamplificateur +
oscillo

Traitement de signaux par FIR : filtre adaptatif à réponse impulsionnelle

finie + algorithme des moindres carrés récursifs (RLS) ou des moindres
carrés moyens (LMS) + un filtre variant dans le temps (TVF) par
transformation de Gabor + filtre de seuillage des ondelettes.

Utilisation de l’EA pour détecter les signaux.

Adaptation du codeur selon le signal, on compresse sans perte

Entropie de l’information : mesure du désordre et d’incertitude.

Si la valeur augmente alors incertitude augmente + probabilité de prévoir

l’information utile diminue

Le codage par arithmétique est un modèle fixe et modèle adaptatif mais

la compression est faible si aucunes pertes

Development of new heart pulse detection technology

Detecting pulses by signal processing of a raw ECG signal registered from
4 electrodes located on the upper left arm. Implemented that may affect
an electrical development equipment and methods of how major parts of
the interference can be reduced by different shield.

Heart has four chambers : the upper 2 chambers called the atria and thin-
walled, low-pressure pumps that receive blood from the venous
circulation.

The first ECG wave of the cardiac cycle is the P wave, which represents
activation of the atria. Conduction of the cardiac impulse proceeds from
the atria through a series of specialized cardiac cells.

Fourier analysis represents signals as linear combination of sine and

cosine waves and therefore the representations are localized in frequency.

MRA : Give good time resolution and poor frequency resolution at high
frequencies and good frequency resolution and poor time resolution at low
frequencies.

With MATLAB :

- Designing templates for signals that need to be simulated

- Development of the patient/ subject simulator

- Display of the simulated signals in either the analog or digital form

Create a basic pattern : Each of the simulated signals has a template.

The QRS complex in the ECG signal corresponds to the depolarization of

the ventricles.

On wavelet filtering the original respiratory waveform, the third level

approximation was chosen to form the template. The multi-resolution
analysis by the biorthogonal wavelet got rid of noise present in the
original signal.

The dZ/dt signal were simulated simultaneously using their respective

templates.
Wavelet based technique can also be extended to accomplish data
compression and feature extraction for each of the simulated
physiological signals ( popularity in both stationary and not stationary )

Estimation paramétrique à partir du signal bruité. Transmission

d’information sur un canal bruité.

Segmentation des signaux cardiaques

Méthode directe : pas besoin d’ECG et s’applique au signal ECG.

Méthode indirecte : besoin de l’ECG, après l’intervalle QRS (tps) il y a

utilisation de la méthode LSC directe car il ne nécessite pas de source
d’information supplémentaire sur le patient.

- Extraction de l’enveloppe d’amplitude : maximiser la distance entre

les sons cardiaques et les sons de bruits

Utilisation de la transformation mathématique :

- Transformation temporelle et morphologique

- Transformation fréquentielle
- Transformation temps-fréquence
- Transformation sur complexité du signal

Puis filtre de lissage pour éliminer les composantes à hautes fréquence

Seuil et Algorithme de détection des maxima locaux sont sur l’enveloppe

du signal.

Via domaine temporelle : Entropie de Shannon : Energie de Shannon

Enveloppe de Shannon ( moyenne de l’énergie de Shannon )

Transformation fréquentielle

Filtrage homomorphique : traitement d’images ( logarithme non linéaire à

linéaire )

Filtre passe bas de Tchebychev avec fréquence de coupure de 20 Hz.

Valeur maximale de l’enveloppe est détectée.

Utilisation de l’enveloppe de Hilbert : permet d’obtenir l’enveloppe

analytique d’un signal donné
Transformation Temps-Fréquence

- Produit multi-échelles

Localisation de régularité locale d’une fonction. La transformée

d’ondelette discrète du signal pulmonaire est calculée par ondelettes
mère

Transformée sur la complexité du signal

Dimension fractale est mesuré physiquement, donnant une idée sur la

complexité géométrique du signal

Entropie est la mesure statistique issue de la théorie de l’information.

Spectre singulier de la matrice de l’état des phases reconstruit.

Les spectres d’amplitudes et de la phase par la FFT peuvent être visualisé

autour d’un échantillon précis.
Il faut prendre en compte les fenêtres de pondération due aux ordinateurs
qui imposent une durée limitée à la transformée de Fourier.

On le superpose à la fenêtre temporelle & pondération

- Largueur de base du pic centrale

- Rapport d’amplitude du 1er lobe secondaire à celle du pic centrale
qui est exprimé en “dB”

MatLab : Opérations usuelles sur 1024 échantillons successifs

Utilisation du Spectrum Scope pour déterminer le bruit ( fréquence ) et on

observe sa FFT. Un filtre RIF est un système linéaire invariant discret dont
le comportement entrée-sortie est caractérisée par les coefficients hi de
sa réponse impulsionnelle.

Filtre à moyenne mobile : opération à réaliser, pour passer à l’instant

suivant, on décale d’une période d’échantillonnage à la fenêtre
d’acquisition ( la moyenne pondérée est mobile )

Traitement des signaux biomédicaux réels ECG-EMG

- Composition basique ECG

- Prétraitement afin d’avoir des signaux adéquats et utilisables.

Utilisation de la base de données Daisy ( DAtabase for the Identification of

SYstem )

Utilisation de la méthode de la transformée en ondelettes discrètes ( basé

sur des filtres à réponse impulsionnelle finie (FIR) + filtre passe bas, filtre
passe-haut puis un sous-échantillonnage par deux

Permet d’atténuer à la fois bruit et ligne de base servant à prétraiter les

données cardiaques utilisées

Technique : Décomposition du signal ECG en ondelettes, Seuillage sera

appliqué sur les coefficients détails des L premiers niveaux de la
transformation en ondelettes ainsi qu’une mise à zéro.

Donc la décomposition, la mise à zéro des approximations, seuillage des

détails et reconstruction
Analyse de la variance : Savoir si la différence significative entre le
traitement de plusieurs groupes de données.