Artificial Neural Network Analysis and ERP in Intimate Partner Violence Sara Invitto1*, A. Mignozzi 1, G.Piraino 3, Gianbattista Rocco2, Irio De Feudis2, Antonio Brunetti2, Vitoantonio Bevilacqua2** 1 Dept. of Biological and Environmental Sciences and Technologies, University of Salento *corresponding author: [email protected] 2Dept. of Electrical and Information Engineering, Technical University of Bari ** second corresponding author: [email protected] 3 Istituto Santa Chiara, Lecce

Abstract. The aim of this work was to analyze, through artificial neural network models, cortical pattern of women with Intimate Partner Violence (IPV) to investigate representative models of sensitization or habituation to the emotional stimulus in IPV. We investigated the ability of high emotional impact images, during a recognition task, analyzing the electroencephalogram data and event related potentials. Neural network analysis highlighted an impairment in IPV group in cortical arousal, during the emotional recognition task. The alteration of this capacity has obvious repercussions on people's lives, because it involves chronic difficulties in interpersonal relationships. Keywords: Intimate Partner Violence, ERP, Artificial Neural Network, Student t-test Features Reduction, Binary Classification.

Introduction Intimate partner violence (IPV) is an important problem of public and social health and involves men and women all over the world regardless of culture, religion and demographic characteristics. This term is often used interchangeably with the term "domestic violence." The biological perspective also has provided information about the factors that might constitute IPV indicators [1] [2]; skin conductance, as well as the levels of testosterone, and the levels of salivary immunoglobulin in the authors of "domestic violence" represent the main marker of a high reactivity to stressful situations, and correlating to impulsive and aggressive behaviors; also in this perspective these indicators may be useful both within prevention programs of the phenomenon, both in terms of

analysis of the risk of recurrence of the phenomenon, as well as in terms of a greater understanding of the underlying causes of the phenomenon. In addition, the biological perspective has highlighted not only the correlation between the location of the violence and the consequent impairment of cognitive functions of the women involved, but also made it possible to highlight how these alterations refer in post-traumatic stress disorder (PTSD) [3]. Numerous data, in fact, have found that 90% of the lesions caused by IPV are around the area of the head, face and neck of the victims; such injuries lead often to brain damage (TBI), both structural and functional, of multiple severity. Small TBI involve mental status abnormalities, neurological and neuropsychological deficits. Women with IPV histories show alterations in the regulation of emotions, in the regulation of pain sensitivity and in management control of fear; such alterations correspond with functional deficits, which correlate with structural alterations and also represent the main characteristics of PTSD. [4]. It’s possible catalog the focal characteristics of PTSD symptoms in three main categories: intrusive symptoms; avoidance symptoms; hyper-excitability symptoms. In intrusive symptoms the patient relives in persistent and persecutory form the traumatic experience, through recurring thoughts and feelings, that recur vividly in the form of flashbacks; while in the symptoms of avoidance, the patient tends to avoid stimuli and situations which can evoke memories of the traumatic experience; finally, the symptoms of hyper-excitability involve manifestation of high levels of arousal and physiological agitation [5]. Other secondary symptoms can be an impoverishment of autobiographical memory for positive events [6], an impaired working memory functioning [7] and also an alteration of attentional capacity; for this reason, studies that have analyzed the structural abnormalities of the brain resulting from PTSD have focused on the hippocampus, a gray matter structure of the limbic system involved in declarative memory, working memory and episodic memory [8]. The importance of PTSD, and its correlation with the phenomenon of domestic violence, it connects to the consequences of such phenomena, corresponding to an alteration of the victims’ routine. These consequences are manifested although a late and acute onset, consequent to stressful events or threatening and catastrophic situations, causing serious emotional suffering protracted in time. Precisely for the importance of the implications, PTSD is recognized as mental disorder by the American Psychiatric Association in 1980 and subsequently introduced in DSMIII. This disorder is then conceptualized as the inability to integrate the traumatic experience, which generates fear, stress and helplessness action [9]. In recent years a growing number of studies have tried to improve the understanding of the relationship between IPV and PTSD, starting from cognitive impairments mentioned above; for this reason, it has been particularly used a survey technique capable of accurately detecting, from the point of view of spatial and temporal, cognitive alterations: EEG. This technique permits to analyze the event potential related, and to dwell mainly on the N200 and P300 components; both associated with attentional and mnemonic skills but refer to different processes [10]. The N200 component appears to be related to the presence of stimuli with a localization waiting; it seems to reflect a stimulus discrimination process, and not a simple detection of the same. The process of discrimination that correlates with the elicitation

of the component is also manifested in a greater level of arousal, Which the corresponding portion to the physiological found in women with histories of IPV and resulting PTSD [11]. About the P300 component, a meta -analysis of studies examining the response of the P3a elicited by destroyers related to trauma found a higher amplitude of the component in the subjects with PTSD compared with control subjects [12].

1

Method

1.1

Subjects

The research sample is constituted by a total of 28 women. The experimental group was composed of 14 women (mean age=39), recruited in a Center against Women Violence. The control group was composed of 14 women (mean=33) without PTDS or IPV nor depressive/ anxiety symptomatology. 1.2

Experiments

The project of this research included the EEG recording of subjects, through BrainAmp device, with the Brain Vision Recorder Software (© 2010 Brain Products GmbH), during the execution of a Go-NoGo Task with Emotional Visual Images, the results of which were subjected to analysis using the Brain Vision Analyzer Software (© 2010 Brain Products GmbH). Before the execution of the task all subjects filled informed consent in according to the Helsinki Declaration, Anamnestic data, the Beck Depression Inventory-II, the Beck Anxiety Inventory - Coradeschi and a PTSD questionnaire. The Beck Anxiety Inventory is one of self-report instrument consisting of 21 items, which allows to make an assessment of the chronicity and severity of anxiety disorder in adults, including those symptoms only minimally overlap with those of depressive nature. In this experiment we administered the Italian version of Coradeschi. The Beck Depression Inventory is a self-assessment instrument consisting of 21 multiple-choice items. It allows to measure the severity of depression in adults and adolescents aged 13 to 80 years of age. The PTSD questionnaire was drafted by invoking diagnostic and temporal criteria of the DSM-IV-TR, which is also one self-report instrument consisting of 19 items. Each item is rated on a scale of four answers, for nothing, from less than a month, for over a month, for more than three months. After this first step was preparing the subject to the EEG recording, which was through the international system 10-20. In our study, the subjects performed a Go- NoGo task, consists of the continuous repetition of two types of stimuli, one of which was presented repeatedly and the other presented only sporadically. Each subject had to press a button every time it appeared the image emotionally significant. For this task we used E-prime 2.0, an application of Psychology software tools, Inc. The stimuli presented were 60, of these 12 target, called T2, had negative valence and 48 non- target, referred to as T1, with positive or neutral value. The task required 20 minutes and consists of the random image playback, with a interstimulus duration of 1000 ms.

2

Data Analysis and results.

2.1

EEG datasets

The dataset, object of study, included the features of the ERP wave, such as amplitude and latency of the N200, P300 and N500 components. As already mentioned, the collected data concerned 28 individuals divided into 2 classes, 14 with PTSD and 14 in the control group. For each observation we had 15 attributes, corresponding to 15 electrodes (F1, F2, ..., F15). The correspondence between features and EEG channels is shown in Table 1. Table 1. Mapping of Features and EEG Channels Feat Eeg

2.2

F1

F2

F3

F4

F5

F6

F7

F8

F9

F10

F11

F12

F13

F14

F15

Fp1

Fp2

Fz

Cz

Pz

F3

F4

F7

F8

C3

C4

P3

P4

O1

O2

Data analysis

We faced the problem of understanding which features were significant to discriminate against individuals belonging to the 2 classes, so as to reduce the number of the 15 features using the two tailed Student t-test, and then applying a supervised learning paradigm with an artificial neural network, to validate these subsets with statistical indexes. Once we selected the statistically significant features we examined whether and how these subsets of features could be considered typical of the phenomenon under study. For each data collection (both the amplitude and the latency of N200, P200, N500) we did the following. For each index i, we computed the average of the two classes for the i-th feature. 1 𝜇𝑖,𝑐 = ∑14 𝑥 , 𝑖 = 1,2, … ,15; 𝑐 = 1,2. (1) 14 𝑗=1 𝑖,𝑗 Then we used the standard “Student” t-test to determine whether or not there was a statistically significant difference between the two means [13]. Table 2. Significant features with the corresponding p-value, for each ERP wave component. ERP

N200 latency

P300 latency

N500 latency

Significant features F1 F2 F8 F9 F13 F4 F5 F10 F11 F13 F14 F5 F11

p-value < 0.05 0.0148 0.0131 0.0099 0.0060 0.0290 0.0028 0.0058 0.0420 0.0286 0.0013 0.0074 0.0493 0.0065

ERP

P300 amplitude

N500 amplitude

Significant features F4 F5 F13 F15 F4 F5 F10 F11 F12 F13

p-value < 0.05 0.0309 0.0036 0.0083 0.0314 0.0020 0.0010 0.0293 0.0262 0.0059 0.0011

The significance level could be set at any level we wish, but in biology it is common accepting the difference as being significant if the probability of it occurring by chance

is less than 0.05, that is, the null hypothesis that the two means were equal could be rejected at a 95% level of confidence. This produced a reduction of the number of features, since only features with a p-value < 0.05 were taken. This inferential statistical analysis revealed the following: all the datasets showed a certain number of significant features, except for the one relating to amplitude of N200 component. The significant features are shown with the corresponding p-value in Table 2. 2.3

Binary Neural Classifier

The aim of adopting artificial neural networks consists of having a virtual instrument of measure for distinguishing automatically the individuals between the two given classes (PTSD or control) and for estimating the goodness and the representative degree of a subset of features, for each ERP wave component (in Table 2). In this way, the ability of each of those five significative subsets to discriminate the two classes was estimated by measuring the performances of a specific classification process. For evaluating these performances, we used an iterative scheme in which, for each input pattern, 10 neural networks are trained, and the average indexes are calculated: Avg Accuracy, Avg AUC, Avg Sensitivity and Avg Specificity. Table 3. A confusion matrix allows visualization of the classification algorithm performances. Expert Advices

Class 1

PTSD

Control

Total

TP

FP

Class 1 (TP+FP) Class 2 (FN+TN)

Classifier Class 2

FN

TN

Total

PTSD (TP+FN)

Control (FP+TN)

𝑠𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 =

𝑠𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 =

𝑇𝑃 𝑇𝑃+𝐹𝑁

𝑇𝑁 𝑇𝑁+𝐹𝑃

𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑇𝑃+𝑇𝑁

(2)

(3)

(4)

𝑇𝑃+𝐹𝑃+𝐹𝑁+𝑇𝑁

In Table 3, the confusion matrix showing the number of false positives (FP), false negatives (FN), true positives (TP), and true negatives (TN). Sensitivity is the ability of the classifier to identify the positive events among the truly positive. Specificity is the ability of the classifier to identify the negative between the truly negative and both indexes are used in the ROC Analysis. Accuracy is the global concordance of the true positive and negative results in the whole set of subjects [14] [15]. AUC index (Area Under the ROC Curve Index) is used to measure the performance of the ROC curve, which is a kind of measure for classification systems and diagnostic systems. The topology used to implement the neural network [16] [17] was a Multilayer Perceptron, in particular a feedforward two-layer perceptron, with 𝑛𝑖𝑛𝑝𝑢𝑡 (5) neurons in the input layer and 2 neurons in the output layer, where for convention one neuron

corresponds to the class 1 and represented the PTSD individuals and the other one corresponds to the class 2 and represents the control group. For all neurons was used the logistic activation function (6). 𝑛𝑖𝑛𝑝𝑢𝑡 = 𝑐𝑒𝑖𝑙𝑖𝑛𝑔 (

𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑓𝑒𝑎𝑡𝑢𝑟𝑒𝑠+1

𝑓(𝑥) =

1

2

)

(5) (6)

1+𝑒 −𝑥

The input data for the neural network were the six ERP wave components (amplitude and latency of N200, P300 and N500, independently). In training we used the following settings: random initialization of synaptic weights with values between ± 1, learning rate of 0.3 and momentum of 0.2, method of cross-validation with a training time of 500 epochs [18].

2.4

Results

The comparison of the results in classification obtained for each of the five ERP wave components in Table 2, using the full data set input and the dataset restricted to statistically significant features, allowed us to observe that only N200 latency and P300 amplitude produced the best classification performances employing the corresponding restricted datasets, as shown in Table 4. In all other cases (P300 latency, N500 both latency and amplitude), the performances were worse than using the full datasets with 15 features. On average, the PTSD group showed, for each significant feature, higher values of N200 latency and lower values of P300 amplitude than the control one. In Tables 5 and 6, respectively, the results of mean and variance of the samples with and without PTSD are shown, for both the N200 latency and the P300 amplitude (see Fig.1 and Fig.2 for ERPs Waves Components). Table 4. Performances of the N200 latency and P300 amplitude, using the full data set input and the dataset restricted to the statistically significant features. ERP N200 Latency P300 Amplitude

Dataset Full Set [F1,F2,F8,F9,F13] [F2,F8,F9] Full Set [F4,F5,F13,F15] [F5]

Avg Accuracy [%] 70.83 81.67 85.00 68.33 75.00 80.00

Avg AUC 0.75 0.85 0.86 0.57 0.59 0.77

Avg Sensitivity 0.55 0.95 0.95 0.50 0.55 0.55

Avg Specificity 0.80 0.75 0.80 0.82 0.87 0.97

The corresponding electrode locations to the statistically significant features of the ERP wave components represent the ROIs (Regions of Interest) of the specific phenomenon. In the case of N200 latency the ROI was composed by F1, F2, F8, F9, F13; in the case of P300 amplitude the ROI covers the fully connected area composed by F4, F5, F13,

F15. Then we looked for a smaller set of features, following the improvement in performance of the neural classifier. So we obtained the optimum sets and concluded that the significant features for discriminating individuals in the two classes could be reduced, respectively, to F2, F8, F9 and to F5 (Table 4).

Fig.1 Matching ERPs: Red Line IPV group; Black Line Control Group on Occipital Right Channel (O2 Electrode, Feature F15)

Fig.2 Matching ERPs: Red Line IPV group; Black Line Control Group on Parietal Right Channel (P4 Electrode, Feature F13)

3

Discussion

The results showed as much the presence of IPV and subsequent development of PTSD as may impair the ability of the people to recognize emotional stimuli. The alteration of this capacity has obvious repercussions on people's lives, because it involves daily difficulties in interpersonal relationships. In light of the findings it is clear that individuals with histories of IPV and suffering from PTSD exhibit a higher sensitivity threshold than to stimuli involving violence in general or specifically to women. Consider the constant stress that these people receive or have received, it is likely that for them is necessary a more specific input to reset the threshold of sensitivity, for this type of stimulus, at a lower level; it is relevant to consider that the specific input depends on the traumatic experience of the person. Considering the experience of these people is possible to suppose that a higher sensitivity threshold may represent the physiological part of an avoidance behavior; this behavior makes the person unable to take on a specific stimulus or that in some way may be associated with an event so traumatic, for instance not remember it. Although the P300 seems to be the elective component for the study of PTSD, specifically for the attention processes, the results of different studies reported difference

regarding the methodological and the sampling. In addition, it is important to emphasize that the P300 can be influenced by factors such as medications, the chronicity of the disorder, the absence or presence of correlation between stimulus and trauma, or even if the themselves stimuli may or may not act as a distractor during the task [19]. These considerations deserve attention because other studies have shown different results with those of our experiment [20]. Future studies may help to clarify the contribution of the component in question and quantify the interference with this by the factors mentioned above. Some studies suggest that possible attentional difficulties may be associated with the trauma itself rather than to PTSD [21] [22] [23]. The results therefore represent the different faces of the same coin; in other words, the difficulties consistent with the development of PTSD in women with IPV come to light in many forms, but these difficulties are in equal measure intrusive in people's daily life; it therefore seems necessary to continue toward an ever greater understanding of the phenomenon care about all the variables related to the problem. Acknowledgments: ‘Università del Salento - ‘5 for Thousand Research Fund’

References [1]

A. Romero-Martinez, M. Lila, R. Conchell, E. Gonzalez-Bono e L. MayaAlbiol, «Immunoglobulin a response to acute stress in intimate partner violence perpetrators: The role of anger expression-out and testosterone.,» pp. 66-71, 2014.

[2]

A. Romero-Martinez, M. Lila, R. Williams, E. Gonzalez-Bono e L. MoyaAlbiol, «Skin conductance rises in preparation and recovery to psychosocial stress and its relationship with impulsivity and testosterone in intimate partner violence perpetrators,» international journal of psychophysiology , pp. 329333, 2013.

[3]

J. Y.-H. Wong, D. Y. T. Fong, V. Lai e A. Tiwari, «Bridging intimate partner violence and the Human Brain: A Literature Review,» Trauma,Violence & Abuse, vol. 15, pp. 22-33, 2014.

[4]

J. Y.-H. Wong, D. Y. T. Fong, V. Lai e A. Tiwari, «Bridging intimate partner violence and the Human Brain: A Literature Review.,» Trauma,Violence & Abuse, 15, pp. 22-33., 2014.

[5]

L. Almli, N. Fani, K. Ressler e A. Smith, «Genetic approaches to understanding post-traumatic stress disorder,» Int. J. Neuropsychopharmacol., pp. 355-370, 2014.

[6]

G. Harvey, R. Bryant e S. Dang, «Autobioghraphical memory in acute stress disorder.,» Journal of consulting and clinical psychology, pp. 500-506, 1998.

[7]

J. Vasterling, L. Duke, K. Brailey, J. Constans, A. Allain Jr e P. Sutker, «Attention, learning, and memory performance and intellectual resources in Vietnam veterans; PTSD and no disorder comparations.,» Neuropsychology, vol. 16, pp. 5-14, 2002.

[8]

M. Wheeler e R. Buckner, «Functional-anatomic correlates of remembering and Knowing.,» Neuroimage, pp. 1337-1349, 2004.

[9]

C. Nemeroff e J. Sherin, «Post traumatic stress disorder: the neurobiological impact of psychological trauma,» Dialogues in Clinical Neuroscience, vol. 13, n. 3, pp. 263-278, 2011.

[10]

S. Luck, «An introduction to the Event- Related Potential Tecnique.,» The MIT Press, pp. 1-357, 2005.

[11]

E. K. Vogel e S. J. Luck, «The visual N1 component as an index of a discrimination process.,» Psychophysiology, vol. 37, pp. 190-203, 2000.

[12]

A. Karl, L. Malta e A. Maercker, «Meta-analytic review of event-related potential studies in post-traumatic stress disorder,» Biological Psychology, vol. 71, pp. 123-147, 2006.

[13]

M. E. Ahsen, N. K. Singh, T. Boren, M. Vidyasagar e M. A. White, «A new feature selection algorithm for two-class classification problems and application to endometrial cancer,» in 51st IEEE Annual Conference on Decision and Control (CDC), Maui, Hawaii, USA, IEEE, 2012, pp. 29762982.

[14]

G. R. Scolaro e F. M. De Azevedo, «Classification of epileptiform events in raw EEG signals using neural classifier,» in 3rd IEEE International Conference on Computer Science and Information Technology (ICCSIT), IEEE, 2010, pp. 368-372, Vol. 5.

[15]

V. Bevilacqua, G. Mastronardi, F. Menolascina, P. Pannarale e A. Pedone, «A novel multi-objective genetic algorithm approach to artificial neural network topology optimisation: the breast cancer classification problem,» in International Joint Conference on Neural Networks, Vancouver, BC, Canada, IEEE, 2006, pp. 1958-1965.

[16]

V. Bevilacqua, F. Cassano, E. Mininno e G. Iacca, «Optimizing FeedForward Neural Network Topology by Multi-objective Evolutionary Algorithms: A Comparative Study on Biomedical Datasets.,» in Advances in Artificial Life, Evolutionary Computation and Systems Chemistry, Springer International Publishing, 2015, pp. 53-64.

[17]

M. A. Sovierzoski, F. I. M. Argoud e F. M. de Azevedo, «Evaluation of ANN classifiers during supervised training with roc analysis and cross validation.,» in International Conference on BMEI, IEEE, 2008, pp. 274-278, Vol. 1.

[18]

J. F. Jekel, D. L. Katz, J. G. Elmore e D. Wild, Epidemiology, biostatistics and preventive medicine. Elsevier Health Sciences., Elsevier Health Sciences, 2007.

[19]

M. Kimble, D. Kaloupek e M. Kaufman, «Stimulus Novelty Differentially Affects Attentional Allocation in PTSD,» Biol Psychiatry, pp. 880-890, 2000.

[20]

A. Javanbakht, I. Liberzon, A. Amirsadri, K. Gjini e N. Boutros, «Eventrelates potential studies of post-traumatic stress disorder: a critical review and synthesis.,» Biology of mood & anxiety disorder, pp. 1-5, 2011.

[21]

A. Karl, M. Schaefer, L. Malta, D. Dorfel, N. Rohlender e A. Werner, «A meta-analysis of structural brain abnormalities in PTSD,» Neuroscience and Biobehavioral Reviews, pp. 1004-1031, 2006.

[22]

M. Kimble, B. Frueh e L. Marks, «Dose the modified Stroop effect exist in PTSD? Evidence from dissertation abstract and the peer reviewed literature.,» Journal of Anxiety Disorder., vol. 23, n. 5, pp. 650-655, 2009.

[23]

M. Kimble, K. Fleming, C. Bandy, J. Kim e A. Zambetti, «Eye tracking and visual attention to threating stimuli in veterans of the Iraq war.,» Journal of anxiety disorders, vol. 24, pp. 293-299, 2010.

[24]

W. WHO, «Violence against woman. A priority health issue,» Womens's Health and Development, pp. 1-28, 1997.

[25]

W. WHO, «World report on violence and health: Summary,» World Health Organization , pp. 1-37, 2002.

[26]

P. A. Ali e P. Naylor, «Intimate partner violence: a narrative review of the biologica and psychological expanations for its causation,» Aggression and Violent Behavior , pp. 373-382, 2013.

[27]

H. Soler, P. Vinayak e D. Quadagno, «Biosocial Aspects of domestic violence,» Psychoneuroendocrinology, n. 25, pp. 721-739, 2000.

[28]

R. Jewkes, «Intimate partner violence : cause and prevention.,» The lancet, vol. 359, n. 9315, pp. 1423-1229, 2002.