lynx   »   [go: up one dir, main page]
Academia.eduAcademia.edu

Figure from:Automatic stress classification with pupil diameter analysis

See full PDFdownloadDownload figure
Parameter Values of the Neural Network Training Algorithm Note: epochs = maximum number of iterations; time = time limit before algorithm stops; goal = target gradient value; grad min = Minimum gradient magnitude; 4 = convergence factor (see Barman & Chowdhury, 2012). TABLE 3

Table 3 Parameter Values of the Neural Network Training Algorithm Note: epochs = maximum number of iterations; time = time limit before algorithm stops; goal = target gradient value; grad min = Minimum gradient magnitude; 4 = convergence factor (see Barman & Chowdhury, 2012). TABLE 3

Related Figures (12)

FIG. 1. (A) Lane Change Test scenario. (B) Experimental design. Our rationale for stress manipulation implies the repeated performance of a simple driving task, to which we added different external stressful stimuli (see section 2.3). The protocol was approved by the French National Board of Informatics and Freedom (declaration n. 0727429; www.cnil.fr). Participants performed a simulated Lane Change Test (LCT), which consists of driving on a traffic-free straight three-lane road (see Fig. 1-A), changing lanes according to the information appearing on two identical road signs displayed concurrently every 150m, on both sides of the road (ISO, 2010; Minin et al., 2011; Mitsopoulos- Rubens et al., 2011).
FIG. 1. (A) Lane Change Test scenario. (B) Experimental design. Our rationale for stress manipulation implies the repeated performance of a simple driving task, to which we added different external stressful stimuli (see section 2.3). The protocol was approved by the French National Board of Informatics and Freedom (declaration n. 0727429; www.cnil.fr). Participants performed a simulated Lane Change Test (LCT), which consists of driving on a traffic-free straight three-lane road (see Fig. 1-A), changing lanes according to the information appearing on two identical road signs displayed concurrently every 150m, on both sides of the road (ISO, 2010; Minin et al., 2011; Mitsopoulos- Rubens et al., 2011).
FIG. 2. Pupil output in mm (top line, left-hand scale) and px (bottom line, right-hand scale). Note. Dashed vertical lines indicate blink onsets
FIG. 2. Pupil output in mm (top line, left-hand scale) and px (bottom line, right-hand scale). Note. Dashed vertical lines indicate blink onsets
FIG. 3. (A) Pupillary response to stressful sounds (solid line, left-hand scale) under constant illumination (dashed line, right-hand scale). (B) Pupillary response factor loadings (only loadings >0.7 are considered for interpretation).
FIG. 3. (A) Pupillary response to stressful sounds (solid line, left-hand scale) under constant illumination (dashed line, right-hand scale). (B) Pupillary response factor loadings (only loadings >0.7 are considered for interpretation).
Several mathematical approaches could be used for this purpose: we chose the Discrete Wavelet Transform (DWT) because a) it allows removing noise yet preserving the original shape of the signal and b) it encompasses a down- sampling procedure which reduces computation time. FIG. 4. (A) Preprocessed pupil diameter from trial tz of participant #7. (B) Signal approximation after 5 wavelet (Haar) decomposition levels. Any wavelet decomposition consists of two parts: approximation and detail. Approximation refers to the overall, general form of the signal (e.g. low frequency component), while detail better explains the high frequency information such as edges, discontinuities, sharp points, etc. purpose: we chose the Discrete Wavelet Transform (DWT) Combination of Wavelet and Neural Networks has been accepted as an accurate method for feature extraction and classification (Minu et al., 2010). Any noisy signal imposes some uncertainty to the calculation and - consequently - to the results. Therefore, before using PD signal as input for the neural network classifier, we need to remove noise from the signal. De-noising could improve classification performance, since it would increase the signal-to-noise ratio. Moreover, it would reduce computational costs, i.e. shorter time to obtain results: this latter aspect is important in a real-time application perspective, although we focus on offline analysis for the present study.
Several mathematical approaches could be used for this purpose: we chose the Discrete Wavelet Transform (DWT) because a) it allows removing noise yet preserving the original shape of the signal and b) it encompasses a down- sampling procedure which reduces computation time. FIG. 4. (A) Preprocessed pupil diameter from trial tz of participant #7. (B) Signal approximation after 5 wavelet (Haar) decomposition levels. Any wavelet decomposition consists of two parts: approximation and detail. Approximation refers to the overall, general form of the signal (e.g. low frequency component), while detail better explains the high frequency information such as edges, discontinuities, sharp points, etc. purpose: we chose the Discrete Wavelet Transform (DWT) Combination of Wavelet and Neural Networks has been accepted as an accurate method for feature extraction and classification (Minu et al., 2010). Any noisy signal imposes some uncertainty to the calculation and - consequently - to the results. Therefore, before using PD signal as input for the neural network classifier, we need to remove noise from the signal. De-noising could improve classification performance, since it would increase the signal-to-noise ratio. Moreover, it would reduce computational costs, i.e. shorter time to obtain results: this latter aspect is important in a real-time application perspective, although we focus on offline analysis for the present study.
FIG. 5. (A) Skin Conductance Response to stressful sounds. (B) Skin Conductance Response factor loadings (only loadings >0.7 are considered for interpretation). Skin conductance (SC) was recorded using the exosomatic method with Direct Current (Boucsein, 2012). Two Biopac EL507 disposable circular electrodes (Ag/AgCl, 1cm diameter circular contact area, 0.5% Chloride) were attached to the participant’ s forehead. Although palmar and plantar zones would be preferable for EDA recording - since they have higher sweat glands density (Dawson et al., 2000; Sato et al., 1989) - the driving task employed in the present study required both hands and feet to be completely free. The electrodes were fixed to the skin upon participant’ s arrival at the lab, assuring a minimum delay of 15min before the recordings started. This time is sufficient to allow good electrical contact between the skin and the electrode surface. An elastic headband was used to prevent artifacts due to wire movements (Boucsein et al., 2012).
FIG. 5. (A) Skin Conductance Response to stressful sounds. (B) Skin Conductance Response factor loadings (only loadings >0.7 are considered for interpretation). Skin conductance (SC) was recorded using the exosomatic method with Direct Current (Boucsein, 2012). Two Biopac EL507 disposable circular electrodes (Ag/AgCl, 1cm diameter circular contact area, 0.5% Chloride) were attached to the participant’ s forehead. Although palmar and plantar zones would be preferable for EDA recording - since they have higher sweat glands density (Dawson et al., 2000; Sato et al., 1989) - the driving task employed in the present study required both hands and feet to be completely free. The electrodes were fixed to the skin upon participant’ s arrival at the lab, assuring a minimum delay of 15min before the recordings started. This time is sufficient to allow good electrical contact between the skin and the electrode surface. An elastic headband was used to prevent artifacts due to wire movements (Boucsein et al., 2012).
FIG. 6. Average VAS scores (stress dimension) for each group and experimental trial. Note. Vertical bars denote 95% confidence intervals (mean +2SE). N = 32 (17 experimental + 15 controls).
FIG. 6. Average VAS scores (stress dimension) for each group and experimental trial. Note. Vertical bars denote 95% confidence intervals (mean +2SE). N = 32 (17 experimental + 15 controls).
FIG. 7. Normalized average pupil diameter (left-hand scale) and average illumination (right-hand scale) for each group and experimental trial. Note. Vertical bars denote 95% confidence intervals (mean +2SE). N = 29 (16 experimental + 13 controls). NS.EDR freq. scores were analyzed with a rmANOVA using trial (t;, tz, tz, tz) as within-factor and group (experimental control) as between-factor. No significant effects were found. The effect of group was also present (F(1,27) = 4.01, p=.05, n> = .13), in that normalized average pupil diameter was higher for the experimental with respect to the control group. When t, was removed from the within-factors (since stress manipulation effectively started at t,), the effect of group emerged more clearly (F(1,27) = 6.31, p<.05, n> = .19).
FIG. 7. Normalized average pupil diameter (left-hand scale) and average illumination (right-hand scale) for each group and experimental trial. Note. Vertical bars denote 95% confidence intervals (mean +2SE). N = 29 (16 experimental + 13 controls). NS.EDR freq. scores were analyzed with a rmANOVA using trial (t;, tz, tz, tz) as within-factor and group (experimental control) as between-factor. No significant effects were found. The effect of group was also present (F(1,27) = 4.01, p=.05, n> = .13), in that normalized average pupil diameter was higher for the experimental with respect to the control group. When t, was removed from the within-factors (since stress manipulation effectively started at t,), the effect of group emerged more clearly (F(1,27) = 6.31, p<.05, n> = .19).
FIG. 8. (A) Skin conductance (solid line) and Tonic Level (dashed line) estimated by deconvolution (Benedek & Kaernbach, 2010). (B) Skin conductance minus Tonic Level, showing a zero baseline. The time-integral of this time series is used as EDA area measure.
FIG. 8. (A) Skin conductance (solid line) and Tonic Level (dashed line) estimated by deconvolution (Benedek & Kaernbach, 2010). (B) Skin conductance minus Tonic Level, showing a zero baseline. The time-integral of this time series is used as EDA area measure.
Note: Architecture details of the four binary classifiers (see Fig. 9). Each classifier is designed to maximize recognition of one class with respect to all the other classes (e.g. maximize recognition of t; with respect to ty, t3, t,) . Neural Network Classifiers Architectures TABLE 2
Note: Architecture details of the four binary classifiers (see Fig. 9). Each classifier is designed to maximize recognition of one class with respect to all the other classes (e.g. maximize recognition of t; with respect to ty, t3, t,) . Neural Network Classifiers Architectures TABLE 2
FIG. 9. Schematic representation of the one-versus-all binary neural network classifier.
FIG. 9. Schematic representation of the one-versus-all binary neural network classifier.
FIG. 10. Schematic representation of the four-way parallel neural network classifier. Note. Number in parentheses indicate frequencies.
FIG. 10. Schematic representation of the four-way parallel neural network classifier. Note. Number in parentheses indicate frequencies.
Four-Way Classifier Confusion Matrix
Four-Way Classifier Confusion Matrix
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved
Лучший частный хостинг