An Approach for Real-Time Stress-Trend Detection Using Physiological Signals in Wearable Computing Systems for Automotive Drivers Rajiv Ranjan Singh, Sailesh Conjeti, Rahul Banerjee
Abstract— Fast and credible identification and estimation of driver’s stress-level and stress-type from sensed physiological signals has been one of the critical research areas in the recent past. Several good metrics and mechanisms involving bioelectric signals like the Galvanic Skin Response (GSR), Electrocardiogram (ECG) and the Photoplethysmography (PPG) have been identified by the scholars over the years. This paper discusses the features extracted from physiological data collected in five different scenarios and their usefulness with the help of statistical trend analysis methods. The algorithm developed comprises of a novel shape-based feature weight allocation approach and a technique for credible online realtime stress-trend detection. Such a stress-trend detection by the mesh of embedded sensory elements residing in the e-fabric of a wearable computing system will help in reducing chances of fatal driving errors by the way of in-time activation of alerts and actuation of corresponding safety / recovery procedures.
I.
INTRODUCTION
A
CCORDING to the World Health Organization (WHO) 2009 report, an alarming number of 105725 fatalities and over 2 million disabilities resulted due to road accidents in India. The overall global losses were estimated at US$ 518 billion which cost the governments over 3% of their gross national product [1]. These figures will significantly reduce if a driver assisting machine capable of recognizing his affective state and respond interactively is built [2]. Keeping this in view, Birla Institute of Technology and Science (BITS), Pilani started the BITS-LifeGuard research initiative which aims at building a robust safety-critical wearable computer for automotive drivers. The system should be capable of intelligent and adaptive local processing of signals. The challenges in building such a resource constrained system involve efficient use of computing capacity, battery life, communication resources and on-board memory [3]. In addition to these, the device
Manuscript received May 11, 2011. Rajiv Ranjan Singh is with the Department of Electrical and Electronics Engineering, Birla Institute of Technology and Science (BITS) Pilani, Rajasthan, PIN 333031, India ( phone: +91-9414648289; fax: +91-1596244183; e-mail:
[email protected]). Sailesh Conjeti is with the Department of Electrical and Electronics Engineering, Birla Institute of Technology and Science (BITS) Pilani, Rajasthan, PIN 333031, India (e-mail:
[email protected]). Rahul Banerjee is with the Department of Computer Science, Birla Institute of Technology and Science (BITS) Pilani, Rajasthan, PIN 333031, India (e-mail:
[email protected]).
978-1-4577-2197-7/11/$26.00 ©2011 IEEE
should be driver friendly and proactively alert as and when required. II.
RELATED WORK AND METHODOLOGY
Healey et al. presented methods to evaluate driver’s relative stress level using physiological signals. The study indicated that Galvanic Skin Response (GSR) and heart rate (HR) signals were significantly correlated to driver stress [4]. Melek et al. investigated into trend detection algorithms developed using fuzzy logic, statistical, regression and wavelet techniques. Out of these it was observed that Triggs’ statistical approach accurately identified signal trends in the monitored segments of HR and blood pressure (BP) signals [5]. Applications of this approach in adaptive trend change detection and pattern recognition for physiological monitoring was explored in a study by Ping Yang [6]. The study concluded that Triggs’ statistical approach achieved comparable results with data that had many sudden changes and both high and low degrees of noise. But, for data with gradual changes the approach gave increased false positive results. Ying Zhang emphasized on the need for improved specificity and sensitivity in physiological monitoring systems used in intensive care units. To reduce false alarms and to improve data integration and knowledge representation, a well annotated data set is crucial [7]. In continuation with the sensory data collection already completed [8], this work presents the algorithms for extraction and analysis of features from the GSR and Photoplethysmography (PPG) physiological signals. Furthermore, a novel approach for feature weight allocation and selection based on trend shapes is discussed. Detecting stress-trends during the drive would enable the wearable computer to activate and respond. Triggs’ statistical approach was adopted for trend analysis and feature tracking. Optimal threshold for subsequent signal segment classifications was found using the desirability function approach which is used in multi-response optimization [9]. Successful stress-trend detection would aid in the subsequent stage of stress pattern classifier design. We further envisage that the driver stress pattern recognition algorithm subsequently obtained will be ported on a wearable computer for real-time field testing and calibration. In a real-time driving scenario, it is important to detect the incremental changes in the emotions as well as the fatigue or stress level of drivers. Features derived from physiological
1477
signals if tracked adaptively will help in identification of alarming situations. To satisfy the above criteria, we are using GSR and PPG signals collected from 20 drivers under five carefully chosen scenarios. The data collection methodology, choice of test subjects and preliminary data analysis of the collected signals has been explained in [8]. The advantages of real-time data collection in different driving scenarios over simulated environments may include (a) training classifier on real-time data makes it robust to noise, motion artifacts, device errors etc. (b) effect of factors like environmental, vehicle’s characteristics and driver’s physiological conditions can be considered (c) correlation of stressful events are more accurate than in simulated conditions. Out of the 20 drivers, to bring variability in the test subjects, around 9 drivers were carefully chosen as subjects for the present analysis. The total driving time for each driver lasted approximately 20 minutes. For compatibility PPG and GSR were downsampled to 32 Hz from 128 Hz and 32 Hz respectively. An optimal time window of 10 seconds resulting in 320 samples of GSR and PPG each was chosen because of the limited processor memory available for such applications. We have also annotated carefully, the time of the possible stress-trend markers like sharp left turns, sharp right turns, busy market area, bad road stretches, circular turns, speed breakers, unanticipated pedestrian crossing, abrupt lane change by another vehicle etc. These will be used to validate our stress-trend detection algorithms. Preprocessing of physiological signals collected in realtime scenarios is necessary because of occurrences of high frequency noise, motion artifacts and sensor errors. For the present work, relevant segments in signals collected from the 9 test drivers which were free of motion artifacts and sensor errors have been used. III.
FEATURE EXTRACTION
We have selected three most commonly used feature extraction methods (a) statistical (b) syntactic and (c) transform based [10]. A.
features calculated from the extracted phasic GSR. The algorithm for extracting these features involved detection of peaks, the point of onset of responses and the half recovery points using techniques adopted from [4] and [14]. The mathematical formulae of the syntactic features (shown in Fig. 1) are described in Table I. The importance of such features was investigated in [4] and [13] - [15].
Fig 1.
Galvanic Skin Response (GSR) Syntactic Features.
2) Photoplethysmogram (PPG) Syntactic Features: Morphology of the pulsatile component of PPG signal has been proven to be helpful in obtaining certain clinical parameters such as pulse, HR and heart rate variability (HRV) etc. [16]. In addition to syntactic features of PPG (shown in Fig. 2), spectral and statistical features derived from heart rate (HR) also contribute to the stress detection. For extracting syntactic features, it was required to detect peaks (maxima / systole) and troughs (minima / diastole) in the preprocessed PPG signal [16]. These features along with their mathematical formulae and description have been tabulated in Table I. The physiological significance of such features was investigated in [16] - [20].
Statistical Features
Physiological signals can be classified as random signals as there is always some degree of uncertainty involved in their occurrences [11]. Hence their statistical characteristics, shown in Table I, are important for the present analysis. B. Syntactic (Structural) Features Syntactic features are derived from the geometry of the signals. Useful structural information when extracted would help in classification and description. 1) Galvanic Skin Response (GSR) Syntactic Features: GSR is a bio-electric physiological signal controlled by the sympathetic nervous system [12]. The GSR signal comprises of two components: phasic and tonic. The splitting of GSR into its components was done using special filters as suggested by [13]. In [4] it was observed that as ions filled the sweat glands, a sudden rise in the skin conductance was recorded. This rise was characterized using skin conductivity
1478
Fig 2.
Photplethesmographic (PPG) Syntactic Features
C. Heart Rate Variability (HRV) Features HRV features are found to be a selective and sensitive measure of stress caused by both physical and mental workload. A decreasing HRV indicates abnormality in the autonomic nervous system function and is a sign of inability of a subject to react to stress-trends [19]. 1) HRV Spectral Features using Lomb Periodogram: The instantaneous heart rate time series calculated from the PPG was used to calculate the heart rate power spectrum. The lomb periodogram used by [4] was chosen because of its robustness to missed heart beats [21]. The features calculated are tabulated in Table I. These features carry vital information about the balance between the parasympathetic and sympathetic nervous systems of the subject as proved by [19]-[22]. 2) HRV Statistical Features: The time domain statistical features of HRV are shown in Table I. These features are useful in analyzing the interbeat changes in the HR and emotions like frustration, boredom etc [23]. Feature vector as defined by [10] is a reduced dimensional representation of a pattern comprising of the set of features used to describe it. After feature extraction routines, each signal segment produced an array of 39 features which were concatenated to form the feature vector for further analysis.
B. Feature Weight Allocation The percentage of each trend observed for a feature in each driver was calculated. A particular trend is allotted to a feature if it is observed in at least 50% of the drivers. The trend shapes and the corresponding features weights for some selected features are illustrated in Fig. 3.
IV. SHAPE BASED FEATURE WEIGHT ALLOCATION In this approach, a 160 sec window was selected in each of the five scenarios from the feature vector matrix. The segments are chosen such that the stress-trends manually recorded for each driving scenario are of highest concentration, thus giving an estimate of feature value under constant stress. We indexed each scenario in the order of their occurrence as 1- Pre-driving, 2- Relaxed-driving, 3Busy-driving, 4- Return-driving and 5- Post-driving [8]. It is observed that stress progressively increased from the start of the driving towards Busy-driving and decreased during Return-driving and Post-driving. A. Classification of Trend Shapes When individual feature shapes were manually observed and tabulated for each driver, the following significant trend shapes were noticed: 1) Concave / Convex Trend: This trend indicates that the feature is of high significance level as it attains a peak / trough in a high stress scenario. It also shows that such a feature reflects the transient and scenario-dependent component of the signals collected. Such features were given a weight of 5. 2) Monotonically Increasing / Decreasing Trend: This trend is related to the long term effects of stress and fatigue. It relates to the global component of the signals collected. Owing to steady increase / decrease observed over time, these are of lesser significance than concave / convex trends. Such features were given a weight of 3. 3) Linear / Other Trends: These features carry very less information about the stress state of the driver or the scenario of operation. These were given zero weight.
Fig 3.
Features Shapes Observed and their Weight Allocation.
The trends of features for each driver was observed and tabulated. The weight-sum of a feature is calculated as the cumulative sum of the individual weights of the trend-shape observed (concave / convex: 5, monotonically increasing / decreasing: 3, others: 0) for each driver. The feature weights allotted to the feature was the quotient of weight-sum divided by the total number of drivers. These calculated weights for the extracted features have been tabulated in Table I. V.
TRIGG’S STATISTICAL STRESS-TREND DETECTION METHOD
When a stress-trend occurs, change in physiological signal base-level, morphology and other parameters are observed. The significance of this change is a quantitative measure of the amount of the stress experienced by the subject. In applications where real-time tracking of signals and trend detection is important, Triggs’ Statistical Tracking approach has been very popular [5]. The Triggs’ Tracking Variable (TTV) can be used to determine and track alarmable trends by instant-by-instant real-time analysis of the features extracted. The TTV variable is calculated using the difference between the actual value of a feature and the value predicted using the exponential weighted moving average (EWMA) of the previous values. The value of TTV indicates the significance of change observed. The algorithm assigns a value of +1 to TTV if there is 100% certainty that a feature is increasing and a value of -1 if there is 100% certainty that a feature is decreasing.
1479
TABLE I FEATURES: FORMULAE AND WEIGHT ALLOCATION No.
Feature Name
Abbreviation
Description/Formula
SO
P(S)
Wt.
GSR Statistical Features 1
GSR Mean [11]
GM
where xi is signal value and N is number of samples
CC
66
4
2
GSR Energy [11]
GE
where fs is signal sampling frequency
CC
66
4
3
GSR Time Duration [11]
GTD
LIN
77
0
4
GSR Bandwidth [11]
GB
NC
--
1
GTBP
CC
55
3
CC
55
3
CC
100
5
5 6 7 8 9 10 11
GSR Time Bandwidth Product [11] GSR Dimensionality [11] GSR Peak Rise Time Sum [13] GSR Peak Amplitude Sum [13] GSR Peak Energy Sum[4] GSR Half Recovery Sum [13] GSR First Derivative Average [4]
GD GPRTS
GSR Syntactic Features Peak Rise time = Time of Occurrence of Peak - Time of Point of Onset
GPAS
Peak-Amplitude = GSR value at Peak- GSR value at Point of Onset
CC
100
5
GPES
Peak Energy = 0.5 * Peak Amplitude * Peak Rise Time
CC
100
5
CC
88
4
NC
--
1
CC
55
4
MINC
55
3
CC
50
2
CC
100
5
CV CV MINC CC CV CC
60 77 55 66 66 66
2 4 3 3 5 5
CC CC NC CV NC CV
66 88 -65 -55
4 4 3 3 2 3
CV CC
66 88
4 4
CC
88
4
GHRS GFDA
Half-Recovery Time = Time of Occurrence of Half AmplitudeTime of occurrence of Peak Average First Derivative= Average of the First Derivative observed in the given segment Average Rise Rate = Sum Average of 1st derivative of points with 1st derivative > Positive Threshold (0.025) Average Decay Rate = Sum Average of 1st derivative of points with 1st derivative < Negative Threshold (-0.025) GSR Percentage Decay = Percentage of Time samples in given segment with 1st derivative < Zero (0). Number of peaks in a given segment. PPG Syntactic Features Average of (Time of Peak- Time of Preceding Trough) in a segment Maximum and Minimum of (Value of PPG peak- Value of PPG trough) in a segment Average of (Time of Trough- Time of Preceding Peak) in a segment Average of Period of PPG signal in a segment 60 / (Time Difference between two consecutive peaks) HRV Spectral Features derived from PPG 60* Frequency maximum in range of 0.5-2.5 Hz in HRV spectrum 60* Frequency maximum in range of 0.1-0.25 Hz in HRV spectrum Power in range of 0.003-0.04 Hz in HRV spectrum Power in range of 0.04-0.15 Hz in HRV spectrum Power in range of 0.15-0.4 Hz in HRV spectrum LF Power/ HF Power HRV Statistical Features derived from PPG Mean of all NN intervals Standard deviation of all NN intervals RMS of the sequential differences of the IBI calculated for the whole trial
12
GSR Rise Rate Avg. [15]
GRRA
13
GSR Decay Rate Avg. [15]
GDRA
14
GSR % Decay [15]
GPD
15
GSR No. of Peaks
GNP
16 17 18 19 20 21
PPG Rise Time [16] Pulse Height Min. [18] Pulse Height Max.[18] PPG Fall Time Cardiac Period [16] Inst. HR [19]-[20]
PPGRT PPGPHmin PPGPHmax PPGFT PPGCP PPGIHR
22 23 24 25 26 27
PPG Spectral HR Respiration Rate [22] V. Low Freq. Power [19] Low Freq. Power [19] High Freq. Power [19] LF/HF Ratio [19]
PPGSHR RSP VLFP LFP HFP LFHF
28 29
AVNN [23] SDNN [23]
AVNN SDNN
30
rMSSD [23]
rMSSD
31
pNN20 [23]
pNN20
% of the number of sequential IBI differences that are over 20 ms
CC
100
5
32
pNN50 [23]
pNN50
% of the number of sequential IBI differences that are over 50 ms PPG Statistical Features
CC
100
5
MINC MINC NC NC MINC
55 66 --55
3 3 2 2 3
MINC
55
3
MINC
55
3
33 34 35 36 37
PPG Mean [11] PPGM PPG Energy [11] PPGE PPG First Moment [11] PPGFM PPG Time Duration [11] PPGTD PPG Bandwidth [11] PPGBW Same as used in case of GSR PPG Time Bandwidth 38 PPGTBP Product [11] 39 PPG Dimensionality [11] PPGD Legend: Feature Names: GSR- Galvanic Skin Response; PPG- Photoplethysmography; IBI- Inter Beat Interval Shapes Observed (SO): CC- Concave; LIN-Linear; MINC- Monotonically Increasing; CV- Convex; NC- No Conclusion P(S) = Probability of a Shape = No of Drivers exhibiting a particular shape / Total number of drivers; Wt. - Weights
1480
A. TTV Calculation The parameters of TTV analysis are the smoothing constant α and the confidence interval between TTV lower control limit (TTVl) and TTV upper control limit (TTVu). The smoothing constant α, which can take values between 0.0 and 1.0, determines the time constant for exponential weighing. It also determines the number of control observations to be included in the exponential weighing [24]. The smoothing constant value of 0.3 was chosen as the number of observations included for smoothing are 6. Hence the first six segments were discarded as transients and not considered in the analysis. For α = 0.3, the 90% confidence interval gave the TTVl as -0.63 and TTVu as +0.63. This interval included all the observed significant trends. The algorithm for calculating Triggs’ tracking variable is illustrated in [24]. B. Segment Weight Calculation for TTV Analysis The Segment Weight is defined as the cumulative sum of weights of features whose TTV values were found to violate the control limits. The feature weights used were found using the shape based feature weight allocation algorithm as described in Section IV B. The TTV values for each feature for the given segment was obtained using the TTV calculation method as described in Section V A. If for a particular feature the corresponding TTV value violates the control limits (90% confidence interval), the corresponding feature weight was added to the segment weight (initialized to zero). C. Optimal Threshold Selection using the Desirability Function Approach For classification of a particular segment as a ‘Stressful’ segment, choice of a critical threshold for segment weights is necessary. The objective of the threshold selection algorithm must be to maximize the number of true alarms and minimize the number of false alarms. In order to find the threshold we adopted the desirability function approach which is used for optimization of multiple response processes [9]. Desirability of a response takes values between 0.0 and 1.0, where 0.0 corresponds to a completely undesirable value of response whereas 1.0 corresponds to ideal response value. In such one-sided transformations, to maximize the percentage of true alarms and minimize the percentage of false alarms, an upper and lower limit of the responses was appropriately chosen. The individual desirability functions and the overall desirability for each control value were calculated using formulae used in [9]. The control value corresponding to the maximum desirability was chosen as the threshold. Here the threshold of segment weight was noticed to lie in the range of 18 to 27. A particular alarm for an ith segment is considered as a true alarm if there was a manually recorded stress-trend marker in its vicinity i.e. in [i-2, i+2] segments, otherwise that alarm is deemed as a false alarm. The grand mean of percentage of true and false alarms calculated for each individual threshold were used as responses in this approach. Here the value of the desirability exponent was
chosen as 1 for a linear increase of the desirability function. The overall desirability was calculated from individual desirability for each threshold value. The threshold value with the maximum desirability was chosen as the optimum threshold. VI. RESULTS & CONCLUSION The average number of stress-trend markers recorded during drive is 42. Analyzing from the individual number of True and False alarms detected, it was found that 34 (corresponds to 80% successful detection rate) and 15 were the threshold levels for True and False alarms respectively. Fig. 4 below shows the results when individual desirabilities were calculated for each threshold value. The approach resulted in an optimal threshold value of 24 with overall desirability reaching a maximum of 0.549. The efficiency of detection for a threshold of 24 can be validated by crosschecking with the events annotated during the drive.
Fig 4.
Optimum Threshold Identification using Desirability Function.
The algorithm resulted in successful detection of approximately 71% of stress-trends recorded (reaching a maximum of 80% for 8 drivers only because some recorded events went missing for a driver). An average of approximately 58% of total alarmable trends observed corresponded to true alarms and 42% to false alarms (shown in Table II). The success rate can be more effectively evaluated if the stressors were recorded in more number of scenarios and for larger data set. They should be devoid of any error of judgment, synchronization errors, machine errors, unavailability of all relevant stress-trends.
1481
TABLE II RESULTS OF STRESS TREND DETECTION % of True % of False Driver Index % Detected Alarms Alarms D1 82 (Max.) 59 40 D2 64 45 55 D3 54 42 57 D4 75 65 34 D5 71 60 39 D6 76 66 33 D7 68 50 50 D8 77 70(Max.) 29(Min.) Avg.(Approx.) 71 58 42
The stress-trend markers recorded and the stress-trends detected using the above approach for three sample drivers are marked in Fig. 5.
[6]
[7]
[8]
[9]
[10]
[11] [12]
[13]
Fig 5.
[14]
Stress-Trends Detected
A proper refinement of these methods along with appropriate clustering techniques will be helpful in assessing the degree of stress in the identified stressful signal segments. This is thus expected to help evolve a stressidentification and calibration mechanism that would allow designing custom wearable computing fabric for vehicular drivers and help save loss of precious lives by the way of providing fast yet credible real-time alerts to the drivers and their coupled cars [8].
[15]
[16]
[17]
ACKNOWLEDGMENT The author wishes to thank Dr. Stephen P. Linder of Dartmouth College, Hanover, USA for his technical support in morphological analysis of PPG signals. We also want to thank Dr. B. K. Rout at BITS Pilani for his valuable suggestions.
[18]
[19]
[20]
REFERENCES [1] [2]
[3]
[4]
[5]
World Health Organization, “Global Status Report on Road Safety: Time for Action,” Geneva, WHO, 2009. R. W. Picard, E.Vyzas and J. Healey, “Towards Machine Emotional Intelligence: Analysis of Affective Physiological States,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 23, no. 10, pp. 1175– 1191, October 2001. R. Jafari, H. Noshadi, S. Ghiasi and M. Sarrafzadeh, “Adaptive Electrocardiogram Feature Extraction on Distributed Embedded Systems,” IEEE Trans. Parallel Distributed Systems, vol. 17 no. 8, pp. 797–807, August 2006. J. A. Healey and R.W. Picard, “Detecting Stress During Real-World Driving Tasks Using Physiological Sensors,” IEEE Trans. Intelligent Transportation Systems, vol. 6, no. 2, pp. 156-166, June. 2005. W. W. Melek, Z. Lu, A. Kapps, and W. D. Fraser, “Comparision of Trend Detection Algorithms in the Analysis of Physiological TimeSeries Data,” IEEE Trans. Biomedical Engineering, vol. 52, no. 4, pp. 639–651, April 2005.
[21]
[22]
[23]
[24]
1482
P. Yang, “Adaptive Trend Change Detection and Pattern Recognition in Physiological Monitoring,” Ph.D. Thesis, Univ. Of British Columbia, June 2009. Y. Zhang, “Real-Time Analysis of Physiological Data and Development of Alarm Algorithms for Patient Monitoring in Intensive Care Unit,” Ph.D. Thesis, MIT, USA, August 2003. R. R. Singh and R. Banerjee, “Multi-parametric Analysis of Sensory Data collected from Automotive Drivers for Building a Safety-Critical Wearable Computing System,” Proc. 2nd Int. Conf. on Computer Engineering and Technology (ICCET 2010), pp. V1-355 - V1-360, Chengdu, China, April 2010.. G. Derringer and R. Suich, "Simultaneous Optimization of Several Response Variables," Journal of Quality Technology, vol. 12 no. 4, pp. 214-219, October 1980. E. J. Ciaccio, S. M. Dunn and M. Akay, “Biosignal Pattern Recognition And Interpretation Systems, Part 2 of 4: Methods for Feature Extraction and Selection,” IEEE Engg. Medicine Bio., pp. 106-113, December 1993. C. S. Lessard, “Signal Processing of Random Physiological Signals,” pp. 11-17, Morgan & Claypool, 2006. A. Gorini and G. Riva1, “ The potential of Virtual Reality as anxiety management tool: a randomized controlled study in a sample of patients affected by Generalized Anxiety Disorder,” Trials, 9:25, 2008. S. Schmidt and H. Walach, “Electrodermal Activity (EDA) – State-ofthe-Art Measurement and Techniques for Parapsychological Purposes,” The Jour. of Parapsychology, vol. 64, pp. 139-163, June 2000. J. Zhai, A. B. Barreto, C. Chin and C. Li, “Realization of Stress Detection using Psychophysiological Signals for Improvement of Human-Computer Interactions,” Proceedings of the SoutheastCon 2005, pp. 415-420, 2005. M. Soleymani, G. Chanel, J.J.M. Kierkels, T. Pun, “Affective Ranking of Movie Scenes Using Physiological Signals and Content Analysis,” Proceedings of the 2nd ACM workshop on Multimedia semantics 2008, pp. 32-39, 2008. S. P. Linder, M. Suzanne, W. Edward Wei and S. P. McGrath, “Using the Morphology of Photoplethysmogram Peaks to Detect Changes in Posture,” Journal of Clinical Monitoring and Computing, vol. 20, no. 3, pp. 151-158, Springer, 2006. J. Weng, Z. Ye, and J. Weng, “An Improved Pre processing Approach for Photoplethysmographic Signal,” IEEE Engineering in Medicine and Biology 27th Annual Conference Shanghai, China, pp. 41-44, Sept. 2005. M. Shamir, L. A. Eidelman, Y. Floman, L. Kaplan and R. Pizov, “Pulse oximetry plethysmographic waveform during changes in blood volume,” British Jour. of Anaesthesia, vol. 82 (2), pp.178-181, 1999. N. Hjortskov, D. Risse´n, A. K. Blangsted, N. Fallentin ,U. Lundberg and K. Søgaard, “The effect of mental stress on heart rate variability and blood pressure during computer work,” Eur. Jour. Appl. Physiology vol. 92, pp. 84–89, 2004. D. W. Ryoo, Y. S. Kim and J. W. Lee, “Wearable Systems for Service based on Physiological Signals,” IEEE Engineering in Medicine and Biology 27th Annual Conference, Shanghai, China, pp. 2437-2440, Sept. 2005. P. Laguna, G. B. Moody, and R. G. Mark, “ Power spectral density of unevenly sampled data by least-square analysis: performance and application to heart signals,” ,” IEEE Trans. Biomedical Engineering, vol. 45, no. 6, pp. 698–715, June 1998. D. L. Partin, M. F. Sultan, C. M. Thrush, R. Prieto and S.J. Wagner, “Monitoring Driver Physiological Parameters for Improved Safety,” 2006 SAE World Congress Detroit, Michigan April 3-6, 2006. D. Giakoumis, A. Vogiannou1, I. Kosunen, K. Moustakas, D. Tzovaras and G.Hassapis, “Identifying Psychophysiological Correlates of Boredom and Negative Mood Induced During HCI,” BIOSTEC 2010, Valencia, Spain, pp. 3-12, 2010, Jan. 2010. G. S. Cembrowski, J. O. Westgard, A. A. Eggert, and E. Clifford Toren, Jr., “Trend Detection in Control Data: Optimization and Interpretation of Trigg’s Technique for Trend Analysis,” Clinical Chemistry, vol. 21, no. 10, pp. 1396–1405, May 1975.