Classification of Startle Eyeblink Metrics using Neural Networks

Transcription

1 Classification of Startle Eyeblink Metrics using Neural Networks Christopher T. Lovelace, Reza Derakhshani Member, IEEE, Sriram Pavan Kumar Tankasala Member, IEEE, and Diane L. Filion Abstract-In this paper, we show the feasibility of using high-speed video for measurement of startle eyeblinks as a new augmentative modality for biometric security, as blinks can reveal emotional states of interest in security screenings using nonintrusive measurements. Using neural network as classifiers, this initial study shows that upper eyelid tracking at 250 frames per second can categorize startle blinks with accuracies comparable to those of the well-established but intrusive EMGbased measures of muscles in charge of eyelid closure. Index Terms- Neural Networks, Signal Detection, Biomedical Signal Analysis, Image Processing, Pattern Classification, Biometrics, Psychology. I. INTRODUCTION t is possible to assess a number of psychological processes Iby observing an individual s response to a sudden, intense, and unexpected sensory event. The ensuing startle reaction engages a set of fast, reflexive muscle contractions that includes a robust eyeblink. Numerous studies have shown this startle eyeblink response to be reliably modulated, both in size and latency, by psychological factors such as attention, emotion, cognitive load, and sensory processing [1]. While there are many other approaches to psychological measurement (e.g., questionnaires, functional brain imaging, etc.), startle eyeblink modulation (SEM) has a distinct advantage in that it is a reflexive response. No cognition or overt response is required of the participant, making it difficult to purposely manipulate the test results, which makes SEM a viable modality for security applications. This is especially useful when evaluating whether a person is being deceptive or otherwise in a negative emotional state, arguably the two most important features currently absent in biometric systems which Manuscript received January 15, This work was supported in part by This work was supported in part by a research grant from Center for Identification Research Center (CITeR), a National Science Foundation Industry/University Cooperative Research Center in biometrics. Christopher Lovelace is with the University of Missouri Kansas City, Kansas City, MO USA ( ; lovelacec@ umkc.edu). Reza Derakhshani is with the University of Missouri Kansas City, Kansas City, MO USA, ( derakhshanir@umkc.edu). Sriram Pavan Kumar Tankasala is with the University of Missouri Kansas City, Kansas City, MO USA ( ; Diane Filion is with the University of Missouri Kansas City, Kansas City, MO USA ( ; filiond@ umkc.edu). in large part rely only on recognition of unique, identifiable physiological features such as iris patterns and fingerprints [2]. Two additional factors make SEM useful in detecting imposters or high-risk claimants being vetted by a biometric system. First, as we will show, it is possible to measure SEM unobtrusively using high-speed video. Such recordings are also amenable to long-range acquisition using high-powered optics. Second, the startle response may be modulated by psychological factors related to deception and hostile intent; as it is well-established that features such as startle response intensity and latency, reflected in electromyographic (EMG) recordings, are reliable indicators of emotional valence in response to a preprogrammed stimulus [3]. One can envision a video-based SEM system being used accordingly for humanmachine interactions in future psychometrically enhanced biometric systems. II. MEASUREMENT OF STARTLE EYEBLINK For SEM to be a useful psychological measure, one must be able to quantify the robustness of an eyeblink and its timing relative to the startle-eliciting stimulus. While there are a number of ways to measure the eyeblink response, this is most often done by using contact electrodes to record the EMG activity of the orbicularis occuli muscle just below the eye, whose contraction produces eyelid closure [4]. While the EMG response is a sensitive index of the startle response, this technique has the disadvantage of being fairly obtrusive, requiring cleaning of the skin and adhesion of electrodes. For some applications, such as next generation psychometricallyenhanced biometric security systems, it would be advantageous to have available a non-contact technique where the eyeblink response could be measured without the need for electrodes in order to detect potentially dangerous or deceptive individuals based on their SEM reaction. We are presently developing two non-contact techniques for measuring startle eyeblink. The first uses near infrared light reflected from the surface of the eye as an indicator of lid closure. Eyelid movement produces a measurable change in the amount of reflected near infrared light, with larger blinks producing greater changes. While this technique has revealed that measurement of physical eyelid movement provides nearly as sensitive an index of SEM as EMG [5], it has the disadvantage that, given current instrumentation, the sensor

2 must be positioned very close to the eye. A second noncontact approach to eyeblink measurement uses high-speed video recording. While this has the advantage of allowing for unobtrusive recording at a greater distance, so far the video record has been more difficult to quantify and score. An answer to this challenge is the subject of this paper. Recording of the EMG and reflected light measures both involve digitization of a continuously changing voltage signal. Extraction of the relevant features (e.g., response onset and peak) for individual eyeblink responses from these signals usually involves an initial pass with computer algorithms that programmatically identify the onset and peak of a response. However, due to noise and other considerations, these values must then be verified by eye by the researcher. This can be a time-consuming process, given that there are usually a few dozen trials to be scored for each testing session. Feature extraction for the video signals is much more challenging. One must track not only eyelid movements, but also head movements so that these may be subtracted out of the final eyelid movement vectors. We present here both a method for extracting eyelid movement information from an eyeblink video record as well as a way to quantify SEM in both video and EMG recordings without user intervention using neural networks. procedures were approved by the UMKC Social Sciences Institutional Review Board. B. Stimuli Acoustic stimuli were presented binaurally using Sennheiser HD590 headphones to participants comfortably seated in a sound- and light-attenuating room. Acoustic stimuli and trigger pulses were controlled using the Presentation software program (Neurobehavioral Systems, Inc.; Albany, CA) on a Pentium class desktop computer. The startleeliciting stimulus was a 50 ms, 105 db SPL(A) (a perceptually-weighted decibel scale) burst of white noise (<1 ms rise/fall). The prepulse was a 20 ms, 70 db SPL(A), 1 khz tone (5 ms rise/fall). The onset of the prepulse preceded the onset of the startle stimulus by 120 ms (see Figure 1). Participants were asked to keep their head stationary and their eyes pointed straight ahead at all times. Stimuli were presented in two sets of trials, with each set containing 20 prepulse trials and 20 control trials (pseudorandom order). Individual trials were separated by sec (random, uniform distribution). III. PREPULSE INHIBITION For the purpose of demonstrating the utility of neural networks to the scoring of eyeblinks recorded using highspeed video and EMG, we elicited startle eyeblinks using a standard stimulus paradigm known to produce robust SEM. This paradigm is based on the observation that the startle response to a loud, sudden sound may be reliably modulated by an innocuous sound that occurs just before the loud sound. When the initial sound (prepulse) precedes the startling sound by about 100 ms, this has the effect of speeding eyeblink latency and reducing eyeblink amplitude [6]. This effect, termed prepulse inhibition (PPI), has been used extensively to study sensory processing and cognition in both healthy controls and patients with various psychiatric disorders [7]. PPI is computed by comparing metrics for eyeblinks that follow a startle stimulus that was preceded by a prepulse (prepulse trials) with those that follow a startle stimulus presented alone (control trials). While PPI is (likely) unrelated to detection of deception, this is a robust effect and will help us evaluate the utility of our approach, as the main aim of this study is to establish the utility of neural networks in measuring SEM using high-speed video in lieu of traditional EMG-based methods. IV. METHODS A. Participants Nineteen healthy adult members of the University of Missouri Kansas City community were recruited for this study. All had self-reported normal hearing and no diagnosis of any psychiatric or neurological condition, or a first-degree relative diagnosed with any such condition. Each participant completed informed consent prior to participating, and all Fig. 1. Graphic depiction of the acoustic stimuli in a control trial (top) and prepulse trial (bottom). C. Eyeblink Measurement and Analysis Ocular EMG was recorded using two Ag/Ag-Cl electrodes adhered to the skin just below the left lower eyelid, with a ground electrode attached at the left temple. The EMG data were digitized at 1 khz using the BIOPAC MP-150 system (BIOPAC Systems, Inc.; Goleta, CA) with a gain of 1000 and filter passband of Hz. The raw data were stored and analyzed offline using in-house software. Prior to analysis, EMG waveforms were filtered using 4 th -order Butterworth filters with -3 db cutoff frequencies at 30 and 400 Hz, respectively. Waveforms were additionally smoothed using a 5-point boxcar filter applied 2 6 times in succession. For each trial that contained a visible response and was not contaminated by a pre-stimulus blink, we measured response onset and peak latency relative to stimulus onset and peak amplitude relative to the 50 ms pre-stimulus baseline. High-speed video recordings were made in a subset of 8 participants using a Fastcam 512 PCI camera (Photron USA, Inc.; San Diego, CA) recording at 250 frames per second with a pixel resolution of 256 x 240. The camera was focused tight on the right eye (opposite the side on which EMG recordings were made; see Figure 2 for a sample frame). Video recording on each trial was triggered 100 ms prior to startle stimulus onset (providing a pre-stimulus baseline) and continued for 1000 ms (250 frames). Because camera memory was limited, video recordings were made on only the first 35 trials for each run.

3 Fig. 2. Single frame from a high-speed video record. V. RESULTS A. Participants Of the 19 participants recruited for this study, 4 were excluded because they showed too few trials containing scorable blinks in either the video or EMG. The final sample consisted of 15 participants (11 women) with a mean age of 22.7 years (SD = 4.4 years). High-speed video was recorded from 4 of these participants (4 women), with a mean age of 22.5 years (SD = 1.0 years). B. Feature Extraction For EMG data, which were recorded simultaneously with the high-speed video and for the purpose of validating the efficacy of video methods for differentiating prepulse from control trials, we used a three-dimensional feature vector containing the numerical values for response onset latency (stimulus onset to response onset time difference), response peak latency (stimulus onset to response peak time difference), and response intensity (peak EMG voltage relative to baseline). These measurements have been traditionally used to quantify SEM [4]. We will refer to these onset-peak three dimensional feature sets, which were garnered by visual inspection, as scheme A (please see Figure 3). For high-speed video data we used two types of features. For the first type, the onset-peak features were extracted by visual inspection in a fashion similar to the aforementioned scheme A for EMG, i.e. a triplet of response latency, peak latency, and intensity for each trial (Figure 3). For the second set of features (scheme B), we extracted the displacement of upper eyelid automatically as follows, and used a snapshot of the whole 1-second displacement signal (250 integer values) for each SEM as their feature vector. Biometric applications of the psychometric information contained in startle eyeblinks suggest the need for nonintrusive measurements through video. Thus accurate isolation and tracking of the eyelid movements without any attached markers or additional lighting is desired. Adding to this already challenging image processing task, typical real world applications also require (near) real-time processing. For this purpose, we chose Motus markerless motion capture and image pattern tracking software (Vicon Motion Systems, Denver, CO). We targeted two regions of interest for Motus pattern tracking: a mid upper eyelid point and a reference eye corner area (Figure 4). The reason for the selection of the eye corner as a reference point is the fact that subjects head movements need to be cancelled out by subtracting the displacement of the latter point from the eyelid movements. Figure 5 shows an example of this method, which provides a vertical trace for blink motion free of head movement artifacts. The depicted signals are the y coordinates for the tracked points on the upper eyelid, eye corner reference point, and their difference. The latter yields the video-based SEM signal used for classification in Scheme B, with each feature vector being a 1-second (250 point) eyelid movement signal. Stimulus Onset Peak latency Blink Onset Peak Intensity Fig. 4. Markers are placed on the eye region in high-speed videos for automatic tracking of the eyelid movements. Marker s y pixel count Blink latency Fig. 3. Blink features for Scheme A. Vertical axis represents either EMG intensity or upper eyelid displacement in high-speed video, and horizontal axis depicts time. A triplet of blink latency, peak latency, and peak intensity is used to describe each SEM. Frame number (time, 4ms) Fig. 5. Eyelid movement tracking using high-speed video. Depicted signals from top: upper eyelid location, eye-corner location (reference point), and the relative movement of eyelid (y upper_lid -y eye_corner ). Vertical axis shows markers y coordinate (pixel count). Horizontal axis depicts the corresponding video frame number at 250 FPS (4ms/frame).

4 C. Neural Network Analysis In order to classify the aforementioned blink features into control and prepulse, we examined a variety of different feedforward neural networks as classifiers with bi-valued targets and hyperbolic tangent nonlinearities. For training, we used gradient descent scaled conjugate gradient algorithm [8], which provides a good balance between performance and speed. The SEM feature sets (EMG scheme A and video schemes A and B) were randomly divided into three subsets: 60% for training, 20% for early-stopping/validation, and 20% for testing. Each neural network configuration was initialized by randomizing its weights and subsequently trained 20 times to evaluate the expected performance of each configuration, given the stochastic nature of the local minima convergence of gradient descent with random starting points [9]. The earlystopping maximum fail criterion was set to 5 iterations. All the experiments were performed using MATLAB 2008a software (Mathworks, MA) running on a Pentium-class personal computer. We performed a semi-exhaustive search across different hidden layer sizes in an attempt to canvass the architecture space for best network configurations for our feature sets (EMG scheme A and video schemes A and B). Given the scarcity of training data due to the limited enrollment exposures in typical biometric applications, we chose a regularized error function given in (1) to penalize extraneous weights and thus avoid over-parameterization which is characteristic of smaller training datasets: SSE Regularize d = 1 2 N i= 1 P 2 λ e i + w 2 k = 1 Here N is the total number of training samples, e i is the classification error for i th data point, P is the total number of free parameters (i.e. network weights w k ), and λ is the regularization constant which was set to 0.2 in our experiments. This parameter dictates the weight decay pressure, which is the result of weight-shrinking gradient of the second term in (1). Gradient descent learning lowers the above sum of squared errors (SSE) which creates sparse and thus low variance models on limited training datasets resulting in better generalization capabilities (another justification of the above augmented error criterion comes from derivation of (1) via a Bayesian framework and assuming Gaussian distribution for both the neural network weights and target data noise [10]). Here we chose 2 hidden layer networks especially for scheme B high-speed video features, because not only the addition of another hidden layer will introduce higher computational capabilities and mapping flexibility to the network, but also the first hidden layer can act as a feature map (feature extractor) similar to convolutional networks [11]; explaining the better experimental results garnered using an extra hidden layer. The following are the results of the neural network classifiers, also depicted in Figures 6 through 11. EMG: For the latency-peak feature sets (scheme A), we varied the number of hidden nodes in hidden layer 1 and 2 from 1 to 51, and with a step size of 2. We initialized each possible 2 k (1) combination of the aforementioned layer sizes 20 times by randomizing the weights prior to each training run. The highest achieved average performance (correct classification of unseen prepulse vs. control trials) was 63.7%, which occurred at hidden layer 1 size of 37 and hidden layer 2 size of 39 nodes (Figure 6). At this configuration, and out of the 20 training runs, the maximum attained accuracy was 72.4% (Figure 7). Fig. 6. Average test-set performance of neural network blink classifiers vs. number of nodes in their hidden layer 1 (horizontal, 1-51 nodes) and hidden layer 2 (vertical, 1-51 nodes); used for picking the best architecture. Network inputs are EMG latency-peak triplet features (scheme A).Lighter shades show better test performance. High-speed video, scheme A: For these visually scored triplets features, similar to the previous case, we varied both hidden layer sizes from 1 to 51 nodes with a step size of 2. Using the 20 training runs per configuration, best average performance (60.5%, Figure 8) occurred at layer 1 size of 41 and layer 2 size of 53 nodes. At this configuration, and out of the 20 training runs, the maximum attained accuracy was 100% (Figure 9). High-speed video, scheme B: In this mode we essentially used all the 250 samples that made up each 1-second upper eyelid displacement signals per SEM as the input to our 2 hidden layer neural networks, with no further feature extraction. Again, similar to the previous case, we varied both hidden layer sizes to find a suitable configuration. Given the increased input feature dimensionality (250 vs. 3), and to save computation time, we varied hidden layer 1 size from 1 to 101 nodes, but with a step size of 5. For hidden layer 2, we varied the size from 1 to 22, with a step size of 3. Using the same 20 training runs per configuration, best average performance (54.7%, Figure 10) occurred at a layer 1 size of 31 and layer 2 size of 4 nodes. At this configuration, and out of the 20 training runs, the maximum attained accuracy was 80% (Figure 11). It is interesting to note that the best result, regardless of average best configuration prescreening across all the combinations of layers 1 and 2 sizes was 87%, which occurred at layer 1 and layer 2 sized of 81 and 13 nodes, respectively. These results are summarized in Table 1.

5 Fig. 7. Best test-set performance (out of 20 initializations per configuration) of neural network blink classifiers vs. number of nodes in their hidden layer 1 (horizontal, 1-51 nodes) and hidden layer 2 (vertical, 1-51 nodes). Network inputs are EMG latency-peak triplet features (scheme A). Lighter shades show better test performance. VI. DISCUSSION Using EMG and video recordings, we have shown that eyeblinks measured using high-speed video contain sufficient information to quantify SEM (at least with respect to determining whether or not a startle stimulus was preceded by a prepulse) compared to the more frequently used, but less convenient, EMG measurements. We have also demonstrated that neural networks may be effectively used for direct classification of control vs. PPI SEMs without the need for manual scoring and feature extraction. Fig. 8. Average test-set performance (out of 20 initializations per configuration) of neural network blink classifiers vs. number of nodes in their hidden layer 1 (horizontal nodes) and hidden layer 2 (vertical nodes), used for picking the best architecture. Network inputs are high-speed video latency-peak triplet features (scheme A). Lighter shades show better test performance. Fig. 9. Best test-set performance (out of 20 initializations per configuration) of neural network blink classifiers vs. number of nodes in their hidden layer 1 (horizontal nodes) and hidden layer 2 (vertical, 1-51 nodes), used for picking the best architecture. Network inputs are high-speed video latencypeak triplet features (scheme A). Lighter shades show better test performance. As intimated in the Introduction, a video-based SEM psychometric mechanism could be used, either by itself or in conjunction with next-generation psychometrically-enhanced biometric systems, for noninvasive measurement of psychological constructs relevant to deception. Thus not only the apparent characteristics of the claimants, such as their iris patterns, but also their credibility could be assessed before allowing access to the secured entity protected by a biometric system. While the startle PPI paradigm used in the present study has not yet been specifically tested for sensitivity to deception, it has been suggested elsewhere that other forms of SEM may be useful in this arena [12]. The reason for this is that startle can be reliably modulated by ongoing emotional and attentional states as well as to cognitive load. Therefore, the startle response could have utility in the detection of deception in two ways. First, threatening situations and a state of fear have both been shown to increase the size of the startle response [13]. For example, the startle eyeblink response is larger in the presence of a cue that predicts a painful electric shock [14]. So, to whatever degree being presented with crime-relevant details or deception-related questions induces a state of threat or fear, startle eyeblinks elicited while a person is presented with such stimuli would be assumed to be larger than those presented during innocuous cues. Second, there is evidence to suggest that the act of lying is more cognitively demanding than telling the truth [15] [16], and cognitive effort has been shown to influence PPI. For example, PPI is more robust when a person is actively attending to a prepulse stimulus than when they are not [1]. TABLE I CLASSIFICATION RESULT SUMMARY (TEST DATASET, % CORRECT) Input Type Feature set: Scheme A Feature set: Scheme B EMG Average: 63.7% Best: 72.4% NA High-speed video Average: 60.5% Best: 100% Average: 54.7% Best: 80%

6 ACKNOWLEDGEMENTS The authors wish to extend their gratitude to Wade Elmore and Elizabeth Duval (Department of Psychology, University of Missouri-Kansas City) for their help with participant recruitment, data collection and analysis, as well as Dr. Judee Burgoon and her team (University of Arizona) and Dr. Arun Ross (West Virginia University) for their contributions to this research. REFERENCES Fig. 10. Best Average test-set performance of neural network blink classifiers vs. number of nodes in their hidden layer 1 (horizontal, nodes) and hidden layer 2 (vertical, 1-22 nodes), used for picking the best architecture. Network inputs are high-speed video latency-peak triplet features (scheme A). Lighter shades show better test performance. Increased cognitive load during deception, and/or heightened attention to deception-related cues, may cause a measurable difference in the startle response elicited when a false answer is being formulated and when a truthful answer is about to be given. Future work will include analysis of SEM video captures acquired during deceptive/truthful questions and answers. We also would like to analyze different, smaller feature sets as compared to scheme B in order to be able to utilize smaller data driven classifiers, for instance smaller neural networks, which require smaller datasets enabling more robust training. This should also allow us to calibrate a classifier to each subject, compared to the currently presented results using one classifier for all subjects. Also, while these data were collected with a rather high frame rate of 250 FPS, we would like to find out whether a lower frame rate could be adequate. Fig. 11. Best test-set performance (out of 20 initializations per configuration) of neural network blink classifiers vs. number of nodes in their hidden layer 1 (horizontal, nodes) and hidden layer 2 (vertical, 1-22 nodes), used for picking the best architecture. Network inputs are high-speed video latencypeak triplet features (scheme A). Lighter shades show better test performance. [1] D. L. Filion, M. E. Dawson, and A. M. Schell, "The psychological significance of human startle eyeblink modification: A review," Biological Psychology, vol. 47, pp. 1-43, [2] A. K. Jain, R. Bolle, and S. Pankanti, Eds., BIOMETRICS: Personal Identification in Networked Society. Kluwer Academic Publishers, [3] J. Stanley and R. G. Knight, "Emotional specificity of startle potentiation during the early stages of picture viewing," Psychophysiology, vol. 41, p , [4] T. Blumenthal and et al., "Committee Report: Guidelines for human startle eyeblink electromyographic studies," Psychophysiology, vol. 42, pp. 1-15, [5] C. T. Lovelace, W. R. Elmore, and D. L. Filion, "An alternative to EMG for measuring prepulse inhibition of startle eyeblink," Pyschophysiology, vol. 43, pp , [6] F. K. Graham, "Control of reflex blink excitability," in Neural mechanisms of goal-directed behavior and learning. New York, USA, Academic Press, 1980, pp [7] D. L. Braff, M. A. Geyer, and N. R. Swerdlow, "Human studies of prepulse inhibition of startle: Normal subjects, patient groups, and pharmacological studies," Psychopharmacology, vol. 156, pp , [8] M. F. Møller, A Scaled Conjugate Gradient Algorithm for Fast Supervised Learning. Denmark, Aarhus University Publishing, [9] J. C. Principe, N. R. Euliano, and W. C. Lefebvre, Neural and Adaptive Systems:Fundamentals through Simulations, ISBN ed. New York City, USA, John Wiley and Sons, [10] C. M. Bishop, Pattern Recognition and Machine Learning. New York, USA, Springer Science, [11] S. Haykin, Neural Networks - A Comprehensive Foundation, Second Edition ed. Upper Saddle River, New Jersey, USA, Prentice Hall, [12] B. E. Verschuere, "Startling secrets: Startle eye blink modulation by concealed crime information," Biol Psychol, vol. 76, pp , [13] M. M. Bradley, T. Silakowski, and P. J. Lang, "Fear of pain and defensive action," Pain, vol. 137, pp , [14] A. O. Hamm and E. al., "Emotional learning, hedonic change, and the startle probe," Journal of Abnormal Psychology, vol. 102, pp , [15] R. J. Johnson, "The self in conflict: The role of executive processes during truthful and deceptive responses about attitudes," Neuroimage, vol. 39, pp , [16] S. E. Leal, "the time of the crime: Cognitively induced tonic arousal suppression when lying in a free recall context," Acta APsychol (Amst), vol. 129, pp. 1-7, [17] G. J. Siegel, N. Ichikawa, and S. Steinhauer, "Blink before and after you think: Blinks occur prior to and following cognitive load indexed by pupillary response," Psychophysiology, vol. 45, pp , [18] Discovering Vicon Motus 9 Manual. [19] Vicon Online Help [Online]

7 Christopher T. Lovelace, Ph.D., has been a faculty member in the UMKC Department of Psychology since His research focuses on the use of startle eyeblink and psychophysics to measure sensory and cognitive processing both within and across sense modalities in humans. Dr. Lovelace earned his B.A in Psychology from Wake Forest University, and his Ph.D. in Psychology from American University. He has published research in the areas of startle eyeblink, synesthesia, and multisensory integration. Reza Derakhshani, Ph.D.,(M 98) - joined the UMKC School of Computing and Engineering as an assistant professor in His research focuses on computational intelligence with applications in biometrics and biomedical signal analysis. He earned his Ph.D. and Master s degrees in Computer Engineering and Electrical Engineering respectively from West Virginia University. He earned his bachelor's degree in Electrical Engineering from Iran University of Science and Technology. His work has been mainly funded by the National Science Foundation, and has resulted in a number of peerreviewed publications and a U.S. patent. Sriram Pavan Kumar Tankasala M.S.E.E., (M 08) - is a Ph.D. student at the University of Missouri - Kansas City majoring in Electrical Engineering. His research interests include biometrics and biomedical signal processing. He received his Master s degree in Electrical Engineering from University of Missouri Kansas City, and his bachelor's degree in Electrical and Electronics Engineering from Jawaharlal Nehru Technological University Hyderabad, India. Diane Filion, Ph.D., -is associate professor and chair of the Psychology Department at UMKC. Her research focuses on the use of psychophysiological measures, primarily startle eyeblink, to investigate stimulus, task, and participant factors that improve and impair early attentional processing and sensorimotor gating. Dr. Filion earned her Ph.D. and master s degree from the University of Southern California, and her bachelor s degree from Eastern Washington University.