Eye Gaze Metrics and Analysis of AOI for Indexing Working Memory towards Predicting ADHD

Ey e Gaze Metrics and Analysis of A OI for Indexing W orking Memory to w ards Predicting ADHD Ga vindy a Ja y aw ardena 1 , Anne Mic halek 2 , and Sampath Jay arathna 1 1 Departmen t of Computer Science 2 Departmen t of Communication Disorders & Sp ecial Education Old Dominion Univ ersity , Norfolk, V A 23529 gavindya@cs.odu.edu, aperrott@odu.edu, sampath@cs.odu.edu Abstract : ADHD is b eing recognized as a diagnosis whic h p ersists into adulthoo d impacting economic, o ccupational, and educational outcomes. There is an increased need to accurately diagnose and recommend in terven tions for this population. One consideration is the developmen t and implementation of reliable and v alid outcome measures whic h reflect core diagnostic criteria. F or example, adults with ADHD ha ve reduced working memory capacit y when compared to their p eers (Michalek et al., 2014). A reduction in w orking memory capacit y indicates attentional con trol deficits whic h align with man y symptoms outlined on b eha vioral c hecklists used to diagnose ADHD. Using computational metho ds, suc h as eye trac king technology , to generate a relationship b etw een ADHD and measures of working memory capacity w ould b e useful to adv ancing our understanding and treatmen t of the diagnosis in adults. This c hapter will outline a feasibilit y study in which eye tracking was used to measure eye gaze metrics during a working memory capacit y task for adults with and without ADHD and mac hine learning algorithms w ere applied to generate a feature set unique to the ADHD diagnosis. The chapter will summarize the purpos e, metho ds, results, and impact of this study . Keyw ords: ADHD, eye tracking, w orking memory capacity 1 In tro duction A ttention-Deficit/Hyperactivity Disorder is b eing recognized as a diagnosis which p ersists into adultho o d impacting economic, o ccupational, and educational outcomes. Estimates indicate that 3-5% of adults hav e a diagnosis of ADHD [48] with prev alence estimated to hav e increased from 6.1% of the United States p opulation in 1997 to 10.2% of the p opulation in 2016 [49]. The disorder is behaviorally marked b y difficulty with attention to imp ortan t details, difficult y initiating and completing tasks, and difficult y mo dulating b eha viors appropriately in relation to the situation [14, 16]. According to Barkley [3], adult ADHD symptoms result from impairments of inhibition or the inabilit y to regulate and mo dulate prep otent resp onses. While a diagnosis of adult ADHD presumes disinhibition, little is known ab out the physiological underpinnings of that cognitive skill in relation to an adult ADHD diagnosis. There is an increased need to accurately diagnose ADHD through the dev elopment and implementation of ob jectiv e and reliable outcome measures whic h reflect core diagnostic criteria, lik e inhibition. Researc hers in cognitive psyc hology evidence attention con trol as the measurable psyc hological construct whic h facilitates inhibitory resp onses by allo cating attention according to task demands, esp ecially in the presence of distracting stimuli [9, 12, 21]. A ttention con trol differentiates success during tasks requiring in tentional and sustained constraints for effective inhibition, lik e dichotic listening [7] or pro cessing sp eec h in noise [43]. Measurements of attention con trol are demonstrated through differences in w orking memory capacit y (WMC) accounting for appro ximately 60% of the v ariance seen across p eople on measures of WMC, lik e complex span tasks [13]. Adults with ADHD hav e reduced WMC when compared to their p eers [32] and, despite the understanding that disinhibition is central to an ADHD diagnosis and differences in WMC 1 mathematically represent the resource which makes inhibition p ossible, there is a paucity of research inv es- tigating ph ysiological responses during measures of WMC which could differen tiate adults with and without ADHD. The tw o goals of this w ork is to determine the feasibility of iden tifying and in tegrating ey e gaze metrics from a WMC task using mac hine learning to generate a v alid and reliable feature set whic h indexes and predicts an ADHD diagnosis and understand the utility of area of in terest (A OI) in order to capture eye gaze metrics and predict ADHD in the context of a WMC task using machine learning algorithms. This chapter in vestigates gaze measures that map onto these v alid neuro cognitiv e deficits that are cen tral to ADHD within the con text of a WMC task. In addition, this chapter inv estigates the differences b etw een ADHD and non-ADHD participants by analyzing t wo main sequence relationships of saccade amplitude to wards saccade duration and p eak velocity . The developmen t of these ob jectiv e measures of ADHD will facilitate its diagnosis and reveal strategies that can enhance the future design effective interv en tion strategies and accessible classro om en vironments. 2 Bac kground 2.1 W orking Memory Capacit y (WMC) T asks W orking memory is the cognitiv e system which mak es it p ossible to mentally hold and manipulate information sim ultaneously . Ov er a decade of work b y Engle and colleagues [8] supp orts the use of complex span tasks as not only measures of w orking memory but as a reflection of individual differences in WMC. Performance on complex span tasks generates a comp osite working memory score numerically indicating ho w well someone can manipulate and hold information. Ho wev er, differences in WMC or that comp osite score represent a p ersons ability to mo derate and control attention. Adults with ADHD hav e reduced working memory when compared to their p eers demonstrating significant differences in WMC [2]. This finding suggests that adults with ADHD hav e a reduced abilit y to monitor and con trol attention, esp ecially during situations with comp eting stimuli [32] or that require resp onse inhibition [26, 41]. Little is known ab out the underlying co vert pro cesses engaged during inhibitory tasks which rely on attention allo cation. Using physiological measures during a task which v alidly reflects attention control, like a complex span task, pro vides ob jective diagnostic information for adults with ADHD. The reading span task (R-Span) is one complex span task widely used as a v alid measure of w orking memory yielding a WMC score [9]. The R-span was originally dev elop ed by Daneman and Carp en ter (1983) [10] as a predictor of reading comprehension and was subsequently mo dified by Engle and colleagues [13]. During the R-Span people see one sen tence on a computer screen, read the sentence out loud, determine if the sen tence is meaningful with a y es or no resp onse, and then verbally identify a letter t yp ed at the end of each sen tence. After a set of sen tences, the p erson is asked to verbally recall as many letters as p ossible in order of presentation. This task represents the p ersons ability to hold and manipulate information sim ultaneously . T o date, there hav e b een no empirical studies in vestigating eye gaze metrics collected during this task whic h migh t differen tiate p erformance and further explain diagnostic differences for adults with and without ADHD. 2.2 Mac hine Learning and ADHD T ypically , exp erts diagnose ADHD using sub jective chec klists and related academic and cognitive p erfor- mance measures [18]. These comprehensive assessments can b e time consuming, inconsistent, and can inaccurately represent deficits making differential diagnosis challenging. Ideally , it would b e more efficient and reliable to develop predictive algorithms based on physiological metrics reflecting core diagnostic crite- ria. How ev er, designing this t yp e of computer program can b e difficult b ecause there is not an existing set of confirmed mathematical features which accurately differenti ate betw een adults with and without ADHD. Mac hine learning principles offer a solution to this barrier. While it is not practical to develop algorithms b y pro viding a sp ecific set of instructions, mac hine learning uses numeric features represen ting a cognitiv e skill to teach the computer data patterns and inferences which can b e applied to groups of data for predicting accurate classifications. In machine learning there are several v arieties of subcategorical learning algorithms. Sup ervised learning is an example of suc h a sub category . The core core ob jectiv e of sup ervised learning is to build a mathematical 2 mo del whic h can be used to predict the outputs of new samples using the training data. Usually , the training data set is stored in a matrix. Eac h row of the training data matrix corresp onds to one training instance whic h also con tains the desired output. F or example, in the ADHD/Non-ADHD literature par sup ervised learning algorithms are iterated through the training dataset to learn a mathematical mo del to predict or classify the output asso ciated with unseen inputs. When determining whether a p erson do es or do es not ha ve ADHD using eye gaze as the outcome metric, the training data would consist of eye tracking data of eac h p erson and each p erson w ould hav e a class specifying whether that p erson is identified as ha ving ADHD or not having ADHD. Sup ervised learning algorithm will build a general mathematical mo del whic h co vers the training data space. When a previously unseen p erson’s e y e gaze data is entered, the model will use its past exp erience to accurately predict whether or not that p erson has ADHD. Some exp erts may question the reliability and v alidity of using machine learning for predicted outputs b ecause those outputs do not consider exp erts from psychology or medicine in the decision pro cess. Ev en with the inv olv ement of expert physicians, it is rep orted that diagnostic errors contribute to approximately 10 p ercen t of patient deaths, by Institute of Medicine at the National Academies of Science, Engineering and Medicine[38]. Causes for such diagnostic errors could b e communication errors b etw een patients and ph ysicians and other failures of the healthcare system. These challenges could b e addressed by identifying patterns of the symptoms patien ts confirm and use them to predict p oten tial diagnostic co des. Ev en if there is a lack of communication betw een the patient and the physician or there is a failure in a healthcare system, the symptom patterns of the patien t would b e highlighted so that accurate diagnosis prediction could b e facilitated. Curren tly , mac hine learning is b eing used for diagnostic prediction not only based on reported symptoms, but also based on patien t history and data extracted from w earable devices. Classification algorithms compare the symptoms pattern of the patient and other related data with the other patien ts in the training dataset in order to predict an accurate diagnosis. In literature using machine learning for ADHD diagnosis and classification, the first attempt to classify adult ADHD patients and healthy controls using a mac hine learning algorithm is [36]. They conducted research to classify ADHD patients and healthy controls using supp ort vector machine (SVM) learning based on even t related potential (ERP) components. They examined data from 148 adult participants. Among them, 50% [36] were diagnosed as ADHD while the rest did not ha ve a diagnosis of ADHD. Both groups of adults w ere selected in a manner that age and gender did not v ary b et ween the tw o groups [36]. Each participant p erformed a visual tw o stimulus GO/NOGO task [36] and ERP resp onses of participants were decomp osed in to indep endent comp onents and created the feature set. Classification accuracy of assigning ADHD participan ts and healthy controls to the corresp onding groups using a non-linear SVM with 10-fold cross-v alidation was 92% [36], whereas it w as 90% [36] for linear SVM. This researc h suggests that classification by means of non-linear methods is more accurate for exp erimen ts conducted in a clinical context [36]. Ev en-though this study uses machine learning approaches to predict ADHD, it do es not use eye gaze metrics nor working memory capacit y . [39] shows that extreme learning machine (ELM), a machine learning algorithm, achiev es 90.18% [39] accuracy when predicting ADHD using structural MRI data. This study confirmed that linear support v ector mac hine and support vector machine-RBF achiev es an accuracy of 84.73% and 86.55% [39] respectively when the same structural MRI dataset is used. Both extreme learning mac hine and supp ort v ector mac hine ha ve b een ev aluated to find the classification accuracy using cross-v alidation. The goal of their study was proposing an ADHD classification mo del using the extreme learning machine (ELM) algorithm for ADHD diagnosis. They assessed the computational efficiency and the effect of sample size on b oth extreme learning machine and supp ort vector machine. They acquired MRI images from 110 participants with 50% of them ha ving a diagnosis of ADHD [39]. This study gives us insight ab out how applying a mac hine learning mo del can accurately predict ADHD. Moreo ver, [31] aimed to classify p eople with ADHD and wih tout ADHD using autoregressiv e mo dels. They used EEG data collected using 26 electrodes from a group of children b etw een the ages of 6 and 8 [31] to discriminate b et ween ADHD and Non-ADHD. Children participated in multiple exp erimental conditions, suc h as eyes op en, eyes closed, and quiet video baseline tasks while collecting EEG data [31]. This study ferified that KNN classifier is able to provide high classification accuracy when classifying c hildren as either ADHD or t ypically delv eoping ADHD. The accuracy achiev ed in this study w as high and v arying b et ween 85% and 95% [31]. Finally , [1] used EEG data with semi-sup ervised learning in order to predict a ADHD and Non-ADHD 3 diagnosis. They had 10 c hildren participan ts with 7 of them ha ving a diagnosis of ADHD and 3 were typically dev eloping [1]. They trained and tested support vector mac hine with EEG data of eac h participan t achieving an accuracy of 97% [1] for ADHD prediction using supp ort v ector mac hine learning. T aken together, the literature affirms the successful use of machine learning to accurately predicate a diagnosis of ADHD. Ho wev er, an empirical gap exists with regard to the training dataset used to train the mac hine learning mo del. The ma jority of studies conducted hav e primarily used MRI, fMRI, or EEG data to train the machine learning mo del used to discriminate b etw een ADHD and Non-ADHD. One of our goal of this pro ject is to predict diagnosis of ADHD using ey e gaze metrics and measures of working memory as the training data for the machine learning algorithms. In this pro ject we predicted a diagnosis of ADHD using eye gaze metrics collected during a WMC task. The output of our task is limited to tw o classes, ADHD and Non-ADHD. W e used sup ervised learning classification algorithms and w e ev aluated the predicted output in terms of ho w close it is to the actual output. W e ev aluated outcomes using an accuracy , precision, recall, and f-measure ev aluation metrics. T ogether these metrics ga ve us an go od understanding ab out the p erformance of the classifier. 2.3 Area of In terests and A ttention T asks The studies [34] and [45] hav e considered A OIs of stimuli for statistical analysis of the eye tracking. The study [20] compared b orderline p ersonalit y disorder (BPD) patien ts to Cluster-C p ersonality disorder (CC) patien ts and non-patients (NP) regarding emotion recognition in ambiguous faces and their visual atten tion allo cation to the ey es which is the A OI. The authors hav e found BPD ha ve a biased visual attention to wards the eyes. [15] is a another study which inv estigates the immediate effects of coloured ov erlays on reading p erformance of presc ho ol children with ASD. The authors of the study hav e used eye trac king and concluded that coloured o verla ys ma y not b e useful to improv e reading and o cular p erformance in c hildren with ASD in a single o ccasion [15]. According to the literature, AOIs hav e b een used in m ultiple studies related to v arious attention tasks. Ev en though AOIs ha ve b een used, there is a paucit y of using eye mo vemen t fixations and saccades occurred within the A OIs when completing a WMC measure for ADHD classification using machine learning approac h. W e look at the consistency and stability of ey e mo vemen t fixations and saccades occurred within the A OIs of stim uli when completing a WMC measure. This is an imp ortan t line of inquiry because it in vestigates ho w relev ance ma y be reflected in ey e mov ements features for atypical and complex atten tional systems, such as in the con text of ADHD. 2.4 Ey e Mov emen ts and ADHD Ey e mo vemen t b ehavior is a result of complex cognitiv e processes; therefore, eye gaze metrics can reveal ob jective and quan tifiable information ab out the qualit y , predictability , and consistency of these cov ert pro cesses [47]. Ey e gaze measurement includes a num b er of metrics relev an t to o culomotor control [22] including saccadic tra jectories, fixations, and other relev an t measures - suc h as v elocity , duration, amplitude, pupil dilation [25]. A saccade (rapid ey e mov ement from one fixation point to another) itself may not be an informativ e indicator of cognition since visual p erception is suppressed during a saccade. How ev er, fixations require preceding saccades to help place the gaze on target stim uli to gather salien t and relev ant information. W e b eliev e that analysis of these eye mo vemen ts can pro vide imp ortan t cum ulativ e clues ab out the underlying ph ysiological functions of atten tion con trol during a WMC task whic h can differen tiate a diagnosis of ADHD for adults. There is substantial ov erlap in brain systems that are inv olved in o culomotor control and cognitive dysfunction in ADHD. The precise measurements of eye mov emen ts during cognitively demanding tasks pro vide a window into underlying brain systems affected by ADHD. The neural substrates of o culomotor con trol are well established [28] and show pro ximity to and ov erlap with the cortical and sub cortical structures in volv ed in cognitive dysfunction in ADHD. F or example, the cortical structures that mediate saccadic programming as well as a n umber of saccadic b eha viors include frontal-p arietal areas such the frontal eye field (FEF), supplementary eye field (SEF), parietal ey e field (PEF), and DLPC. These areas are also affected during cognitiv e control and WM in ADHD [44]. With resp ect to sub cortical structures, the accuracy of saccades is main tained via cerebellum. F or example, saccadic h yp ometria is an undershooting of a saccade to 4 a target that is t ypically seen in normal sub jects, whereas saccadic hypermetria, ov ershooting the target, is a hallmark feature of cereb ellar dysfunction [29]. A study of saccades during visuo-spatial WM has rep orted significan t diagnostic group differences in under- versus o ver-shooting to the target b etw een b o ys with ADHD and non-affected controls, suc h that the ADHD group tended to ov ersho ot the target and the control group tended undersho ot the target [42]. Studies ha ve shown that individuals with ADHD also hav e deficits in the suppression of saccades relative to controls [35, 37, 42]. Similarly , p eople with ADHD demonstrate difficulties with inten tionally inhibiting o cular resp onses when compared to their p eers during tasks which require purp oseful an ti-saccade b eha viors [26, 41]. Ey e gaze metrics, esp ecially saccade features, reliably rev eal imp ortan t differences b et ween adults with and without ADHD. Based on the diagnostic utility of eye mov ements, [5] inv ented a metho d which determines whether an individual has ADHD by sampling the eye mo vemen ts of participants when they are in an inactiv e state. P atented as ”ADHD detection by eye saccades” [5], their pro cedure includes a sampling device which has infrared radiation for brightening the eye of a participant and detecting reflections from the eye. The eye mo vemen t data collected using their device determines the v alue of a pre-selected parameter whic h has a threshold v alue indicating whether the participant has ADHD or not. According to their study , the most significan t feature of the eye mov ement data is the angular acceleration of the eyeball [5]. They hav e measured ocular angular acceleration for the participants by asking them to stare at a blank screen [5]. The measuremen t data of the angular acceleration of the eye b elo w the threshold v alue indicates diagnosis of ADHD and the data ab ov e the threshold v alue constitutes a classification of health y/normal. Although the study conducted by [5] used on a mec hanical device with infrared radiation and not machine learning, the fact that it acquired a patent supports its strong impact and confirms that ey e measurements could b e used to diagnose ADHD. 2.5 Ey e Gaze and Machine Learning and ADHD There is a paucity of empirical studies which implement machine learning to predict ADHD classification using a measures of atten tion control or WMC. How ev er, there are a few inv estigations whic h use maching learning in com bination with measures of inhibition. [19] measured activ ation patterns using functional mag- netic resonance imaging while adolescen ts with ADHD performed a Stop T ask. During this task, participan ts had to suppress or inhibit the motoric resp onse of pushing a button. The researchers used Gaussian pro cess classifiers and whole activ ation pattern analysis and were able to predict the ADHD diagnosis with 77% accuracy [19]. Lik ewise, in a study with adults with and without a diagnosis of ADHD, mac hine learning predicted the diagnosis with a specificity of .91 and sensitivity of .76 based on EEG metrics during a NoGo task measuring inhibition [4]. These results supp ort collecting ph ysiological metrics during tasks required atten tion con trol to generate pattern recognition analysis for the accurate classification of ADHD. Similarly , only one study was lo cated which used eye mo vemen ts in conjunction with mac hine learning to predict ADHD [46]. This study included participants diagnosed with ADHD, fetal alcohol sp ectrum disorder (F ASD), and P arkinson’s disease (PD). Researchers presen ted short video clips to each participant and analyzed the resulting data sets for three sp ecific types of eye mov ement features: 1) o culomotor-based features such as fixation durations and distributions of saccade amplitudes; 2) saliency-based features; and 3) group-based features [46]. Results confirmed that saliency based features b est differentiated c hildren with ADHD and F ASD from t ypically dev eloping c hildren. Mac hine learning algorithms predicted ADHD in the sample of c hildren with 77.3% accuracy [46]. T aken together, these empirical findings suggest that diagnostic biomark ers of ADHD could b e generated from eye gaze metrics during a WMC task using mac hine learning. As such, in this feasibilit y study , w e examined patterns of saccades and stabilit y of fixations generated when completing a measure of WMC to create a feature set which could b e used to differentiate a diagnosis of ADHD for adults. Based on the evidence that WMC is reduced in adults with ADHD [32], measurement of eye mo vemen ts during a measure of WMC will address the following researc h question: 1) do eye gaze feature v alues indexing a WMC task predict the classification of ADHD in adults? 5 3 Metho dology 3.1 P articipan ts A total of 14 adult participants without (n = 7) and with a diagnosis (n = 7) b etw een the ages of 18-35 w ere recruited for this study from an higher education institution in the mid-Southeastern United States. The sev en adult participants w ere (6 F, 1 M, M age=22.85, SD age=3.01) diagnosed with ADHD b y medical practitioners and that diagnsosis was confirmed through formal and verified do cumentation. Eac h ADHD participan t also completed an informational interview v erbally confirming their diagnosis. Moreo ver, adults with ADHD remained medication free for the 12 hours prior to study participation. Prior to b eginning study tasks, all adults were informed of their risks regarding remaining medication free and participating in the study . Participan ts provided their consent by signing forms outlining costs and b enefits of participation appro ved the Universit y’s Institutional Review Board (IRB) in accordance with the Helsinki Declaration. P articipants who completed the protocol w ere giv en a ten dollar Amazon or Chick-Fil-A gift card. 3.2 W orking Memory Capacit y T ask WMC is reflected through complex span tasks, including the Reading Span (R-Span). The R-Span is a v alidated task designed to reflect the cognitive system’s ability to main tain activ ated representations [13, 12]. In the R-Span task, participants are asked to read a sentence and letter they see on a computer screen. Sen tences are presented in v arying sets of 2-5 sentences. Participan ts are ask ed to judge sentence coherency by saying ’yes’ or ’no’ at the end of each sentence. Then, participan ts are ask ed to remem b er the letter prin ted at the end of the sentence. After a 2-5 sentence set, participan ts are asked to recall all the letters they can remember from that set. WMC scores are generated based on the num be r of letters accurately recalled divided by the total num b er of p ossible letters recalled in order. Ho wev er, this pro ject fo cused on measures of visual attention whic h could differentiate adults with and without ADHD. 3.3 Apparatus Ey e gaze metrics were recorded and analyzed using the T obii Pro X2-60 computer screen-based eye track er with T obii Studio analysis soft ware. The T obii Pro X2-60 records ey e mov emen ts using infrared corneal reflectiv e technology at a sampling rate of 60 Hz (i.e. approximately once every 16.23 milliseconds). Gaze data accuracy was within 0.4 degrees of visual angle and precision was within 0.34 degrees of visual angle. T obii’s eye tracking technology is effective for generating reliable and v alid brain/b eha vior outcomes for c hildren and adolescents [40]. All of the participan ts fulfilled the following inclusion criteria: 1) betw een 18 and 65 years of age, 2) sp ok e English as their first language, 3) self-rep orted normal vision with or without corrective lenses, 4) no history of psychotic symptoms; and 5) no comorbid cognitive impairments (e.g. do cumen ted learning disabilities, reading disabilities). Figure 1: Comparison of Eye Fixatoins for ADHD (Left) and Non-ADHD (Right) particiapnt during WMC T ask. 6 3.4 Ey e Mov emen t F eatures The human o culomotor plant (OP)[23] consists of the eye glob e and six extrao cular m uscles and its sur- rounding tissues, ligaments each con taining thick and thin filamen ts, tendon-lik e comp onents and liquids. In general there are six ma jor eye mo vemen t types: fixations, saccades, smo oth pursuits, optokinetic reflex, v estibule-o cular reflex and vergence[29]. An eye-trac ker provides eye gaze p osition information as well as other gaze related parameters (pupil dilation etc.) so that algorithmic deriv ation in terms of t w o primary ey e mo vemen ts, fixations (relative gaze p osition at one p oint on the screen) and saccades (rapid eye mo vemen ts of gaze from one fixation p oint to another) can b e analyzed to deriv e the users atten tion patterns. W e are interested in inv estigating num b er of eye fixation based features in the current framework. W e dev elop ed a detailed saccade and fixation feature set using the following qualifiers: gender, n umber of fixa- tions, fixation duration measured in milliseconds, a verage fixation duration in milliseconds, fixation standard deviation in milliseconds, pupil diameter left, pupil diameter right, and diagnosis label or class. Due to the sampling rate of the trac king system, w e were not able to calculate microsaccades and ov ersho ot/undershoot saccades as comp onen ts of the feature set. 3.5 Measuring Atten tion during WMC The data for this study w as collected during a larger pro ject in volving adults with and without ADHD and an audiovisual listening in noise task where WMC scores were measured and used as a cognitive co v ariate without eye tracking metrics. The en tire testing session for the pro ject to ok approximately 45 min utes. The session b egan with the participant interview, explanation of the purp ose of the study , and review of the consen t form. During the interview, the participan ts provided demographic information and were screened to confirm that all inclusion criteria were satisfied. Once participants indicated they understo od their righ ts and ga ve consen t, they entered the testing area to b egin the study . Participan ts sat at a desk in fron t of a Dell Computer with a 21 inc h monitor. The distance and p osition of each participan t was mo dified in order to maintain a 45 degree viewing angel of the monitor. F or each participant, the exp erimen tal tasks b egan with eye gaze calibration. Once calibration was confirmed, participan ts view ed a welcome screen follow ed b y the random presentation of several exp erimen tal tasks, including the RSP AN task. The lo cation of the RSP AN in the order of exp erimental tasks was randomized and counterbalanced across participants in order to main tain v alidit y . Participan ts w ere randomly assigned to a group determining the order of experimental task presentation prior to b eginning the study . F or all of the exp erimen tal tasks, participants were giv en practice trials. 4 Mac hine Learning on Data W e chose precision, recall, f-measure, and accuracy as the ev aluation measures for our w ork. Prior studies [30] ha ve already prov en that these measures are indep endent of category distributions provided that precision and recall are measured at the same time. Intuitiv ely , precision measures exactness of the system (i.e., out of all predicted data instances for a sp ecific category lab el how man y are predicted correctly) while recall indicates the completeness of the system (i.e., out of all lab eled data for a sp ecific a category lab el how man y are predicted correctly). F v alue measures the balance b etw een precision and recall in a single v alue. In our tables with results assessing classifiers, precision, and recall refers to their weigh ted av erage v alues. Accuracy sp ecifies the fraction of the predictions that the classifier predicated correctly . W e emplo yed a grid searc h mechanism to identify the b est parameter com bination for optimal result. The optimal parameters are selected based on p erformance for eac h classifier after a 10-fold cross v alidation. T able 1 sho ws the RSP AN score for the participants in the current study . An independent t-tes t statis- tical analysis (p=0.07) confirms that for this feasibility study there are no significant group differences on WMC scores and are predicted to b e a result of the small sample size. The RSP AN is an individual differ- ences measure and significance in v ariance is detected with large sample sizes. Additionally , WMC scores are typically generated through a comp osite score of t wo or m ore span tasks [9], for example, a previous in vestigation by one of the authors confirmed group differences in WMC using the RSP AN and op eration span (OSP AN) to generate a WMC comp osite score for adults with ADHD [32]. 7 T able 1: RSP AN Score of the ADHD Vs Non-ADHD P articipant Age Gender RSP AN Classification 3 18 F emale 0.86 Non-ADHD 7 35 Male 0.88 Non-ADHD 9 19 F emale 0.60 Non-ADHD 17 23 Male 0.55 Non-ADHD 20 21 F emale 0.57 Non-ADHD 25 32 Male 0.88 Non-ADHD 26 20 F emale 0.74 Non-ADHD 30 21 F emale 0.51 ADHD 34 19 Male 0.67 ADHD 35 26 F emale 0.76 ADHD 36 29 F emale 0.71 ADHD 37 21 F emale 0.60 ADHD 38 21 F emale 0.40 ADHD 47 23 F emale 0.62 ADHD 4.1 Visual Analysis Figure 1 presents images of ey e gaze patterns from t w o adults participants, one with and one without ADHD. Informal visual analysis indicates that the adult with ADHD is fixating primarily b elo w the stimulus items with little direct fixation to sentence components including: the words, the decision point, or the item to b e remem b ered. Unlike the adults with ADHD, the adult without ADHD has a ma jority of fixations which are in-line with all sen tence comp onents. Although this is a conclusion generated from informal visual inspection, it rev eals that adults with ADHD are not visually scanning stim ulus items in a path similar to adults without ADHD. The fixation cluster pattern is just below the stimulus sentence comp onen ts. This is consisten t with the findings of Krejtz et al. (2015) who suggest that while adults with ADHD had similar fixations to salient visual cues when compared to adults without ADHD, they demonstrated less structured and more chaotic scan patterns [24]. 4.2 Mac hine Learning for Classification Prediction W e generated three feature sets categorized according to metric t yp e: 1) fixation feature set; 2) saccade fea- ture set; and 3) saccade and fixation com bination feature set. Each of the three feature sets w ere individually en tered into 43 different classifiers yielding precision rates, recall rates, F1 scores, and p ercen t accuracy . W e iden tified six of the top p erforming classifiers for each of the three feature sets: J48, LMT, RandomF orest, REPT ree, K Star, and Bagging. Results for each feature set are discussed individually . Six of the top p erforming classifiers for the fixation feature set are listed in the T able 2. The Bagging classifier (ensemble meta-estimator) yielded the highest p ercen t accuracy with 78.48% indicating that a fixation feature set collected during a RSP AN task classifies a diagnosis of ADHD with greater than 70% accuracy . The REPT ree classifier yielded the low est p ercent accuracy at 76.77%. T able 2: Classification of Ey e Fixation F eatures during WMC Classifier Precision Recall F1 Accuracy J48 0.77 0.76 0.77 77.79 LMT 0.77 0.77 0.77 77.92 RandomF orest 0.76 0.76 0.76 76.79 REPT ree 0.75 0.76 0.75 76.77 K* 0.76 0.76 0.76 76.92 Bagging 0.77 0.78 0.77 78.48 8 Figure 2: R OC Graph of the T op P erforming Classifiers for Fixation F eature Set. T o further inv estigate the performance metrics for the 6 most effective classifiers for the fixation feature set w e generated a Receiver Op erating Characteristics (ROC) graph (see Figure 2). The R OC graph displa ys the relativ e trade-off b et ween b enefits (true p ositiv e) rates on the Y axis and the costs (false p ositive) rate on the X axis. The graph sho ws the Bagging as our top p erforming classifier offering the b est trade-off in terms of the cost and the b enefits. T able 3 outlines six of the top p erforming classifiers for the saccade feature set. The Random F orest classifier yielded the highest p ercen t accuracy at 91.14% indicating that the saccade feature set collected during a RSP AN task classifies a diagnosis of ADHD with greater than 90% accuracy . The J48 classifier yielded the lo west p ercen t accuracy at 88.95%. T able 3: Classification of Saccade F eatures during WMC Classifier Precision Recall F1 Accuracy J48 0.89 0.89 0.89 88.95 LMT 0.89 0.89 0.89 89.51 RandomF orest 0.91 0.91 0.91 91.14 REPT ree 0.89 0.89 0.89 89.16 K* 0.86 0.86 0.86 85.98 Bagging 0.91 0.91 0.91 90.82 W e generated a ROC graph (see Figure 3) for the classifiers we selected for the saccade feature set to in vestigate the p erformance metrics. The R OC curve shows that Random F orest classifier has the largest Area Under the Curve (AUC) meaning that it has the lo west error. The AUC of Random F orest classifier is 0.9114. Therefore, it has 91.14% c hance of correctly distinguishing betw een ADHD and Non-ADHD. Random F orest is our top p erforming classifier offering the b est trade-off in terms of the cost and the benefits for the saccade feature set. Finally , table 4 provides results of the six top p erforming classifiers for the com bination of saccade and fixations feature set. The Random F orest classifier yielded the highest p ercent accuracy at 91.11% indicating 9 Figure 3: R OC Graph of the T op P erforming Classifiers for Saccade F eature Set. that the combination of fixation and saccade features collected during a RSP AN task classified a diagnosis of ADHD with greater than 90%accuracy . The K Star classifier yielded the low est p ercent accuracy at 77.21%. Figure 4: ROC Graph of the T op P erforming Classifiers for the Combination of Fixation and Saccade F eatures Set. 10 T able 4: Classification of Ey e Fixation and Saccade F eatures during WMC Classifier Precision Recall F1 Accuracy J48 0.89 0.89 0.89 89.19 LMT 0.89 0.89 0.89 89.91 RandomF orest 0.91 0.91 0.91 91.11 REPT ree 0.89 0.89 0.89 89.16 K* 0.77 0.77 0.77 77.21 Bagging 0.91 0.91 0.91 90.83 W e generated a ROC graph (see Figure 4) to inv estigate the p erformance metrics for the classifiers w e selected for the combination of saccade and fixations feature set as well. The graph shows that ev en for the combination of saccade and fixations feature set, Random F orest is the top most p erforming classifier offering the b est trade-off in terms of the cost and the b enefits. The AUC of Random F orest classifier is 0.9111 meaning that it has 91.11% c hance of correctly distinguishing b etw een ADHD and Non-ADHD. 5 Main-Sequence Relationship W e inv estigated the main sequence relationships among saccade amplitude, saccade duration, and saccade p eak velocity for b oth ADHD and non-ADHD groups. F eature sets for the main sequence relationships are based on the following qualifiers: saccade amplitude measured in degrees, saccade duration measured in milliseconds, and saccade p eak velocity measured in degrees p er second. Saccade amplitude is the size of a saccade. Saccade p eak velocity is the highest velocity reached during a saccade. Saccade duration is the time taken to complete the saccade. Saccade p eak v elo cit y is calculated b y the Equation. 1 in degrees/second [27] ˙ θ peak velocity = ˙ θ max × (1 − e − θ amplitude /C ) (1) where ˙ θ peak velocity is the saccade p eak velocity , ˙ θ max is the asymptotic p eak velocity (500 degree/second), θ amplitude is the saccade amplitude (degrees) and C is the constan t (14 for normal h umans). Figure 6(c) sho ws the relationship b et ween saccade amplitude and saccade p eak velocity for normal h umans. Saccade duration in milliseconds is calculated b y Equation. 2 [6]. Figure 7(c) sho ws the relationship b et ween saccade amplitude and saccade duration for normal humans. t duration = (2 . 2 × θ amplitude + 21) (2) 6 Area of In terests In addition, w e emplo yed a num ber of fixation and saccade based features captured within three AOI groups in sen tences presented in RSP AN task. W e utilized a standard RSP AN task where participan ts are instructed to read a sentence and a letter display ed on a computer screen, judge the sen tence’s coherency , and memorize the letter at the end. W e extracted their eye mov ement features based on three stimuli: 1) area of the sen tence, 2) area of the critical word that determines the coherency of the sentence, and 3) the decision area with the letter to b e remem b ered. Figure. 5 shows the three AOIs drawn on a single sentence using T obii Analysis softw are. Note that the b oundaries of A OIs are drawn man ually (static AOIs). W e derived t wo feature sets for the inv estigation of fixations and saccades within AOIs based on the follo wing qualifiers: num b er of fixations in AOI 1, 2 and 3, fixation duration in AOI 1, 2 and 3, av erage fixation duration in AOI 2, fixation standard deviation in A OI 2, pupil diameter of b oth eyes in AOI 2 and 3, maximum and minimum saccade amplitude in AOI 1, 2 and 3, a verage saccade amplitude in A OI 1, 2 and 3, and standard deviation of saccade amplitude in AOI 1, 2 and 3, respectively . • Sc ene-b ase d : F eature set including the ab o ve qualifiers within the A OIs of sets of 2-5 sen tences. 11 Figure 5: A OIs During WMC T ask from a Sen tence generated using T obii Studio Analysis Softw are. • Sentenc e-b ase d : F eature set including the ab o ve qualifiers within the A OIs of all the sentences. All fixation features and saccade features were calculated using P andas, a Python data analysis library . Prior studies [11] suggested that diagnostic criteria for ADHD should b e adjusted to gender differences. W e find that including gender in the feature set sligh tly increases the p erformance across all our classifiers. 7 Analysis of Main Sequence Relationships (a) Amplitude vs. Peak V elo city of ADHD (b) Amplitude vs. Peak V elo city of Non-ADHD (c) Amplitude vs. Peak V elo city of Normal Figure 6: Main Sequence Relationships (a) the relationship b et ween saccade amplitude (degree) and saccade p eak velocity (degrees/second) of ADHD sub jects, (b) the relationship b et ween saccade amplitude (degree) and saccade p eak v elo cit y (degrees/second) of Non-ADHD sub jects, and (c) the relationship b etw een saccade amplitude (degree) and saccade p eak velocity (degrees/second) of Normal humans. In general, saccades are stereotyped: The relationships b etw een saccade amplitude, saccade peak v elo cit y , and saccade duration are relatively fixed for normal human b eings, and are referred to as main sequence relationships. The t wo main sequence relationships are: 1) the relationship b etw een saccade amplitude 12 (a) Amplitude vs. Dur ation of ADHD (b) Amplitude vs. Dur ation of non-ADHD (c) Amplitude vs. Dur ation of Normal Figure 7: Main Sequence Relationships, (a) the relationship betw een saccade amplitude (degree) and duration (ms) of ADHD sub jects, (b) the relationship b et ween saccade amplitude (degree) and duration (ms) of Non- ADHD sub jects, and (c) the relationship b et ween saccade amplitude (degree) and duration (ms) of Normal h umans. (degree) and duration (ms), and 2) the relationship b etw een saccade amplitude (degree) and saccade p eak v elo city (degree/second). W e hypothesize that any differences encountered in main sequence relationships could lead to the conclusion that the saccade is not normal. Figure. 6 presents the relationships betw een saccade amplitude and saccade p eak v elo cit y in represen ta- tiv e ADHD and Non-ADHD adults during the entire session of WMC task. The data in Figures 6(a) and 6(b) show a similar relationship b et ween saccade amplitude and the saccade p eak velocity with trend line as w ell for normal humans (see Figure 6(c)). These results are consistent with the study [17] which describes the main sequence relationship indexing the test of v ariables of atten tion. The data in Figure. 7 show a similar relationship b et ween saccade amplitude and the saccade duration for ADHD and Non-ADHD adults during the en tire WMC task. In addition, it shows similar trend line with normal humans as w ell (see Figures 7(a), 7(b), and 7(c)). 7.1 Mac hine Learning on Scene-based and Sentence-based F eature Sets W e obtained all p erformance metrics using WEKA by executing the selected classifier with a 10-fold cross v alidation using the b oth feature sets we dev elop ed for the in vestigation of fixations and saccades within A OIs. The reason for using WEKA is that, it facilitates users to execute machine learning algorithms out-of-the-b o x and visualize ho w differen t algorithms p erform for the same data set. Figure 1 presen ts images of eye gaze patterns from tw o adults participants, one with and one without ADHD. According to the Figure 1, the adult with ADHD is fixating primarily b elow the AOIs of stimulus items in sentence including: the words, the decision p oint, and the item to b e remem b ered (see Figure 5). The adult without ADHD has a larger num b er of fixations which are in-line with A OIs. W e selected the same six top performing classifiers listed in T able 2 and w e utilize our feature sets whic h primarily consist of fixation and saccade features within AOIs to train the six classifiers. T able 5 lists the classification results of the scene-based feature set. The RandomF orest classifier yielded the highest p ercent accuracy of with 83.33% indicating that the scene-based feature set alone classifies a 13 (a) Amplitude vs. Dur a- tion of ADHD (b) Amplitude vs. Dur a- tion of Non-ADHD (c) Amplitude vs. Pe ak V elo city of ADHD (d) Amplitude vs. Pe ak V elo city of Non-ADHD Figure 8: Main Sequence relationships obtained from the Scene-based F eature set including eye gaze metrics within the A OIs; (a) Saccade Amplitude vs. Saccade Duration relationship of ADHD participants, (b) Saccade Amplitude vs. Saccade Duration relationship of Non-ADHD participan ts, (c) Saccade Amplitude vs. Saccade Peak V elo city relationship of ADHD participants, and (d) Saccade Amplitude vs. Saccade Peak V elo cit y relationship of Non-ADHD participants (a) Amplitude vs. Dur a- tion of ADHD (b) Amplitude vs. Dur a- tion of Non-ADHD (c) Amplitude vs. Pe ak V elo city of ADHD (d) Amplitude vs. Pe ak V elo city of Non-ADHD Figure 9: Main Sequence relationships obtained from the Sen tence-based F eature set including ey e gaze metrics within the AOIs ; (a) Saccade Amplitude vs. Saccade Duration relationship of ADHD participants, (b) Saccade Amplitude vs. Saccade Duration relationship of Non-ADHD participan ts, (c) Saccade Amplitude vs. Saccade Peak V elo city relationship of ADHD participants, and (d) Saccade Amplitude vs. Saccade Peak V elo cit y relationship of Non-ADHD participants diagnosis of ADHD with greater than 80% accuracy . The Kstar classifier yielded the lo west percent accuracy at 63.04% for the scene-based feature set. T able 6 lists the results of the sentence-based feature set. The RandomF orest classifier yielded the highest p ercen t accuracy of with 86.20% indicating that the sen tence-based feature set classifies a diagnosis of ADHD with greater than 85% accuracy . F or sentence-based feature set, K star classifier yielded the lo west p ercen t accuracy at 71.83%. Since any differences encoun tered in main sequence relationships could lead to the conclusion that the saccade is not normal, we plot Figure. 8 and 9 to analyze the main sequence relationships of the tw o feature 14 sets we generated using AOIs. Figure. 8(c) presents the relationship b et ween saccade amplitude and the saccade p eak velocity in representativ e ADHD adults and 8(d) presents the relationship b etw een saccade amplitude and the saccade p eak velocity in representativ e Non-ADHD adults when using the scene-based feature set during the entire WMC task. The data in Figures 8(c) and 8(d) show a similar relationship b et ween saccade amplitude vs. the saccade p eak velocity and similar saccade amplitude range for b oth ADHD and Non-ADHD sub ject groups. The data in Figure 8(a) and 8(b) show a similar relationship b et ween saccade amplitude and the saccade duration for ADHD and Non-ADHD adults when using the scene-based feature set during the entire WMC task. The results indicate when considering AOIs of scence- based sentences, ADHD and Non-ADHD adults display similarities in main sequence relationships which are similar to normal humans as w ell. The data in Figures 9(c) and 9(d) show a similar relationship b et ween saccade amplitude vs. the saccade p eak velocity for b oth ADHD and Non-ADHD sub ject groups when using the sentence-based feature set during the entire WMC task. The data in Figure 9(a) and 9(b) also show a similar relationship b etw een saccade amplitude and the saccade duration for ADHD and Non-ADHD adults. These relationships are similar to the relationships obtained from the scene-based feature set (see Figure 8). 8 Discussion Since participants are presented with v arying sets of 2-5 sentences, we developed one feature set considering the AOIs in the first sentence of all the sen tence sets (scene-based feature set) and the other feature set considering the AOIs of all the 42 sen tences in the RSP AN task (sen tence-based feature set). W e used common AOIs among all the participants, thus they are static shap es and would not change from one sub ject to another. In the case of static A OIs, the granularit y of the sentence-based feature set is increased when compared to the granularit y of the scene-based feature set. As a result, we observe b etter accuracy in eac h classifier(See T able 6). T able 5: Classification of Ey e Fixation and Saccade F eatures within AOIs of Scence-based during WMC Classifier Precision Recall F1 Accuracy J48 0.75 0.75 0.75 75.36 LMT 0.79 0.79 0.79 79.71 RandomF orest 0.83 0.83 0.83 83.33 REPT ree 0.70 0.70 0.69 70.29 K* 0.63 0.63 0.63 63.04 Bagging 0.80 0.79 0.79 79.71 T able 6: Classification of Ey e Fixation and Saccade F eatures within AOIs of Sen tence-based during WMC Classifier Precision Recall F1 Accuracy J48 0.79 0.79 0.79 79.39 LMT 0.82 0.82 0.82 82.61 RandomF orest 0.86 0.86 0.86 86.20 REPT ree 0.82 0.82 0.82 82.04 K* 0.72 0.71 0.71 71.83 Bagging 0.84 0.83 0.83 83.93 Consideration of fixation as well as saccade features set according to stimulus AOIs, lead us to classify a diagnosis of ADHD with greater than 80% accuracy . Our results confirm the utilit y of eye mo vemen t feature set generated according to stim ulus AOIs indexing WMC as a predictor of a diagnosis of ADHD in adults. RandomF orest classifiers p erformed best in-terms of predicting a clas sification of ADHD with 86.20% percent accuracy by using sentence-based feature set representing a physiological measure of visual attention during a WMC task. 15 Since we are utilizing the ey e gaze metrics calculated by T obii Studio analysis softw are within the manually mark ed static AOI b oundaries for the developmen t of our feature sets, there might b e instances where we ha ve less data p oints. The static AOIs may not b e enough in terms of the area b oundary to capture eye gaze metrics of some of the participants. W e did not consider device error or human error when creating the AOI b oundaries. In the future, we are interested in identifying AOIs dynamically for each participan t in each sentence. 9 Conclusions The purp ose of this feasibility study w as to determine if patterns of saccades and stability of fixations generated when completing a measure of WMC, the RSP AN task, would create a feature set which could b e used to differen tiate a diagnosis of ADHD for adults. W e used accuracy , precision, recall, and f-measure as the ev aluation metrics. While fixation features, saccade features, and a combination of saccade and fixtaion features accurately predicted the classification of ADHD with an accuracy of greater than 78%, saccade features were the b est predictors with an accuracy of 91%. The results of this feasibility study confirmed the utilit y of a combination of fixation, and saccade feature set generated within AOIs while completing RSP AN tasks as a predictor of a diagnosis of ADHD in adults. T ree-based classifiers p erformed b est in-terms of predicting a classification of ADHD with 86% p ercent accuracy using physiological measures of sustained visual attention within A OIs during a WMC task. These results are consistent with previous studies confirming significant differences in saccadic b ehaviors for p eople with ADHD [26, 41]. During the RSP AN task, the rapid mov ement of the eye across the scan path from one fixation p oint to the other yields a more accurate class ification of ADHD than the abilit y to sustain gaze. These preliminary results indicate that detailed and discrete ey e gaze metrics during a measure of attention control (i.e. WMC) provide unique indices of ADHD and offer physiological insight regarding cognitiv e resources underlying WMC, an imp ortan t cognitiv e construct resp onsible for b eha vioral inhibition and attention monitoring. Moreov er, they are consisten t with previous inv estigations finding that adults with ADHD demonstrate similar broad visual attention patterns as adults without ADHD but different scan patterns [24] and different pupillometry metrics as a function of visual cue type [33]. 10 F uture Directions The results of this feasibilit y study confirm the utility of eye mov emen t feature set indexing WMC as a predictor of a diagnosis of ADHD in adults. RandomF orest classifiers performed best in-terms of predicting a classification of ADHD with 91.14% p ercen t accuracy by com bining saccade feature set representing a ph ysiological measure of visual attention during a WMC task. T ree-based classifiers performed b est in-terms of predicting a classification of ADHD with 86% p ercent accuracy using ey e gaze metrics within AOIs during a WMC task. This pro ject is a necessary first step in delineating a feature set of eye gaze metrics captured within AOIs which represent physiological diagnostic criteria, including executive attention in adult ADHD. In the future, we will expand the exp erimen tal studies to further analyze eye gaze metrics according to dynamically changing stim ulus AOIs with resp ect to the participants using a larger sample size. Sp ecifically , creating a b oundary for the AOIs; the sentence, the w ord which determine sentence accuracy , the visual p oin t of decision, and the item to be remembered. Identifying these AOIs dynamically for eac h participant will enable us to generate a detailed feature set whic h could b e utilized to classify a diagnosis of ADHD with a greater p ercen tage of accuracy than of this study . 16 References [1] B. Abibullaev and J. An. Decision supp ort algorithm for diagnosis of adhd using electro encephalograms. Journal of me dic al systems , 36(4):2675–2688, 2012. [2] R. M. Alderson, L. J. Kasper, K. L. Hudec, and C. H. Patros. Atten tion-deficit/hyperactivity disorder (adhd) and w orking memory in adults: a meta-analytic review. Neur opsycholo gy , 27(3):287, 2013. [3] R. A. Barkley . Behavioral inhibition, sustained atten tion, and executive functions: constructing a unifying theory of adhd. Psycholo gic al bul letin , 121(1):65, 1997. [4] J. Biederman, P . Hammerness, B. Sadeh, Z. Peremen, A. Amit, H. Or-Ly , Y. Stern, A. Reches, A. Gev a, and S. F araone. Diagnostic utility of brain activity flo w patterns analysis in attention deficit hyperac- tivit y disorder. Psycholo gic al me dicine , 47(7):1259–1270, 2017. [5] R. N. Blazey , D. L. Patton, and P . A. Parks. Adhd detection by eye saccades, Nov. 25 2003. US Paten t 6,652,458. [6] R. H. Carp en ter. Movements of the Eyes, 2nd R ev . Pion Limited, 1988. [7] G. J. Colflesh and A. R. Conw ay . Individual differences in working memory capacity and divided atten tion in dichotic listening. Psychonomic Bul letin & R eview , 14(4):699–703, 2007. [8] A. Con wa y , C. Jarrold, and M. Kane. V ariation in working memory . Oxford Univ ersity Press, 2008. [9] A. R. Conw a y , M. J. Kane, M. F. Bunting, D. Z. Hambric k, O. Wilhelm, and R. W. Engle. W ork- ing memory span tasks: A metho dological review and users guide. Psychonomic bul letin & r eview , 12(5):769–786, 2005. [10] M. Daneman and P . A. Carp en ter. Individual differences in integrating information b etw een and within sen tences. Journal of Exp erimental Psycholo gy: L e arning, Memory, and Co gnition , 9(4):561, 1983. [11] S. Diamantopoulou, L. Henricsson, and A.-M. Rydell. Adhd symptoms and p eer relations of c hildren in a communit y sample: examining asso ciated problems, self-p erceptions, and gender differences. In- ternational Journal of Behavior al Development , 29(5):388–398, 2005. [12] R. W. Engle. W orking memory capacity as executive attention. Curr ent dir e ctions in psycholo gic al scienc e , 11(1):19–23, 2002. [13] R. W. Engle, S. W. T uholski, J. E. Laughlin, and A. R. Conw ay . W orking memory , short-term memory , and general fluid in telligence: a latent-v ariable approach. Journal of exp erimental psycholo gy: Gener al , 128(3):309, 1999. [14] S. A. Fields, W. M. Johnson, and M. B. Hassig. Adult adhd: Addressing a unique set of challenges. The journal of family pr actic e , 66(2):68–74, 2017. [15] K. N. F ong, W. Ma, H. Pang, P . P . T ang, and L. L. Law. Immediate effects of coloured ov erla ys on the reading p erformance of preschool children with an autism sp ectrum disorder using eye tracking. R ese ar ch in developmental disabilities , 89:141–148, 2019. [16] L. F ostic k. The effect of attention-deficit/h yp eractivit y disorder and meth ylphenidate treatment on the adult auditory temporal order judgment threshold. Journal of Sp e e ch, L anguage, and He aring R ese ar ch , 60(7):2124–2128, 2017. [17] M. F ried, E. Tsitsiash vili, Y. S. Bonneh, A. Sterkin, T. Wygnanski-Jaffe, T. Epstein, and U. P olat. Adhd sub jects fail to suppress eye blinks and microsaccades while anticipating visual stimuli but recov er with medication. Vision r ese ar ch , 101:62–72, 2014. [18] L. L. Greenhill. Diagnosing attention-deficit/h yp eractivity disorder in children. The Journal of clinic al psychiatry , 59:31–41, 1998. 17 [19] H. Hart, K. Chantiluk e, A. I. Cubillo, A. B. Smith, A. Simmons, M. J. Brammer, A. F. Marquand, and K. Rubia. Pattern classification of resp onse inhibition in adhd: tow ard the developmen t of neurobio- logical markers for adhd. Human Br ain Mapping , 35(7):3083–3094, 2014. [20] D. Kaiser, G. A. Jacob, L. v an Zutphen, N. Siep, A. Sprenger, B. T usc hen-Caffier, A. Senft, A. Arntz, and G. Domes. Biased attention to facial expressions of am biguous emotions in b orderline personality disorder: An ey e-tracking study . Journal of p ersonality disor ders , pages 1–28, 2019. [21] M. J. Kane, M. K. Blec kley , A. R. Conw a y , and R. W. Engle. A con trolled-attention view of working- memory capacity . Journal of Exp erimental Psycholo gy: Gener al , 130(2):169, 2001. [22] O. Komogortsev, C. Holland, S. Ja yarathna, and A. Karp ov. 2d linear o culomotor plant mathematical mo del: V erification and biometric applications. ACM T r ansactions on Applie d Per c eption (T AP) , 10(4):27, 2013. [23] O. V. Komogortsev, S. Jay arathna, C. R. Aragon, and M. Mahmoud. Biometric identification via an o culomotor plant mathematical model. In Pr o c e e dings of the 2010 Symp osium on Eye-T r acking R ese ar ch & Applic ations , pages 57–60. ACM, 2010. [24] K. Krejtz, A. Duc howski, T. Szmidt, I. Krejtz, F. Gonz´ alez P erilli, A. Pires, A. Vilaro, and N. Villalob os. Gaze transition en tropy . ACM T r ansactions on Applie d Per c eption (T AP) , 13(1):4, 2015. [25] K. Krejtz, A. T. Ducho wski, A. Niedzielsk a, C. Biele, and I. Krejtz. Ey e tracking cognitive load using pupil diameter and microsaccades with fixed gaze. PloS one , 13(9):e0203629, 2018. [26] Y.-J. Lee, S. Lee, M. Chang, and H.-W. Kwak. Saccadic mov emen t deficiencies in adults with adhd tendencies. ADHD Attention Deficit and Hyp er activity Disor ders , 7(4):271–280, 2015. [27] R. Leigh and D. Zee. The neurology of ey e mo vemen ts: Oxford univ ersity press new york. 1999. [28] R. J. Leigh and C. Kennard. Using saccades as a research to ol in the clinical neurosciences. Br ain , 127(3):460–477, 2004. [29] R. J. Leigh and D. S. Zee. The neur olo gy of eye movements . Oxford Univ ersity Press, USA, 2015. [30] L. Manevitz and M. Y ousef. One-class do cument classification via neural netw orks. Neur o c omputing , 70(7-9):1466–1481, 2007. [31] J. L. L. Marcano, M. A. Bell, and A. L. Beex. Classification of adhd and non-adhd using ar mo d- els. In Engine ering in Me dicine and Biolo gy So ciety (EMBC), 2016 IEEE 38th A nnual International Confer enc e of the , pages 363–366. IEEE, 2016. [32] A. M. Michalek, S. M. W atson, I. Ash, S. Ringleb, and A. Raymer. Effects of noise and audiovisual cues on sp eec h processing in adults with and without adhd. International journal of audiolo gy , 53(3):145–152, 2014. [33] A. P . Michalek and J. Ro c he. Pupil dilation as a measure of atten tion in adhd: A pilot study . Americ a , 116:2395–2405, 2017. [34] N. Mohammadhasani, R. A. F abio, H. F ardanesh, and J. Hatami. The link b etw een visual atten tion and memory in adhd and normally developing studen ts: Seeing is remembering? R eti, sap eri, linguaggi , 2(1):89–102, 2015. [35] S. H. Mostofsky , A. Lasker, L. Cutting, M. Denckla, and D. Zee. Oculomotor abnormalities in attention deficit hyperactivity disorder a preliminary study . Neur olo gy , 57(3):423–430, 2001. [36] A. Mueller, G. Candrian, J. D. Krop otov, V. A. P onomarev, and G.-M. Baschera. Classification of adhd patien ts on the basis of indep endent erp comp onen ts using a machine learning system. In Nonline ar biome dic al physics , volume 4, page S1. BioMed Central, 2010. 18 [37] D. P . Munoz, I. T. Armstrong, K. A. Hampton, and K. D. Mo ore. Altered control of visual fixation and saccadic eye mov ements in attention-deficit hyperactivity disorder. Journal of neur ophysiolo gy , 90(1):503–514, 2003. [38] E. National Academies of Sciences, Medicine, et al. Impr oving diagnosis in he alth c ar e . National Academies Press, 2016. [39] X. Peng, P . Lin, T. Zhang, and J. W ang. Extreme learning machine-based classification of adhd using brain structural mri data. PloS one , 8(11):e79476, 2013. [40] J. Richmond and C. A. Nelson. Relational memory during infancy: evidence from eye tracking. Devel- opmental scienc e , 12(4):549–556, 2009. [41] W. Roberts, M. T. Fillmore, and R. Milich. Separating automatic and in tentional inhibitory mec hanisms of attention in adults with attention-deficit/h yp eractivit y disorder. Journal of Abnormal Psycholo gy , 120(1):223, 2011. [42] N. Rommelse, S. V an der Stigc hel, J. Witlo x, C. Geldof, J.-B. Deijen, J. Theeuw es, J. Oosterlaan, and J. Sergeant. Deficits in visuo-spatial working memory , inhibition and o culomotor con trol in b o ys with adhd and their non-affected brothers. Journal of Neur al T r ansmission , 115(2):249–260, 2008. [43] J. R¨ onn b erg, T. Lunner, A. Zekv eld, P . S¨ orqvist, H. Danielsson, B. Lyxell, ¨ O. Dahlstr¨ om, C. Signoret, S. Stenfelt, M. K. Pic hora-F uller, et al. The ease of language understanding (elu) mo del: theoretical, empirical, and clinical adv ances. F r ontiers in systems neur oscienc e , 7:31, 2013. [44] K. Rubia. Cognitive neuroscience of attention deficit hyperactivity disorder (adhd) and its clinical translation. F r ontiers in human neur oscienc e , 12:100, 2018. [45] V. J. Serrano, J. S. Ow ens, and B. Hallow ell. Where c hildren with adhd direct visual atten tion during emotion knowledge tasks: relationships to accuracy , resp onse time, and adhd symptoms. Journal of attention disor ders , 22(8):752–763, 2018. [46] P .-H. Tseng, I. G. Cameron, G. P ari, J. N. Reynolds, D. P . Munoz, and L. Itti. High-throughput classifi- cation of clinical p opulations from natural viewing eye mo vemen ts. Journal of neur olo gy , 260(1):275–284, 2013. [47] S. V an der Stigchel, N. Rommelse, J. Deijen, C. Geldof, J. Witlox, J. Oosterlaan, J. Sergeant, and J. Theeuw es. Oculomotor capture in adhd. Co gnitive Neur opsycholo gy , 24(5):535–549, 2007. [48] E. G. Willcutt. The prev alence of dsm-iv attention-deficit/h yp eractivit y disorder: a meta-analytic review. Neur other ap eutics , 9(3):490–499, 2012. [49] G. Xu, L. Strathearn, B. Liu, B. Y ang, and W. Bao. Twen ty-y ear trends in diagnosed attention- deficit/h yp eractivity disorder among us children and adolescents, 1997-2016. volume 1, pages e181471– e181471. American Medical Asso ciation, 2018. 19