• Aucun résultat trouvé

Signals acquisition and preprocessing

Chapter 3 Physiological signals recording and processing

3.2 Signals acquisition and preprocessing

3.2.1 Signal acquisition

All the signals were acquired at a sampling rate of 1024 Hz. This high sampling rate was chosen to have high time resolution EEG signals and it was applied to the other signal sources as well because they were plugged on the same acquisition system. At that sampling rate the bandwidth was of 268Hz due to an antialiasing filter that is applied to the signals by the Actiview system before storage on the hard drive. The complete bandwith was not used in the present studies;

nevertheless it allows performing high frequency EEG analysis for future studies.

The physiological activity of each participant was recorded for the complete duration of the protocols. The obtained signals were then segmented in several trials (or epochs) each one being associated to a given stimulus. The starting time of each stimulus could be retrieved thanks to the triggers while the duration of a trial was defined by the protocols. Before extracting physiological features that are related to emotional activity, three preprocessing steps were applied to the signals: EEG and peripheral signals were denoised, EEG signals were re-referenced and the heart rate (HR) signal was computed from the non-filtered BVP signal.

3.2.2 Denoising

As a first step, the EEG signals were filtered by a 2-47 Hz Equiripple band pass filter (Table 3.1).

This filter was applied to remove the DC offset of each electrode, drifts due to the difference of electrode impedance over time and power lines 50 Hz noise. This band pass filter also allows preserving frequency bands of interest for the study of emotional processes.

The peripheral signals were filtered by a moving average filter to remove noise. For this purpose we used filters of length 512 samples for GSR and temperature, 128 for BVP, and 256 for respiration (Table 3.1). Those different lengths were chosen to remove high frequencies without corrupting oscillations of interest in the different signals.

All the signals were filtered using the filtfilt function from the signal processing Matlab toolbox (v. 6.2.1) which processes the input signal in both the forward and reverse directions. This function allows performing a zero-phase filtering.

High pass (-3dB)

Low pass (-3dB)

EEG 2 Hz 47 Hz

GSR - 0.9 Hz

BVP - 3.5 Hz

Respiration - 1.7 Hz

Temperature - 0.9 Hz

Table 3.1. Low and high pass cutoff frequencies at -3dB for the different filters.

3.2.3 EEG re-referencing

Since with the Biosemi system the original reference (signals are originally referenced to the CMS electrode) provides a poor signal to noise ratio, it is necessary to re-reference them afterward. To obtain a Laplacian reference the following Laplacian operator was applied to each electrode i:

( )

( ) ( ) 1 ( )

i i j

j Neig i i

x n x n x n

N (3.1)

where xi is the CMS referenced signal of electrode i, xi the Laplacian referenced signal, n the sample number, Neig(i) the neighbors electrodes of electrode i and Ni the size of this neighborhood. The neighborhood of an electrode was defined according to the Appendix B.

3.2.4 HR computation

As stated in Section 2.2.2.c an HR signal can be inferred from the BVP signal recorded by a pletysmograph. However, computing HR from a BVP signal is less reliable than from an ECG since the vaso-construction can influence the shape and timing of the heart pulses. As can be seen from Figure 3.4, the BVP signal is periodic because the blood pressure changes with heart contractions. Two points of interest can be identified in this signal: the foot of the systolic upstroke and the systolic peak both due to one of the heart contraction.

A method to determine HR from a BVP signal is proposed in [128]. This method is based on a complex analysis that identifies the systolic peaks as heart beats and requires recordings of long duration. For this study, this method was not used because:

- as can be seen from Figure 3.4 (right) it is sometimes difficult to identify which peak is which in a pulse, especially when the blood pressure is strongly increasing and decreasing or if the signal is of bad quality (noise due to movements, sensor badly placed or moved, etc.);

- in our experiments the trials generally lasted less than 10 seconds, the method was thus not adequate for these signals.

Figure 3.4. The heart waves in a BVP signal. (Left) Three pulses of the BVP signal with the different peaks, (rigth) example of a pulse where it is difficult to identify the different peaks.

In order to compute HR from signals of short durations without using the systolic peak, a method based on the detection of the foot of the systolic upstroke was implemented. The use of this point for identification of a heart beat is motivated by its frequent use for pulse wave velocity computation [96] (i.e. the time elapsed between the heart beat and the corresponding wave in a blood pressure signal). The developed method is composed of the following steps:

1. the linear trend of the BVP signal was removed from each trial to attenuate the effects of strong increase and decrease of blood pressure;

2. heart beats were assumed to be the local minima of the signal which were obtained by finding samples were the derivative is zero and the amplitude is switching from a decrease to an increase;

3. in the case where two such beats fall in the same interval of 0.5 second then only the beat that corresponds to the highest increasing BVP derivative is kept. The 0.5 second interval was chosen based on the assumption that the HR will not exceed 120 beats per minutes (BPM) which is somehow reasonable since in all the protocols participants were sitting in front of a computer screen without performing any significant physical activity;

4. the interbeat intervals (IBI) were computed as the time elapsed between two consecutive beats which could be converted to B-1 HR values corresponding to the B detected heart beats. The time stamp of each HR value was placed in the middle of the corresponding IBI interval.

Systolic

peak Tidal peak

Foot of the systolic upstroke

Time (samples) Time (samples)

Figure 3.5. Example of the beat detection and HR computation algorithm on a 9 seconds signal. The HR signal is represented as a staircase function with the length of a step corresponding to the duration of an IBI.

As can be seen from Figure 3.5 this method performed fairly well on short duration trials (less than 10 seconds) but it was found to be less reliable on signals from one of the protocols where the length of a trial was longer (5 minutes) and signals were noisier. To improve the reliability of the peak detection, an algorithm was designed to detect and correct the falsely detected heart beats a posteriori. It is composed of two main steps described below.

1. Detection and correction of false positive peaks (a beat is detected but this is not a true one):

a. for the ith IBI the median mi of the 5 precedent IBI’s was computed (except for the 5 first IBI);

b. if mi IBI i( ) then the ith IBI is considered as corrupted, being a parameter of the algorithm;

c. for each corrupted IBI check if removing one of the two corresponding peak will solve the problem, in this case this peak is removed.

2. Detection and correction of false negative peaks (a beat has not been identified):

a. similarly to step 1.b. if mi IBI i( ) then the ith IBI is considered as corrupted;

b. for each corrupted IBI i the number Pi of peaks to add was determined by ( ( ) / ) 1

i i

P round IBI i m ; if adding those peaks does not violate the constraints

1.a and 2.a then the peaks were added in this IBI to construct Pi+1 new IBI’s with equal duration.

The parameter was empirically set to 0.2 seconds since it corresponds to reasonable changes in HR and it detected nearly all the false positive and false negative peaks.