ORIGINAL RESEARCH published: 04 February 2021 doi: 10.3389/fnagi.2020.603790 Distinctive Oculomotor Behaviors in Alzheimer’s Disease and Frontotemporal Dementia Carmen Lage 1,2*, Sara López-García 1 , Alexandre Bejanin 2,3 , Martha Kazimierczak 1,2 , Ignacio Aracil-Bolaños 3 , Alberto Calvo-Córdoba 4 , Ana Pozueta 1,2 , María García-Martínez 1,2 , Andrea Fernández-Rodríguez 1 , María Bravo-González 1 , Julio Jiménez-Bonilla 5 , Ignacio Banzo 5 , Juan Irure-Ventura 6 , Jordi Pegueroles 2,3 , Ignacio Illán-Gala 2,3 , Juan Fortea 2,3 , Eloy Rodríguez-Rodríguez 1,2 , Alberto Lleó-Bisa 2,3 , Cecilia E. García-Cena 4 and Pascual Sánchez-Juan 1,2 1 Institute for Research Marqués de Valdecilla (IDIVAL), University of Cantabria and Department of Neurology, Marqués de Valdecilla University Hospital, Santander, Spain, 2 Centro de Investigación Biomédica en Red sobre Enfermedades Neurodegenerativas (CIBERNED), Madrid, Spain, 3 Sant Pau Memory Unit, Department of Neurology, Hospital de la Santa Creu i Sant Pau - Biomedical Research Institute Sant Pau (IIB Sant Pau), Universitat Autonoma de Barcelona, Barcelona, Spain, 4 Escuela Técnica Superior de Ingeniería y Diseño Industrial – Centre for Automation and Robotics, Technical Edited by: University of Madrid (UPM) – Consejo Superior de Investigaciones Científicas and Aura Innovative Robotics SL, Madrid, Ana B. Vivas, Spain, 5 Department of Nuclear Medicine, Marqués de Valdecilla University Hospital, Santander, Spain, 6 Department of International Faculty of the University Immunology, Marqués de Valdecilla University Hospital, Santander, Spain of Sheffield, Greece Reviewed by: Sebastian Crutch, Oculomotor behavior can provide insight into the integrity of widespread cortical University College London, networks, which may contribute to the differential diagnosis between Alzheimer’s United Kingdom Manousos A. Klados, disease and frontotemporal dementia. Three groups of patients with Alzheimer’s disease, International Faculty of the University behavioral variant of frontotemporal dementia (bvFTD) and semantic variant of primary of Sheffield, Greece progressive aphasia (svPPA) and a sample of cognitively unimpaired elders underwent an *Correspondence: eye-tracking evaluation. All participants in the discovery sample, including controls, had Carmen Lage clage@idival.org a biomarker-supported diagnosis. Oculomotor correlates of neuropsychology and brain metabolism evaluated with 18F-FDG PET were explored. Machine-learning classification Received: 07 September 2020 algorithms were trained for the differentiation between Alzheimer’s disease, bvFTD and Accepted: 28 December 2020 Published: 04 February 2021 controls. A total of 93 subjects (33 Alzheimer’s disease, 24 bvFTD, seven svPPA, and Citation: 29 controls) were included in the study. Alzheimer’s disease was the most impaired Lage C, López-García S, Bejanin A, group in all tests and displayed specific abnormalities in some visually-guided saccade Kazimierczak M, Aracil-Bolaños I, parameters, as pursuit error and horizontal prosaccade latency, which are theoretically Calvo-Córdoba A, Pozueta A, García-Martínez M, closely linked to posterior brain regions. BvFTD patients showed deficits especially in the Fernández-Rodríguez A, most cognitively demanding tasks, the antisaccade and memory saccade tests, which Bravo-González M, Jiménez-Bonilla J, Banzo I, Irure-Ventura J, Pegueroles J, require a fine control from frontal lobe regions. SvPPA patients performed similarly to Illán-Gala I, Fortea J, controls in most parameters except for a lower number of correct memory saccades. Rodríguez-Rodríguez E, Lleó-Bisa A, Pursuit error was significantly correlated with cognitive measures of constructional praxis García-Cena CE and Sánchez-Juan P (2021) Distinctive Oculomotor and executive function and metabolism in right posterior middle temporal gyrus. The Behaviors in Alzheimer’s Disease and classification algorithms yielded an area under the curve of 97.5% for the differentiation Frontotemporal Dementia. Front. Aging Neurosci. 12:603790. of Alzheimer’s disease vs. controls, 96.7% for bvFTD vs. controls, and 92.5% for doi: 10.3389/fnagi.2020.603790 Alzheimer’s disease vs. bvFTD. In conclusion, patients with Alzheimer’s disease, bvFTD Frontiers in Aging Neuroscience | www.frontiersin.org 1 February 2021 | Volume 12 | Article 603790 Lage et al. Oculomotor Behavior in Dementia and svPPA exhibit differentiating oculomotor patterns which reflect the characteristic neuroanatomical distribution of pathology of each disease, and therefore its assessment can be useful in their diagnostic work-up. Machine learning approaches can facilitate the applicability of eye-tracking in clinical practice. Keywords: oculomotor, Alzheimer’s disease, frontotemporal dementia, biomarkers, antisaccade, smooth pursuit, semantic dementia INTRODUCTION be characterized through diverse features. Although some of them can be described qualitatively through One might think that through the assessment of eye movements clinical examination, modern video-oculography we are just evaluating purely motor responses. Far from this, technology allows to reliably quantify them and to obtain oculomotor behavior is controlled by a widespread and highly additional parameters based on raw data. Thus, this interconnected cortical network that works in a hierarchical new technology emerges as a potential tool to measure manner integrating perception, action planning and response the integrity of the brain regions that support all these generation processes [for a review, see (McDowell et al., oculomotor features. 2008; Coiner et al., 2019)]. The simplest ocular movements Different oculomotor changes have been described in are visually-guided responses to keep an object of interest degenerative dementias. In Alzheimer’s disease, the most fixed in the fovea, either a rapid gaze shift to an appearing characteristic findings are an increment in saccade latencies stimulus or prosaccade, either a pursuit movement which and in the antisaccade error rate (Molitor et al., 2015; allows tracking a moving target. Such visually-guided ocular Kahana Levy et al., 2018). Studies including frontotemporal movements are controlled by a core frontoparietal network dementia (FTD) patients are scarcer and some of their that includes the frontal and supplementary eye fields and results conflicting, which could be due to methodological diverse parietal regions like the intraparietal sulcus and superior differences but also to sample heterogeneity between studies, parietal cortex, and which is supported by subcortical structures since, apart from one work which described patients with and finally connected to the superior colliculus, the saccade autopsy-confirmed diagnosis (Boxer et al., 2012), no others have generator. In contrast, volitional saccades are ocular responses exclusively included patients with either pathology or biomarker- according to contextual commands, as occurs in the antisaccade supported diagnoses. or memory saccade paradigms, which require the recruitment We hypothesized that oculomotor parameters that are of additional cortical regions to carry out and control the related to visuospatial functioning and therefore predominantly execution of an internally-generated plan. In the antisaccade controlled by posterior brain regions would be significantly paradigm, the subject is instructed to look at the mirror position impaired in Alzheimer’s disease patients, where the hallmark of the appearing stimulus. This test has been considered a is parietal and posterior temporal atrophy. On the other sensitive measure of inhibitory control, since it requires to hand, we would expect that, in patients with the behavioral withhold the automatic response of looking toward the stimulus variant of FTD (bvFTD), the most impaired parameters and, instead, to generate a saccade in the opposite direction would be those related to executive functions and which (Heuer et al., 2013). Antisaccade generation is controlled by involve supporting activity from frontal lobe regions; while the same basic frontoparietal network with an additional role overall performance would be preserved in the semantic of the dorsolateral prefrontal cortex (DLPFC) and anterior variant of primary progressive aphasia (svPPA), where cingulate cortex (Munoz and Everling, 2004; Pa et al., 2014). neurodegeneration tends to affect selectively anterior temporal Similarly, the memory saccade paradigm is another cognitively lobes and cognitive functions others than semantic memory are demanding task in which spatial working memory is necessary initially less affected. to retain the precise location where the stimulus has previously Our first aim was to describe the oculomotor behavior appeared and to redirect the gaze to it based on an internal in response to a wide range of tests in a well-phenotyped representation. In this case, neuroimaging and lesion studies sample of patients with Alzheimer’s disease, bvFTD point to a recruitment activation of DLPFC together with and svPPA diagnoses, supported by neuropsychological, basal ganglia and thalamocortical circuitries (Brown et al., neuroimaging and Alzheimer’s disease core biomarkers, 2004). and to compare it with a sample of cognitively unimpaired When assessing visually-guided and volitional controls with negative Alzheimer’s disease biomarkers. Our saccades in clinical practice, oculomotor responses can second aim was to analyze whether oculomotor parameters were correlated with the results of neuropsychological Abbreviations: AUC, area under the curve; bvFTD, behavioral variant assessments and brain metabolism evaluated with 18F- frontotemporal dementia; DLPFC, dorsolateral prefrontal cortex; 18F-FDG, 2- FDG PET to explore if they reflected specific patterns of [18F] Fluoro-2-Deoxy-D-Glucose; MMSE, Mini-Mental State Examination; PiB, Pittsburgh Compound-B; ROC, receiver operating characteristic; ROCFT, Rey- brain dysfunction. Finally, we tested whether a machine Osterrieth Complex Figure Test; TMT, Trail Making Test; svPPA, semantic variant learning classification algorithm could contribute to their of primary progressive aphasia. differential diagnosis. Frontiers in Aging Neuroscience | www.frontiersin.org 2 February 2021 | Volume 12 | Article 603790 Lage et al. Oculomotor Behavior in Dementia MATERIALS AND METHODS (Hernández et al., 2018). Stimuli were bright green dots with a diameter of two centimeters presented on a display at a Participants viewing distance of 60 centimeters. An examination protocol Three groups of patients with Alzheimer’s disease, bvFTD and establishing the sequence of tests and standardized instructions svPPA were recruited from the Cognitive Disorders Unit of the for participants was followed to minimize variability as much Marqués de Valdecilla University Hospital (Santander, Spain). as possible. Definite trials were preceded by practice trials that Participants were evaluated with the Global Deterioration Scale allowed confirming the understanding of the tests. Each definite (Reisberg et al., 1982) and Mini-Mental State Examination test was preceded by a nine point-calibration and began with a (MMSE) (Folstein et al., 1975) as global measures of disease central fixation target. The evaluation comprised a prosaccade severity. Only patients in a mild dementia stage (Global test, an antisaccade test, a memory saccade test and a sinusoidal Deterioration Scale = 4) were included. All Alzheimer’s disease smooth pursuit test. The prosaccade, antisaccade and memory patients displayed the classical amnestic phenotype. Diagnoses saccade tests included 12 trials in the horizontal plane followed by were established according to consensus criteria for probable eight trials in the vertical plane each. Horizontal trials consisted Alzheimer’s disease (McKhann et al., 2011), bvFTD (Rascovsky of the random appearance of targets at 5, 10, or 20 degrees to right et al., 2011), and svPPA (Gorno-Tempini et al., 2011). In or left; and, in vertical trials, at 5 or 12 degrees up or down. The addition to congruent neuropsychological and neuroimaging sinusoidal smooth pursuit test included six horizontal trials and findings (brain CT and/or MRI), all diagnoses were supported six vertical trials. by at least one type of biomarker, amyloid-PET, and/or CSF Alzheimer’s disease core biomarkers. Final diagnoses were Prosaccade Test agreed in multidisciplinary meetings including four neurologists Prosaccades were evaluated by the random appearance of an (PSJ, ERR, SLG, and CL) and two neuropsychologists (MGM eccentric target, subsequently replaced by the reappearance of the and AP). To minimize misclassification or heterogeneity central target. Subjects were instructed to keep their gaze fixed on due to co-pathology, only those cases with a complete the target. concordance between clinicians’ diagnosis and biomarker results were included. Sinusoidal Smooth Pursuit Test Healthy volunteers were participants from the Valdecilla Here, the target moved from one end of the screen to the other Study of Memory and Brain Aging, a prospective cohort and subjects were asked to follow it as accurately as possible. recruiting community-dwelling non-demented people older than 55 years. The baseline protocol includes a comprehensive Antisaccade Test neuropsychological assessment, brain MRI and CSF analysis of In a similar fashion to that in the prosaccade test, the central Alzheimer’s disease core biomarkers. Those subjects selected target was replaced by the appearance of an eccentric target, but for our study had no cognitive complaints and showed normal the command, in this case, was: “When the target appears at one results in all baseline evaluations, including normal levels of CSF side, look at the opposite location, in a mirrored way. If you biomarkers, which allows excluding the influence of preclinical realize that you have looked at the target, try to correct yourself Alzheimer’s disease on their oculomotor performance. looking at the opposite location.” These four groups were used as a discovery sample for the description of oculomotor performance, exploring Memory Saccade Test As in the prosaccade test, the target appeared eccentrically and neuropsychological and brain metabolic correlates and training then at the central position. After that, the target disappeared, machine learning algorithms. The classification accuracy of these leaving the screen blank. The instructions were: “Keep your gaze algorithms was subsequently tested in two independent samples, fixed on the target when it appears at one side and when it comes a sample of 15 patients with Alzheimer’s disease from our center back to the center. When the central target disappears, look at the who did not have a biomarker-supported diagnosis; and a sample location where it had previously appeared.” of 6 bvFTD patients from an independent center, the Memory Unit of Santa Creu i Sant Pau Hospital (Barcelona, Spain) Oculomotor Parameters (Alcolea et al., 2019a; Illán-Gala et al., 2019). Both samples were In order to use the subtle alterations of eye movements for evaluated with the same eye-tracker and an identical examination diagnostic aims, it is crucial to guarantee the reproducibility protocol as the discovery sample. of the measuring, which is described in the OSCANN medical The study was approved by both local Ethics Committees and device user manual and summarized in a related publication all participants gave their written informed consent according to (Hernández et al., 2018). After the automatic analysis of the the Declaration of Helsinki. For those patients who could not images captured by the eye-tracker camera, we extracted features give a reliable informed consent due to their degree of cognitive from each oculomotor test following the published methodology impairment, it was obtained from their accompanying relative. (García Cena et al., 2020). Oculomotor responses can be characterized by diverse Oculomotor Evaluation: Procedure and parameters. For descriptive purposes, such parameters have Paradigms been grouped into three domains: (a) parameters related to Eye movement recordings were carried out with OSCANN, spatial accuracy, as saccade error (the deviation of the final an eye-tracking device based on video-oculography technology position of the gaze from the target, measured as positive or Frontiers in Aging Neuroscience | www.frontiersin.org 3 February 2021 | Volume 12 | Article 603790 Lage et al. Oculomotor Behavior in Dementia negative error) and pursuit error (the difference between the Mapping version 12 (SPM12) software (Well-come Department target position and the gaze position during a pursuit test); (b) of Imaging Neuroscience, Institute of Neurology, London, UK) parameters related to time, as latency (defined by the time delay implemented in MATLAB 9.2 (The MathWorks, Sherborn, MA). between the appearance of a peripheral target and the onset Images were quantitatively normalized using the pons-vermis of the ocular movement) and pursuit gain (the rate between as the reference region, spatially normalized to the Montreal ocular velocity and target velocity during a pursuit test); and Neurological Institute PET template, and smoothed with a (c) parameters related to success, as the percentage of correct Gaussian kernel of full width at half maximum 8 mm. All memory saccades in the memory saccade test, and, in the resulting images were visually inspected to check for possible antisaccade test, the percentage of correct antisaccades, corrected registration errors. erroneous antisaccades (henceforth, corrected antisaccades) and successful antisaccades, which represent the sum of correct and Statistical Analysis corrected antisaccades. Precise definitions can be found in the Oculomotor parameters were compared across the four Supplementary Material. diagnostic groups of the discovery sample by ANOVA with Tukey’s post hoc test. Multivariate analyses using General Linear Neuropsychological Evaluation Models with the oculomotor parameter as the dependent variable We hypothesized that oculomotor parameters would be and age and sex as covariates were performed. The influence correlated with those cognitive domains that are directly of neurodepressant drugs on oculomotor performance was implicated in the performance of each oculomotor response investigated as a potential confounding variable. Spearman’s or, indirectly, which are theoretically controlled by the test was used to look for significant correlations between each same brain region that supports that oculomotor feature, oculomotor parameter and the number of sedatives taken by including visuospatial function, memory and executive function. a patient, which included antidepressants, benzodiazepines, Therefore, we designed a neuropsychological evaluation that neuroleptics or antiepileptics. In those cases where the level assessed verbal memory with the Free and Cued Selective of significance was <0.1, the number of drugs was included Reminding Test (Pena-Casanova et al., 2009a), visual memory in the multivariate analysis. Differences between groups were with the Rey-Osterrieth Complex Figure Test (ROCFT) Free considered significant when P < 0.05. Analyses were performed Delayed Recall (Pena-Casanova et al., 2009a), constructional using The Statistical Packages for Social Sciences (SPSS 19.0.1). praxis with ROCFT Copy (Pena-Casanova et al., 2009a), The most significant oculomotor parameter from each test was ideomotor apraxia with imitation of finger gestures (Pena- selected to explore its cognitive and brain metabolic correlates. Casanova, 2005), visuospatial ability with the Number Location Neuropsychological scores were compared with these parameters subtest of the Visual Object and Space Perception Battery (Pena- with a Spearman’s correlation and linear regression analyses Casanova et al., 2009b) and attention and executive function with age as a covariate. Multiple regressions were performed with the Trail Making Test (TMT) A and B and Symbol digit test voxel-wise to assess the relationship between 18F-FDG standard (Pena-Casanova et al., 2009c). uptake value rate and the selected oculomotor parameters across the whole gray matter. In these models, we regressed out the Biomarker Studies time delay between the PET acquisition and the oculomotor For the CSF study, levels of amyloid-β (Aβ1-42), total evaluation. A threshold of P < 0.001 uncorrected, together with tau and phosphorylated tau (p-tau181) were quantified with a cluster extent k>100 mm3, was used for all these analyses. the LUMIPULSE G600II automated platform (Fujirebio) and interpreted according to established cut-off points (Alcolea et al., Machine Learning Classification 2019b). Algorithms A subset of patients underwent a neuroimaging study Our third aim was to construct three classification algorithms including a 2-[18F] Fluoro-2-Deoxy-D-Glucose (18F-FDG) PET using a machine learning approach for the differentiation and/or an amyloid-PET with Pittsburgh Compound-B (PiB)/CT between (1) Alzheimer’s disease patients and controls; (2) scan, obtained within 1-week interval using a Siemens Biograph bvFTD patients and controls; (3) Alzheimer’s disease and LSO Pico 3D equipment (Siemens Healthcare Molecular bvFTD patients. Due to its small size, the svPPA sample Imaging, Hoffman Estates, Illinois, USA). Participants were was not considered suitable for this analysis. Only the data injected with 3-4 MBq/kg 18F-FDG and 555 MBq of 11C-PiB. from the discovery sample was used for generating the Image acquisition consisted of one static image acquired from 30 classification algorithms. to 45 min after injection for 18F-FDG PET and from 60 to 90 min The data set used in any machine learning classifier must for 11C-PiB PET. The information provided by the CT was be carefully prepared. Firstly, a normalization was performed used to correct the attenuation and images were reconstructed to avoid dispersion in data with different dimensions. The on a 128 × 128 matrix using the ordered subsets expectation normalization allows the different dimensions of the data to be maximization iterative method. scaled to standardize the range of the characteristics, since it can 11C-PiB PET was evaluated exclusively by visual read affect the results in a critical way (Graf and Borer, 2001). as positive or negative cortical amyloid-β deposition. To investigate the metabolic correlates of oculomotor responses, x−x we preprocessed the 18F-FDG data using Statistical Parametric x̂ = σ Frontiers in Aging Neuroscience | www.frontiersin.org 4 February 2021 | Volume 12 | Article 603790 Lage et al. Oculomotor Behavior in Dementia Where x is the value of a feature, while x and σ are the mean value and the standard deviation of the feature set, respectively. In our study, the number of subjects was less than the number of features, so the partial less square regression (PLSR or PLS) technique was applied to reduce the number of significance variables. Let’s X ∈ Rn the set of independent features and Y ∈ Rn the set of dependent features. The relation between each set is given by a score vector. We compute the score vector using the partial minimum square regression (PLSC) (Krishnan et al., 2011). Then, the feature sets are defined by: X = TPT + E Y = UQT + F Where T ∈ Rn×p , U ∈ Rn×p while P ∈ RN×p and Q ∈ RM×p are the weight matrix and E ∈ Rn×N F ∈ Rn×M are the residual matrices. The PLS2 algorithm was used to compute each matrix. Finally, the Fisher discriminant ratio (FDR) was used to select the features to train the machine learning algorithm. One of the main advantages of FDR is that we can associate the sets of features with a label such as: Alzheimer’s disease, bvFTD or controls. The FDR is defined by the µi : and σi are the mean and FIGURE 1 | Flow chart to find the most suitable combination between eye variance of the set i. movement features and machine learning algorithm. AUC, area under the curve; FDR, Fisher discriminant ratio; ML, machine learning; PLSR, partial less (µ1 − µ2 )2 square regression. FDR = σ12 − σ22  Under Matlab environment, we performed multiple tests in order to find the most suitable combination between oculomotor features and type of machine learning algorithm for each pair of groups (Alzheimer’s disease patients vs. controls; bvFTD vs. controls; Alzheimer’s disease vs. bvFTD), as it is shown in Figure 1. The most suitable combination was defined by the biggest area under the curve (AUC) in the receiver operating characteristic (ROC) curve. The classifiers were modeled through a cross-validation process in which the sample was divided in four subsets (80% of the sample) for training and one subset (20% of the sample) for classification. Despite the cross-validation performed by Matlab, we decided to implement the selected algorithm under Microsoft Visual Studio C++ language in order to get a software independent from Matlab and to carry out the cross-validation of the algorithm with the whole set of samples instead of a subset like in the previous case and through a loop of 1,000 iterations. Moreover, under C++ environment, the confidence interval of the algorithm was computed using the following flow chart FIGURE 2 | Loop implemented to test the confidence interval of the selected machine learning algorithm. AUC, area under the curve; I, iterations; ML, displayed in Figure 2. machine learning; ROC, receiver operating characteristic. Finally, we were interested in applying the classification algorithms for the differentiation between controls and each dementia group (Alzheimer’s disease patients vs. controls and bvFTD vs. controls) in independent samples of patients with the RESULTS aim of testing their external validity. For doing so, the previously generated algorithms were applied to the independent samples of Demographics Alzheimer’s disease and bvFTD patients and their classification From the initial sample of evaluated subjects, five patients accuracy was assessed using ROC curves. were excluded due to a biomarker result discordant with their Frontiers in Aging Neuroscience | www.frontiersin.org 5 February 2021 | Volume 12 | Article 603790 Lage et al. Oculomotor Behavior in Dementia clinical group: one subject with a clinical diagnosis of probable 0.00016) and vertical prosaccades (mean difference 50.34 ms, Alzheimer’s disease dementia due to normal levels of CSF P = 0.014) as well as in return saccades (mean difference biomarkers and negative PiB-PET; and two bvFTD patients and for horizontal prosaccades 110.48 ms, P = 0.0042; and for two svPPA patients due to positivity in PiB-PET. After that, the vertical prosaccades 62.82 ms, P = 0.016). Additionally, svPPA discovery population included 18 Alzheimer’s disease patients, 18 patients displayed significantly longer latencies than controls in bvFTD, seven svPPA, and 29 controls. Diagnoses were supported vertical prosaccades and return saccades. Conversely, prosaccade by at least one type of Alzheimer’s disease core biomarker (11C- latencies in bvFTD patients were comparable to controls, which PiB PET in 40/43 patients; CSF study in 15/43 patients and 29/29 distinguished them from the Alzheimer’s disease group in the controls). Groups did not differ in terms of age at evaluation horizontal test (mean difference 143.37 ms, P = 0.011). (P = 0.36) or disease duration (P = 0.65) (Table 1). Significant Alzheimer’s disease patients also showed the longest latencies differences were observed in the MMSE scoring, with the lowest for corrected antisaccades and memory saccades, in this case performance in the Alzheimer’s disease group, as well as in the followed by bvFTD, which could differentiate both groups from intake of potentially sedative drugs, with a higher intake in controls, in the horizontal as well as in the vertical plane bvFTD patients compared to the other three groups. (Tables 3, 4). In these tests, svPPA latencies were similar to controls and the longer latencies in Alzheimer’s disease patients Oculomotor Evaluation made it possible to differentiate the two dementia groups in both Those oculomotor parameters significantly correlated with the corrected antisaccades and memory saccades. number of sedatives taken by a patient were: (a) in the antisaccade In the smooth pursuit test (Table 2), bvFTD patients presented test: correct, corrected and successful antisaccades, horizontal the lowest vertical gain values, showing significant differences and vertical corrected antisaccade latency, horizontal and vertical with controls (mean difference 0.19, P = 0.031) and Alzheimer’s positive error, and vertical negative error; (b) in the memory disease patients (mean difference 0.27, P = 0.034), while there saccade test: correct memory saccades, horizontal and vertical were no clear differences in the horizontal pursuit. latency and horizontal and vertical negative error. For the analysis of these parameters, the number of drugs was included in the multivariate analysis as a covariate. Parameters Related to Success For these parameters, the performance between horizontal Parameters Related to Spatial Accuracy and vertical trials was compared. There were no statistically The poorest accuracy performance was obtained in the significant differences in the percentage of correct, corrected Alzheimer’s disease group, who showed a tendency to make and successful antisaccades nor correct memory saccades, so hypometric saccades with greater values of negative error in their horizontal and vertical saccades were pooled to obtain total prosaccades, antisaccades and memory saccades than the other scores for each parameter to increase statistical power and groups, which resulted significantly different from controls in simplify results. In previous parameters related to time and these three tests (Tables 2–4; Figure 3). For the most cognitively accuracy, we considered that keeping horizontal and vertical demanding tasks, the antisaccade and memory saccade tests, the results disaggregated was more appropriate since the amplitudes bvFTD group also made less accurate saccades than controls, with to the target in horizontal and vertical trials were not the same significantly greater values of positive and negative error in the and this might impact values related to time or distance. antisaccade test and negative error in the memory saccade test. Again, the Alzheimer’s disease group showed the poorest SvPPA patients did not show significant differences from performance, with significantly lower percentages of corrected controls in error values except for a greater positive error in and successful antisaccades and correct memory saccades horizontal antisaccades (mean difference 3.25 degrees, P = than the other three groups (Tables 3, 4). Although not as 0.020). Their lower negative error values also distinguished them low as Alzheimer’s disease patients, the bvFTD group also from Alzheimer’s disease and bvFTD both in the antisaccade and obtained significantly lower percentages of correct, corrected, the horizontal memory saccade test. and successful antisaccades as well as correct memory saccades In the smooth pursuit test, the greatest pursuit error values than controls. There were no significant differences between were observed in the Alzheimer’s disease group, which were Alzheimer’s disease and bvFTD in the percentage of correct significantly different from controls in the horizontal (mean antisaccades, but bvFTD patients were able to make a higher difference 2.92 degrees, P = 0.000065) as well as in the vertical number of corrections than Alzheimer’s disease patients, and test (mean difference 1.43 degrees, P = 0.00098) (Figures 3, 4, this made that bvFTD patients obtained a significantly higher Table 2). This parameter also distinguished Alzheimer’s disease percentage of successful antisaccades (mean difference 33.56%, patients from the other two dementia groups in the horizontal P = 0.047). Additionally, the bvFTD group was superior to test (mean difference 3.14 degrees for bvFTD, P = 0.00072; 2.90 Alzheimer’s disease in the percentage of correct memory saccades degrees for svPPA, P = 0.012). (mean difference 14.36%, P = 0.014). In the antisaccade test, the performance of svPPA patients Parameters Related to Time was comparable to controls and this distinguished them from In the prosaccade test, Alzheimer’s disease patients showed the Alzheimer’s disease and bvFTD groups in the percentage longer latencies than controls (Table 2), with significant of correct and corrected antisaccades (Figure 4). However, at differences in horizontal (mean difference 152.10 ms, P = the memory saccade test, svPPA patients performed better than Frontiers in Aging Neuroscience | www.frontiersin.org 6 February 2021 | Volume 12 | Article 603790 TABLE 1 | Demographics and neuropsychological performance. Frontiers in Aging Neuroscience | www.frontiersin.org Lage et al. Controls Alzheimer’s BvFTD SvPPA Group effect Alzheimer’s disease Alzheimer’s disease BvFTD vs. svPPA disease P-value vs. bvFTD vs. svPPA P-value P-value P-value N 29 18 18 7 Age, years 66.21 (5.51) 68.17 68.83 (8.71) 70.86 (8.11) 0.36 0.99 0.83 0.92 (6.96) Gender, % female 79.31 66.67 22.22 57.14 0.0014 0.018 0.67 0.16 Disease duration, years – 4.94 (1.73) 5.56 (3.15) 6.00 (3.79) 0.65 0.79 0.67 0.93 Intake of sedatives, mean number 0.14 (0.44) 1.06 (0.87) 2.06 (1.11) 0.86 (0.90) <0.00001 0.0022 0.95 0.0073 MMSE (0–30) 28.96 (0.92) 16.72 23.50 (2.73) 22.43 (4.65) <0.00001 <0.00001 0.0015 0.89 (5.23) FCSRT Total Free and Cued Recall (0–48) 42.96 (5.50) 11.70 30.17 (12.55) 22.00 (10.22) <0.00001 0.000017 0.17 0.29 (9.11) FCSRT Delayed Free and Cued Recall (0–16) 15.07 (1.36) 2.60 (2.91) 9.67 (4.64) 6.00 (3.74) <0.00001 <0.00001 0.20 0.10 ROCFT Recall (0–36) 16.04 (5.23) 1.63 (3.11) 6.65 (5.18) 7.25 (3.30) <0.00001 0.091 0.25 0.10 ROCFT Copy (0–36) 32.86 (2.85) 20.13 25.06 (9.24) 28.75 (7.09) 0.000080 0.37 0.20 0.78 (11.14) Imitative praxis (0–8) 7.92 (0.27) 6.40 (1.90) 7.73 (0.59) 8.00 (0.00) 0.00017 0.0024 0.0076 0.93 VOSP NL (0–10) 9.21 (0.96) 6.50 (2.56) 7.72 (2.49) 8.80 (1.10) 0.0018 0.40 0.13 0.65 Trail Making Test A (seconds) 45.93 (12.70) 157.10 119.22 (66.24) 84.80 (25.15) <0.00001 0.19 0.036 0.48 (72.12) 7 Trail Making Test B (seconds) 103.52 (55.05) 179.50 220.60 (100.36) 145.50 (82.73) 0.0052 0.87 0.95 0.52 (95.46) Symbol digit test 39.55 (14.14) 16.43 19.33 (9.25) 21.60 (10.21) <0.00001 0.95 0.88 0.98 (12.33) Mean values of each demographic characteristic and neuropsychological score, with standard deviations in brackets (). Group effect calculated with ANOVA for the four groups. Only post hoc P-values for patient groups are shown. Bold values indicate P < 0.05. BvFTD, behavioral variant frontotemporal dementia; FCSRT, Free and Cued Selective Reminding Test; MMSE, Mini-Mental State Examination; ROCFT, Rey-Osterrieth Complex Figure Test; svPPA, semantic variant of primary progressive aphasia; VOSP NL, Visual Object and Space Perception Battery, Number Location subtest. r, correlation coefficient; P, level of significance. February 2021 | Volume 12 | Article 603790 Oculomotor Behavior in Dementia TABLE 2 | Visually-guided responses: Prosaccade test and sinusoidal smooth pursuit test. Frontiers in Aging Neuroscience | www.frontiersin.org Lage et al. Oculomotor parameter Controls Alzheimer’s disease vs. BvFTD vs. controls SvPPA vs. controls Alzheimer’s disease vs. Alzheimer’s disease vs. BvFTD vs. svPPA controls bvFTD svPPA Crude P-value (adjusted Crude P-value (adjusted Crude P-value (adjusted Crude P-value (adjusted Crude P-value (adjusted Crude P-value (adjusted P-value) P-value) P-value) P-value) P-value) P-value) Prosaccade latency (ms) Horizontal mean value (SD) 257.50 (52.24) 409.60 (226.60) 266.23 (34.20) 336.62 (160.48) 0.012 (0.011) 0.60 (0.28) 0.64 (0.31) 0.0016 (0.00017) 0.99 (0.49) 0.49 (0.094) Vertical mean value (SD) 264.80 (34.70) 315.14 (100.48) 282.79 (42.25) 322.63 (119.10) 0.53 (0.41) 1.00 (0.68) 0.58 (0.29) 0.099 (0.014) 0.84 (0.18) 0.21 (0.028) Return saccade latency (ms) Horizontal mean value (SD) 250.59 (72.96) 361.06 (202.30) 245.48 (28.48) 353.50 (219.70) 0.063 (0.12) 1.00 (0.95) 0.28 (0.24) 0.037 (0.0042) 1.00 (0.38) 0.27 (0.046) Vertical mean value (SD) 250.19 (36.30) 313.01 (133.12) 280.70 (56.47) 327.88 (114.81) 0.68 (0.62) 0.98 (0.62) 0.59 (0.38) 0.082 (0.016) 0.64 (0.10) 0.14 (0.025) Positive error (◦ ) Horizontal mean value (SD) 0.67 (0.73) 1.06 (1.43) 0.69 (0.62) 1.34 (1.08) 0.70 (0.31) 0.91 (0.44) 0.48 (0.13) 0.57 (0.16) 1.00 (0.88) 0.39 (0.074) Vertical mean value (SD) 0.54 (0.39) 0.95 (1.17) 0.60 (0.51) 0.78 (0.45) 0.48 (0.28) 0.95 (0.64) 0.94 (0.68) 0.25 (0.063) 0.99 (0.61) 0.85 (0.38) Negative error (◦ ) Horizontal mean value (SD) −0.41 (0.62) −3.23 (4.38) −1.60 (3.44) −0.68 (0.60) 0.43 (0.26) 0.32 (0.088) 0.43 (0.44) 0.022 (0.0076) 0.61 (0.25) 1.00 (0.95) 8 Vertical mean value (SD) −0.67 (1.44) −1.55 (2.39) −0.47 (0.37) −0.48 (0.34) 0.20 (0.070) 0.53 (0.15) 1.00 (0.94) 0.34 (0.15) 0.98 (0.59) 1.00 (0.63) Pursuit gain Horizontal mean value (SD) 0.79 (0.18) 0.89 (0.78) 0.87 (0.29) 0.67 (0.24) 1.00 (0.63) 0.72 (0.26) 0.78 (0.44) 0.87 (0.53) 0.93 (0.95) 0.94 (0.45) Vertical mean value (SD) 0.75 (0.33) 0.83 (1.02) 0.56 (0.60) 0.73 (0.13) 0.610 (0.034) 0.99 (0.55) 0.95 (0.29) 0.98 (1.00) 0.78 (0.031) 1.00 (0.54) Pursuit error (◦ ) Horizontal mean value (SD) 3.11 (2.01) 6.03 (3.01) 2.89 (1.93) 3.12 (1.51) 0.00055 (0.00072) 0.041 (0.012) 1.00 (0.93) 0.00043 (0.000065) 0.99 (0.88) 1.00 (0.85) Vertical mean value (SD) 2.23 (1.29) 3.66 (1.44) 2.56 (1.34) 2.95 (1.30) 0.090 (0.10) 0.69 (0.30) 0.93 (0.82) February 2021 | Volume 12 | Article 603790 0.0064 (0.00098) 0.87 (0.18) 0.65 (0.21) Bold values indicate P < 0.05. BvFTD, behavioral variant frontotemporal dementia; ms, milliseconds; SD, Standard deviations; svPPA, semantic variant of primary progressive aphasia; ◦ , degrees. r, correlation coefficient; P, level of significance. Oculomotor Behavior in Dementia TABLE 3 | Volitional saccades: Antisaccade test. Frontiers in Aging Neuroscience | www.frontiersin.org Lage et al. Oculomotor parameter Controls Alzheimer’s disease vs. BvFTD vs. controls SvPPA vs. controls Alzheimer’s disease vs. Alzheimer’s disease vs. BvFTD vs. svPPA controls bvFTD svPPA Crude P-value (adjusted Crude P-value (adjusted Crude P-value (adjusted Crude P-value (adjusted Crude P-value (adjusted Crude P-value (adjusted P-value) P-value) P-value) P-value) P-value) P-value) % Correct antisaccades 26.38 (18.75) 4.00 (5.41) 7.50 (10.80) 25.00 (29.33) 0.93 (0.85) 0.045 (0.016) 0.12 (0.038) (SD) 0.00031 (0.00037) 0.0022 (0.018) 1.00 (0.85) % Corrected 96.06 (11.11) 46.08 (34.94) 76.15 (33.96) 100.00 (0.00) 0.0066 (0.0496) 0.00016 (0.000058) 0.19 (0.026) antisaccades (SD) 1.61 × 10−6 (1.64 × 10−6 ) 0.057 (0.026) 0.98 (0.88) % Successful 92.24 (14.92) 36.44 (30.77) 70.00 (34.88) 83.33 (16.63) 0.0024 (0.047) 0.0015 (0.00065) 0.69 (0.12) antisaccades (SD) 1.36 × 10−7 (8.75 × 10−8 ) 0.030 (0.0072) 0.86 (0.26) Corrected antisaccade latency (ms) Horizontal mean value (SD) 587.14 (183.32) 1680.38 (879.79) 1112.34 (959.28) 676.84 (173.65) 0.093 (0.31) 0.0094 (0.0031) 0.48 (0.056) 0.000012 (2.50 × 10−06 ) 0.056 (0.0024) 0.99 (0.24) Vertical mean value (SD) 550.64 (90.90) 1485.63 (1066.28) 937.94 (564.51) 560.45 (139.67) 0.071 (0.33) 0.0090 (0.0034) 0.53 (0.054) 0.000055 (0.0000020) 0.18 (0.0053) 1.00 (0.38) Antisaccade positive error (◦ ) Horizontal mean value (SD) 2.43 (2.73) 4.56 (3.70) 8.40 (3.41) 5.69 (4.25) 0.048 (0.013) 0.92 (0.51) 0.33 (0.081) 0.36 (0.069) 5.20 × 10−05 (0.00019) 0.12 (0.020) Vertical mean value (SD) 1.72 (1.40) 4.87 (2.72) 3.46 (1.26) 2.50 (1.62) 0.22 (0.17) 0.043 (0.013) 0.67 (0.28) 0.000033 (0.000018) 0.022 (0.041) 0.74 (0.27) Antisaccade negative error (◦ ) Horizontal mean value (SD) −1.79 (1.10) −11.15 (10.66) −5.18 (9.81) −3.76 (1.94) 0.12 (0.14) 0.14 (0.035) 0.97 (0.38) 9 0.00099 (0.00013) 0.42 (0.046) 0.92 (0.47) Vertical mean value (SD) −1.58 (0.66) −5.88 (5.32) −5.01 (5.99) −1.48 (0.73) 0.95 (0.97) 0.11 (0.026) 0.22 (0.037) 0.019 (0.0042) 0.047 (0.039) 1.00 (1.00) Bold values indicate P < 0.05. BvFTD, behavioral variant frontotemporal dementia; ms, milliseconds; SD, Standard deviations; svPPA, semantic variant of primary progressive aphasia; ◦ , degrees. r, correlation coefficient; P, level of significance. February 2021 | Volume 12 | Article 603790 Oculomotor Behavior in Dementia Frontiers in Aging Neuroscience | www.frontiersin.org Lage et al. TABLE 4 | Volitional saccades: Memory saccade test. Oculomotor parameter Controls Alzheimer’s disease vs. BvFTD vs. controls SvPPA vs. controls Alzheimer’s disease vs. Alzheimer’s disease vs. BvFTD vs. svPPA controls bvFTD svPPA Crude P-value (adjusted Crude P-value (adjusted Crude P-value (adjusted Crude P-value (adjusted Crude P-value (adjusted Crude P-value (adjusted P-value) P-value) P-value) P-value) P-value) P-value) % Correct memory 94.14 (9.74) 26.00 (19.41) 40.36 (25.15) 64.17 (36.94) 0.30 (0.014) 0.0022 (0.00019) 0.072 (0.14) saccades (SD) 2.71 × 10−11 1.01 × 10−09 (0.00064) 0.0065 (0.025) (1.22 × 10−09 ) Memory saccade latency (ms) Horizontal mean value (SD) 410.26 (197.59) 852.58 (314.01) 695.44 (343.24) 517.37 (245.48) 0.55 (0.92) 0.12 (0.041) 0.58 (0.068) 0.00059 (0.000069) 0.014 (0.0029) 0.83 (0.18) Vertical mean value (SD) 384.39 (115.94) 1163.14 (903.67) 866.05 (461.76) 492.58 (332.35) 0.46 (0.14) 0.056 (0.0093) 0.38 (0.26) 0.00043 (0.00088) 0.0076 (0.26) 0.96 (0.92) Memory saccade positive error (◦ ) Horizontal mean value (SD) 1.48 (0.76) 2.53 (3.39) 1.38 (1.37) 1.53 (0.75) 0.49 (0.25) 0.67 (0.29) 1.00 (0.97) 0.39 (0.11) 1.00 (0.88) 1.00 (0.86) Vertical mean value (SD) 0.98 (0.56) 1.44 (1.71) 0.91 (0.64) 0.88 (0.44) 0.73 (0.32) 0.77 (0.34) 1.00 (0.94) 0.75 (0.32) 1.00 (0.89) 1.00 (0.84) Memory saccade negative error (◦ ) Horizontal mean value (SD) −0.72 (0.67) −3.82 (2.11) −2.37 (1.91) −0.86 (0.54) 0.13 (0.42) 0.0025 (0.00032) 0.18 (0.0055) 0.000019 (2.14 × 10−06 ) 0.011 (0.00019) 1.00 (0.20) Vertical mean value (SD) −0.98 (1.37) −2.17 (1.30) −2.30 (1.47) −1.41 (1.52) 1.00 (0.39) 0.85 (0.57) 0.73 (0.15) 10 0.34 (0.27) 0.14 (0.095) 0.94 (0.73) Bold values indicate P < 0.05. BvFTD, behavioral variant frontotemporal dementia; ms, milliseconds; SD, Standard deviations; svPPA, semantic variant of primary progressive aphasia; ◦ , degrees. r, correlation coefficient; P, level of significance. February 2021 | Volume 12 | Article 603790 Oculomotor Behavior in Dementia Lage et al. Oculomotor Behavior in Dementia FIGURE 3 | Performance of controls and dementia groups in oculomotor parameters related to time (A), success (B), and spatial accuracy (C,D). Asterisks mark those dementia groups which show significant differences (P < 0.05) compared to controls. ms, milliseconds; ◦ , degrees. the other two dementia groups but still significantly worse than to investigate for correlations. Instead, we selected the following controls (mean difference 29.97%; P = 0.025). most significant parameter, horizontal corrected antisaccade latency. The neuropsychological evaluation was available in all controls and bvFTD patients, but only in 10 Alzheimer’s disease Relationship Between Oculomotor and five svPPA patients. Due to the small number, svPPA was not Behavior and Neuropsychological included in this sub-analysis. Performance As a measure of global cognition, MMSE showed mild to As expected, significant differences across the four groups moderate correlations with prosaccade and corrected antisaccade were found for all neuropsychological tests, with the highest latencies and the percentage of correct memory saccades, performance in controls (Table 1). Alzheimer’s disease patients especially in Alzheimer’s disease and controls (Table 5). performed significantly worse than bvFTD in verbal memory and Prosaccade latencies showed moderate to strong negative imitative praxis. No significant differences were found between correlations with ROCFT copy and imitative praxis both in bvFTD and svPPA patients. Alzheimer’s disease and bvFTD patients and with Symbol digit For the comparison with cognitive and brain metabolic test in controls, meaning that longer latencies were associated results, we selected the oculomotor parameter with the lowest P- with worse cognitive performance. Regarding corrected value from each test, which were: horizontal prosaccade latency, antisaccade latencies, moderate correlations were found with horizontal pursuit error and percentage of correct memory executive function tests in controls, including TMT-A and B and saccades. In the antisaccade test, the percentage of successful Symbol digit test. Alzheimer’s disease patients also showed strong antisaccades showed the lowest P-value, but, since this is a correlations between corrected antisaccade latencies and TMT-A composite measure, it was not considered an optimal parameter and ROCFT copy, but it lost significance in the multivariate Frontiers in Aging Neuroscience | www.frontiersin.org 11 February 2021 | Volume 12 | Article 603790 Lage et al. Oculomotor Behavior in Dementia FIGURE 4 | Some clinical examples of oculomotor evaluations. (A,B) Show the performance of a patient with Alzheimer’s disease in an horizontal prosaccade test (A) and an horizontal sinusoidal smooth pursuit test (B), with great difficulties in overlapping his gaze with the moving target in the latter. (C,D) Illustrate the different performance in the horizontal antisaccade test between a patient with bvFTD (C), with some successful responses but that are slow and inaccurate; and a patient with svPPA (D), who also makes some errors but followed by fast corrections in this case. In the ordinate axis, 0 indicates the center, positive values the right side and negative values the left side. Blue lines represent the patient’s ocular movement and dotted lines the expected location of the gaze, which corresponds to the target position in the prosaccade and sinusoidal smooth pursuit tests, and to the opposite position of the target in the antisaccade test. Blinks are marked in yellow and pupil detection failure (usually also due to blinks) in red. deg, degrees; ms, milliseconds. analysis. Pursuit error was negatively correlated with the Symbol the three disease groups. Finally, no significant associations were digit test in bvFTD and controls, and also with ROCFT recall found with the percentage of correct memory saccades. and copy in Alzheimer’s disease and bvFTD groups, respectively. Finally, correct memory saccades were the only oculomotor parameter associated with verbal memory, strongly correlated Classification Algorithm with Total Free and Delayed Recall and borderline significant to Each algorithm was trained using those oculomotor parameters visual recall in Alzheimer’s disease patients. that, after the normalization process, demonstrated significant differences between groups in the univariate analysis at a significance level of P < 0.001. Furthermore, parameters that Relationship Between Oculomotor were susceptible to having unavailable values, as time or Behavior and Regional Brain Metabolism accuracy-related parameters of correct antisaccades or memory 18F-FDG PET was available for 18 patients (six Alzheimer’s saccades, were excluded. The oculomotor parameters that disease, seven bvFTD, and five svPPA). We investigated were finally included in each algorithm can be found in the for negative associations with horizontal prosaccade latency, Supplementary Material. horizontal pursuit error and horizontal corrected antisaccade Based on the accuracy of each algorithm, we selected the latency; and for positive associations with the percentage of best classifier with higher values in ROC curves, which were: correct memory saccades. Brain regions that showed significant Supported Vector Machine for the Alzheimer’s disease vs. associations at an uncorrected threshold of P < 0.001 (k>100 controls and bvFTD vs. controls pairs; and K-Nearest Neighbors mm3) are depicted in Figure 5. for the Alzheimer’s disease vs. bvFTD pair. The cross-validation Prosaccade latency was negatively correlated with the under C++ environment offered the final results, which are metabolism in the left middle temporal gyrus and precuneus. summarized in Table 6. The classifiers demonstrated a mean This association was mainly driven by Alzheimer’s disease cases AUC of 97.5% for the differentiation between Alzheimer’s disease (see scatterplot on Figure 5A). A greater corrected antisaccade vs. controls, 96.7% for bvFTD vs. controls, and 92.5% for latency was associated with decreased metabolism in the right Alzheimer’s disease vs. bvFTD. middle frontal gyrus. Pursuit error was negatively correlated with The classifiers for the differentiation between controls the metabolism in the right posterior middle temporal gyrus. In and each dementia group were subsequently applied in the both cases, the association was more uniformly distributed over independent samples of Alzheimer’s disease and bvFTD to test Frontiers in Aging Neuroscience | www.frontiersin.org 12 February 2021 | Volume 12 | Article 603790 Frontiers in Aging Neuroscience | www.frontiersin.org Lage et al. TABLE 5 | Relationship between neuropsychological performance and oculomotor parameters. Test Prosaccade horizontal latency (r) Corrected antisaccade horizontal latency (r) Horizontal pursuit error (r) % Correct memory saccades (r) Crude P-value (adjusted P-value) Crude P-value (adjusted P-value) Crude P-value (adjusted P-value) Crude P-value (adjusted P-value) Controls Alzheimer’s BvFTD Controls Alzheimer’s BvFTD Controls Alzheimer’s BvFTD Controls Alzheimer’s BvFTD disease disease disease disease MMSE −0.64 −0.58 −0.13 −0.36 −0.55 −0.23 0.014 −0.0062 −0.037 0.29 0.049 0.17 (0–30) 0.00032 (0.013) 0.012 (0.11) 0.65 (0.030) 0.058 (0.0061) 0.049 (0.065) 0.42 (0.76) 0.94 (0.93) 0.98 (0.89) 0.88 (0.31) 0.14 (0.028) 0.89 (0.89) 0.57 (0.50) FCSRT Total Free and −0.0079 0.12 −0.22 −0.32 0.29 0.016 −0.067 −0.15 −0.36 0.24 0.90 0.14 Cued Recall (0–48) 0.97 (0.43) 0.75 (0.48) 0.42 (0.32) 0.10 (0.21) 0.53 (0.96) 0.96 (0.47) 0.75 (0.88) 0.68 (0.062) 0.14 (0.36) 0.22 (0.41) 0.037 (0.016) 0.63 (0.78) FCSRT Delayed Free −0.078 0.12 −0.081 −0.20 0.14 −0.03 −0.22 0.062 −0.40 0.19 0.87 0.18 and Cued Recall (0–16) 0.70 (0.75) 0.73 (0.45) 0.77 (0.37) 0.31 (0.30) 0.76 (0.75) 0.91 (0.60) 0.27 (0.18) 0.87 (0.16) 0.10 (0.17) 0.34 (0.74) 0.053 0.55 (0.65) (0.00031) ROCFT Recall (0–36) −0.20 −0.15 −0.27 −0.22 0.25 0.21 −0.25 −0.48 −0.32 −0.026 0.89 0.26 0.31 (0.69) 0.72 (0.81) 0.31 (0.26) 0.26 (0.11) 0.64 (0.58) 0.49 (0.30) 0.23 (0.33) 0.23 (0.038) 0.22 (0.53) 0.90 (0.92) 0.041 (0.054) 0.38 (0.50) ROCFT Copy (0–36) −0.058 −0.71 −0.62 −0.0053 −0.99 −0.28 −0.19 −0.35 −0.49 0.42 −0.41 0.041 0.77 (0.81) 0.050 0.010 (0.0063) 0.98 (0.085) 0.00031 0.35 (0.90) 0.35 (0.42) 0.40 (0.37) 0.045 0.026 (0.053) 0.49 (0.63) 0.88 (0.84) (0.00092) (0.17) (0.0043) Imitative praxis (0–8) 0.29 −0.72 −0.52 0.000 −0.40 −0.097 −0.26 −0.36 −0.22 0.19 0.71 −0.30 13 0.17 (0.30) 0.020 0.073 (0.038) 1.00 (0.99) 0.38 (0.43) 0.77 (0.85) 0.22 (0.16) 0.32 (0.42) 0.44 (0.43) 0.34 (0.80) 0.18 (0.48) 0.37 (0.77) (0.0019) VOSP NL −0.13 −0.084 0.014 −0.16 −0.76 0.21 −0.10 −0.16 −0.45 0.10 −0.20 0.16 (0–10) 0.52 (0.85) 0.84 (0.19) 0.96 (0.96) 0.42 (0.11) 0.046 (0.18) 0.47 (0.75) 0.61 (0.42) 0.71 (0.91) 0.059 (0.41) 0.61 (0.73) 0.75 (0.71) 0.60 (0.56) Trail making test A 0.15 0.49 0.32 0.50 0.82 −0.033 0.11 0.16 0.45 −0.20 −0.30 0.33 (seconds) 0.46 (0.30) 0.15 (0.31) 0.23 (0.012) 0.0074 (0.0062) 0.024 (0.10) 0.91 (0.90) 0.60 (0.10) 0.65 (0.55) 0.062 (0.21) 0.32 (0.99) 0.62 (0.17) 0.25 (0.50) Trail making test B 0.55 – 0.30 0.64 – 0.20 0.25 – 0.30 0.00 – 0.11 (seconds) 0.0038 (0.21) 0.62 (0.32) 0.00037 (0.000025) 0.75 (0.37) 0.24 (0.18) 0.62 (0.52) 1.00 (0.99) 0.90 (0.97) Symbol digit test −0.61 −0.054 −0.25 −0.60 −0.60 0.070 −0.27 −0.16 −0.63 0.052 0.80 0.19 0.0031 (0.015) 0.90 (0.68) 0.36 (0.29) 0.0040 (0.0016) 0.29 (0.22) 0.81 (0.24) 0.24 (0.046) 0.73 (0.59) 0.0051 0.81 (0.25) 0.20 (0.34) 0.52 (0.72) (0.023) February 2021 | Volume 12 | Article 603790 Bold values indicate P < 0.05. Trail Making Test B could not be calculated in Alzheimer’s disease sample since it was carried out in only two of these patients. BvFTD, behavioral variant frontotemporal dementia; FCSRT, Free and Cued Selective Reminding Test; MMSE, Mini-Mental State Examination; ROCFT, Rey-Osterrieth Complex Figure Test; VOSP NL, Visual Object and Space Perception Battery, Number Location subtest. r, correlation coefficient; P, level of significance. Oculomotor Behavior in Dementia Lage et al. Oculomotor Behavior in Dementia FIGURE 5 | Relationship between brain metabolism and horizontal prosaccade latency (A), horizontal corrected antisaccade latency (B) and horizontal pursuit error (C). SUVR, Standard Uptake Value Rate. their external validity. The independent sample of Alzheimer’s 0.00022) (Table 7). The AUC in this sample was 73.3%. For the disease patients was composed of 15 individuals without a replication of the bvFTD vs. controls algorithm, the independent biomarker-supported diagnosis who were in a mild dementia sample included six patients with similar age, disease duration, stage (Global Deterioration Scale = 4) as the discovery sample MMSE and number of sedatives to that of the discovery sample. A and showed comparable MMSE scores and number of sedatives, biomarker-supported diagnosis was available in two cases. Here, but were significantly older (mean difference 9.50 years, P = the AUC was 83.3%. Frontiers in Aging Neuroscience | www.frontiersin.org 14 February 2021 | Volume 12 | Article 603790 Lage et al. Oculomotor Behavior in Dementia TABLE 6 | Performance of the machine learning classifiers. Pairs of groups Best AUC Worst AUC Mean AUC Best accuracy Worst accuracy Mean accuracy Selected machine learning algorithm Alzheimer’s disease vs. controls 0.9985 0.95 0.9753 0.973 0.8649 0.9478 Supported vector machine bvFTD vs. controls 0.9912 0.9147 0.9674 0.9459 0.8919 0.9323 Supported vector machine Alzheimer’s disease vs. bvFTD 0.9722 0.7778 0.9246 0.9722 0.7778 0.9246 k-Nearest neighbors AUC, area under the curve; BvFTD, behavioral variant frontotemporal dementia. TABLE 7 | Demographics of Alzheimer’s disease and bvFTD independent samples. Alzheimer’s disease independent sample BvFTD independent sample Characteristic Mean difference from discovery Characteristic Mean difference from discovery sample (P-value) sample (P-value) n 15 6 Age, years 77.67 (5.90) 9.50 (0.00022) 71.40 (3.78) 2.57 (0.53) Gender, % female 53.33 – (0.49) 40.00 – (0.58) Disease duration, years 3.91 (1.38) −1.04 (0.10) 2.75 (1.71) −2.81 (0.10) Intake of sedatives, mean number 0.86 (0.86) −0.20 (0.53) 1.50 (0.71) −0.56 (0.50) MMSE (0–30) 19.27 (3.71) 2.54 (0.12) 25.00 (2.55) 1.50 (0.28) Mean values of each demographic characteristic with standard deviations in brackets () and mean differences in comparison with the discovery sample with level of significance (P-value), calculated with T-test. Bold values indicate P < 0.05. BvFTD, behavioral variant frontotemporal dementia; MMSE, Mini-Mental State Examination; svPPA, semantic variant of primary progressive aphasia. r, correlation coefficient; P, level of significance. DISCUSSION the other three groups. Reported findings in Alzheimer’s disease patients in this test are decreased velocity or gain (Molitor In this work, we have compared the oculomotor behavior of et al., 2015). The two studies that have investigated neuroimaging Alzheimer’s disease, bvFTD and svPPA patients, evaluated with correlates in patients with dementia did not find any association a new video-oculography system and quantitatively described by with a specific cortical region (Boxer et al., 2006; Shakespeare a wide range of parameters, grouped through the axes of spatial et al., 2015), but these studies assessed other pursuit parameters accuracy, time and success. Their assessment across the four than pursuit error. According to current literature, the middle oculomotor tests offered a pattern that could differentiate each temporal complex (also named V5), a region localized bilaterally group from the others (Figure 6) and reflected the characteristic in lateral occipitotemporal cortex (Coiner et al., 2019), constitutes neuroanatomical distribution of pathology of each disease. the key cortical area for smooth pursuit since it is implicated Alzheimer’s disease was the most impaired group in the three in extracting motion information from the moving target and axes and, as expected, displayed specific impairments in some afterwards sending it to pursuit-specific regions of the frontal visually-guided saccade parameters that are theoretically closely and supplementary eye fields (Petit and Haxby, 1999; Krauzlis, linked to posterior brain regions, as pursuit error and horizontal 2004; Thier and Ilg, 2005). Here, we found that pursuit error prosaccade latency. Conversely, bvFTD patients showed deficits was associated with decreased metabolism in the right posterior especially in the more challenging volitional saccades, which middle temporal gyrus, in a region located closely to middle require a fine control exerted mainly from frontal lobe regions; temporal complex. One possibility is that this absence of while svPPA patients performed in general similarly to controls complete concordance could be related to the relatively low except from a lower success rate in the memory saccade test. inaccuracy of the spatial normalization and resolution of the PET. In this way, increased values of pursuit error might be considered Parameters Related to Spatial Accuracy: as a surrogated marker of posterior cortical damage and therefore Pursuit Error Distinguishes Alzheimer’s be suggestive of Alzheimer’s disease pathology in a given patient, as opposed to other conditions that predominantly harm frontal Disease lobe structures. Alzheimer’s disease patients showed the least accurate saccades, tending to obtain the highest values of spatial error not only in their volitional saccades but also in their prosaccades, which Parameters Related to Time: Different could suggest a more impaired visuospatial function. However, Meanings of Visually-Guided and Volitional probably one of the most remarkable findings of this work Saccade Latencies is their performance on the smooth pursuit test, where these Alzheimer’s disease patients also showed longer mean latencies patients obtained significantly higher pursuit error values than than the other three groups across all paradigms, which replicates Frontiers in Aging Neuroscience | www.frontiersin.org 15 February 2021 | Volume 12 | Article 603790 Lage et al. Oculomotor Behavior in Dementia et al., 2018), while others not (Boxer et al., 2006, 2012; Garbutt et al., 2008). In our work, bvFTD patients showed prosaccade latencies comparable to those of controls, but longer than them for corrected antisaccades and memory saccades. This underlines the differences between the processing of visually-guided and volitional saccades, where latencies not only represent the time required to perceive a peripheral target and to generate a saccade, but also to build an internal spatial representation of the demanded saccade without visual reference and to process the intended response, steps which finally increase the global amount of time (McDowell et al., 2008). According to this, latencies for corrected antisaccades and memory saccades were prominently longer than for prosaccades in all groups, and especially in Alzheimer’s disease and bvFTD. In controls, corrected antisaccade latencies showed significant correlations with TMT-A, Symbol digit test and, with the greatest coefficient, with TMT-B, a test that assesses visuomotor speed and cognitive flexibility, which agrees with previous works which have found associations between executive function and antisaccade performance (Meyniel et al., 2005; Heuer et al., 2013; Douglass et al., 2018). Previous publications have reported associations between antisaccade parameters and diverse frontal regions in patients with dementia and normal elders (Boxer et al., 2006; Mirsky et al., 2011). Here, we also found a strong negative correlation between corrected antisaccade latency and metabolism in a region of the right middle frontal gyrus located FIGURE 6 | Useful oculomotor parameters for the differential diagnosis in the DLPFC. The DLPFC is considered to play a decisional role between Alzheimer’s disease and the other two dementia groups (A), and in the antisaccade task through the inhibition of prosaccades, between bvFTD and svPPA (B). AS, antisaccade; MS, memory saccade; PS, as well as supporting working memory in maintaining the task prosaccade. goal, which allows performing a successful antisaccade (Pierrot- Deseilligny et al., 2003). Right lateralization of DLPFC activation during antisaccade tasks has been reported before (Ford et al., 2005; Ettinger et al., 2008; Jamadar et al., 2013). Considering that one of the most characteristic findings in the literature. Lesion corrected antisaccade latency measures the time required to make studies have traditionally linked an increment in prosaccade a correction after an erroneous antisaccade, the association found latencies to parietal cortex damage, a region considered to with right DLPFC might be more related to working memory play an important role in saccade triggering, supported by the activity and task monitoring, which has been specifically linked existence of direct projections to the superior colliculus (Pierrot- to right DLPFC (Ford et al., 2005; Kaufman et al., 2010). Deseilligny et al., 1991; Gaymard et al., 2003). Similarly, in Contrarily to the findings in Alzheimer’s disease and bvFTD, Alzheimer’s disease samples, it has been related to increased svPPA latencies of corrected antisaccades and memory saccades parietal atrophy (Shakespeare et al., 2015). In our study, were comparable to controls and significantly shorter than in prosaccade latencies showed the strongest correlations to Alzheimer’s disease. SvPPA constitutes an entity in that initial constructional and imitative praxis in the neuropsychological pathology tends to affect selectively anterior temporal poles, evaluation, which supports this hypothesis. The associations which leads to a specific impairment of semantic memory with a with brain metabolism involved the left middle temporal relative sparing of other cognitive abilities during the first stages and precuneus. Precuneus is considered to be part of the of the disease. Considering that the reported cortical substrate oculomotor network (Coiner et al., 2019) and has previously of oculomotor behavior relies on a core frontoparietal network, demonstrated functional activation during different saccade it is not surprising that the three published works that offer responses, including prosaccades (Jamadar et al., 2013). However, disaggregated data from svPPA subjects (Boxer et al., 2006, 2012; in our study, these associations seem to be driven by some Garbutt et al., 2008) have found performances indistinguishable Alzheimer’s disease cases. Since these regions are those primarily from healthy controls. Therefore, we find intriguing that, in our targeted in Alzheimer’s disease, we cannot exclude that this result study, svPPA patients tended to show longer prosaccade latencies, reflects between-group differences rather than the metabolic in contrast to their preserved volitional saccade latencies. It may substrates of prosaccade latency. be noted that Alzheimer’s disease and svPPA patients made a total In the case of bvFTD, some previous works also observed number of early saccades significantly greater than controls (see significantly longer saccade latencies in comparison with healthy Supplementary Table 1), but statistically significant differences controls (Meyniel et al., 2005; Burrell et al., 2012; Douglass in prosaccade latencies remained significant after accounting Frontiers in Aging Neuroscience | www.frontiersin.org 16 February 2021 | Volume 12 | Article 603790 Lage et al. Oculomotor Behavior in Dementia for it. Since the generation of prosaccades is supported by a all svPPA patients were able to perform some correct memory widespread network, one possible explanation is that, in svPPA saccades, with the lowest percentage of success of 20% in two patients, the increase in prosaccade latencies is not due to parietal patients. Moreover, this sample achieved high rates of success in damage, as reported in Alzheimer’s disease, but by a disruption in the antisaccade test, which may be equally difficult to understand. other nodes of the neural circuitry. Supporting this idea is the fact that, in controls, prosaccade latencies were not correlated Classification Algorithms to constructive and imitative praxis as in Alzheimer’s disease, The last objective of this work was to test whether a classification but with the Symbol digit test, and, in bvFTD, also to TMT- algorithm based on machine learning techniques could be useful A, both tests which assess executive function. Furthermore, for the differential diagnosis of these diseases. Machine learning in the work which includes the largest sample of svPPA with is a subfield of the artificial intelligence disciple that leverages 19 subjects (Garbutt et al., 2008), prosaccade latencies showed numerical techniques to automatically “learn” programs for significant correlations not only to parietal and occipital but also performing these tasks by processing a huge quantity of data. The to right temporal volumes. Consequently, we may hypothesize use of machine learning algorithms applied to different medical that the same impairment in an oculomotor parameter could be data has increased in the last decade and, in particular, those due to different anatomic damage depending on the underlying applications related with medical images (De Fauw et al., 2018; condition. This also suggests that oculomotor parameters which Rajpurkar et al., 2018; Elaziz et al., 2020; Stemmer et al., 2020). rely on more widespread networks could be less useful for However, machine learning techniques arise as a powerful tool the differential diagnosis between pathologies, while parameters to analyses data coming from different sources (Corey et al., more tightly controlled by a particular brain region, as could be 2018; Fontanella et al., 2018), in order to assist clinicians in the the case of pursuit error, might constitute better markers of a diagnosis stage (Kim et al., 2017; Viscaino et al., 2020). Moreover, specific pathology. a previous study described a machine learning model based on smooth pursuit data that discriminated age-matched controls from young-onset Alzheimer’s disease patients with ∼95% Parameters Related to Success: A Pattern accuracy (Pavisic et al., 2017). In our work, we incorporated of Impairment for Each Disease the most significant parameters from the four oculomotor tasks The lowest success rates were found in the Alzheimer’s to train algorithms that were able to distinguish Alzheimer’s disease group, with significantly lower percentages of successful disease, bvFTD, and controls. These results were replicated antisaccades and correct memory saccades than the other three in independent samples, and, in the case of bvFTD patients, groups. According to the literature, the bvFTD group also made also from a different memory clinic. Although the classification fewer correct antisaccades (Meyniel et al., 2005) and memory accuracy of the algorithm for differentiating Alzheimer’s disease saccades (Douglass et al., 2018) than controls. Some previous from controls was lower in the replication sample, it must be works have described that bvFTD patients were able to correct taken into consideration that this group lacked biomarkers. as many errors as controls and more than Alzheimer’s disease Thus, it was potentially a more heterogeneous sample, posing patients (Boxer et al., 2006; Garbutt et al., 2008), while others a greater challenge to the classification algorithm, but that at did not (Boxer et al., 2012; Douglass et al., 2018). Here, bvFTD the same time it was a more representative sample of everyday patients exhibited an intermediate ability to correct antisaccade practice and therefore a more rigorous test of its external validity. errors, significantly greater than the Alzheimer’s disease group Additionally, the discovery sample was composed of younger but not comparable to controls. In conclusion, bvFTD patients patients that tended to be more cognitively impaired, with non- displayed a normal performance in the prosaccade test, but significantly lower MMSE scores. Despite this, the algorithm was showed deficits, including longer latencies, greater values of able to identify 11 out of 15 patients of the replication sample. All spatial error and lower success rates, only in the more cognitively in all, our results should be considered only as an encouraging demanding volitional saccades, which implicate a supportive role first step toward the application of oculomotor evaluations in from frontal regions as the DLPFC or anterior cingulate cortex medical practice, where machine learning approaches offer the (McDowell et al., 2008; Jamadar et al., 2013). possibility of incorporating information from a high number The performance of svPPA patients was comparable to of parameters. controls in the antisaccade test, but lower in the memory saccade test. To our best knowledge, an evaluation of memory saccades Limitations in svPPA has not been described before, so we consider that A potential limitation of this study is the small sample this is a novel finding. This way, svPPA could be differentiated sizes, especially in the case of svPPA patients, where results from controls by the percentage of correct memory saccades; and must be taken with caution. In spite of this, most findings from bvFTD and Alzheimer’s disease patients by the percentage are concordant with previous publications, and we think of correct and corrected antisaccades. Given that the instructions that these results are reinforced by the fact that exclusively for the memory saccade test are more complicated than for patients with Alzheimer’s disease core biomarker-supported the visually-guided tests, it is important to consider that the diagnoses were included, which, to our best knowledge, is comprehension impairment that characterizes svPPA may be the first time both for the case of Alzheimer’s disease and a limiting factor for the successful performance of this test. FTD literature. The only previous work in FTD patients However, we think that, in our sample, this is less probable, since with available Alzheimer’s disease core biomarkers reported Frontiers in Aging Neuroscience | www.frontiersin.org 17 February 2021 | Volume 12 | Article 603790 Lage et al. Oculomotor Behavior in Dementia the results of CSF biomarkers in only six out of 14 patients, DATA AVAILABILITY STATEMENT and oculomotor information was not disaggregated from the whole sample that included bvFTD, svPPA, non-fluent The original contributions presented in the study are included progressive aphasia, and motor neuron disease-FTD (Coppe in the article/Supplementary Material, further inquiries can be et al., 2012). Also, to maximize homogeneity in our discovery directed to the corresponding author/s. sample, all included patients were in the same functional stage and showed similar disease duration, a relevant ETHICS STATEMENT factor considering that it has been described that some oculomotor parameters become increasingly impaired with The studies involving human participants were reviewed and disease progression (Molitor et al., 2015; Kahana Levy et al., approved by Comité de Ética de Investigación Clínica de 2018). Cantabria (CEIC), Marqués de Valdecilla Health Research Additionally, an issue that might limit clinical translation is Institute. The patients/participants provided their written that some patients might not be able to fully understand the informed consent to participate in this study. most cognitively demanding tasks, the antisaccade and memory saccade tests. However, visually-guided saccade tests, which AUTHOR CONTRIBUTIONS could be used in moderately-demented patients, were also useful in differentiating Alzheimer’s disease from the other groups. CL: study design, oculomotor evaluations, data analysis, and Finally, we reported neuroimaging results at an uncorrected drafting of the manuscript. SL-G, MK, and IA-B: oculomotor threshold, since the 18F-FDG PET evaluation was available evaluations and collection of clinical data. AB: neuroimaging only in 18 patients and our statistical power was low. data analysis. AC-C: oculomotor data processing and artificial Additionally, we cannot exclude that the associations found intelligence procedures. AP and MG-M: neuropsychological between prosaccade latency and brain metabolism are driven assessments. AF-R and MB-G: collection of biological samples by some Alzheimer’s disease cases, since increased latencies and coordination of subject recruiting. JJ-B and IB: neuroimaging and hypometabolism in middle temporal gyrus and precuneus evaluations. JI-V: CSF analytical procedures. JP: neuroimaging can be a shared finding in this disease. Conversely, the results processing. CG-C: study concept and design and artificial regarding corrected antisaccade latency and pursuit error are intelligence procedures. PS-J: study concept and design and concordant with the current literature since they involve well- statistical advisement. ER-R, AL-B, JF, II-G, and all authors: known related core regions of the oculomotor system, such critical revision of the manuscript. as the DLPFC and the posterior middle temporal gyrus. In any case, we recommend extreme caution in the interpretation FUNDING of the brain metabolism results and we suggest considering them as supportive findings of the oculomotor outcomes rather PS-J was supported by grants from IDIVAL, Instituto de than firm conclusions about the neuroanatomic correlates of Salud Carlos III (Fondo de Investigación Sanitario, PI08/0139, oculomotor parameters. PI12/02288, PI16/01652), JPND (DEMTEST PI11/03028), the CIBERNED program, Siemens Healthineers (Valdecilla Study of Conclusion and Future Works Memory and Brain Aging) and Aura Innovative Robotics S.L. In this work, we have described differential patterns in The latter provided eye-tracking equipment and collaborated oculomotor behavior in typical Alzheimer’s disease, bvFTD with the data analysis blinded to clinical information. The and svPPA and introduced a machine learning methodology funders were not involved in the study design, collection, that could be applied in clinical practice to patients who analysis, interpretation of data, the writing of this article or the often pose a diagnostic challenge. Some of the reported decision to submit it for publication. II-G is supported by the findings, especially those related to success parameters, are Global Brain Health Institute (Atlantic Fellow for Equity in Brain also accessible for routine clinical examination. Further works Health) and the Rio Hortega grant (CM17/00074) from Instituto should address whether some oculomotor features are early de Salud Carlos III and the European Social Fund. enough to uncover the presence of neurodegenerative changes in prodromal stages, and specific enough of the characteristic ACKNOWLEDGMENTS spreading of cortical and network neurodegeneration of each disease to distinguish them even when the clinical picture We are deeply grateful to all participants who took part in is less defined, as atypical presentations of Alzheimer’s this study. disease. Furthermore, it would be of great interest to analyse whether reported correlations between oculomotor SUPPLEMENTARY MATERIAL parameters and neuroimaging in healthy controls and lesion studies are the same in neurodegenerative diseases, due to The Supplementary Material for this article can be found the deep complexity of the neuroanatomic substrates of online at: https://www.frontiersin.org/articles/10.3389/fnagi. oculomotor behavior. 2020.603790/full#supplementary-material Frontiers in Aging Neuroscience | www.frontiersin.org 18 February 2021 | Volume 12 | Article 603790 Lage et al. Oculomotor Behavior in Dementia REFERENCES Gaymard, B., Lynch, J., Ploner, C. J., Condy, C., and Rivaud-Péchoux, S. (2003). The parieto-collicular pathway: anatomical location and Alcolea, D., Clarimón, J., Carmona-Iragui, M., Illán-Gala, I., Morenas- contribution to saccade generation. Eur. J. Neurosci. 17, 1518–1526. Rodríguez, E., Barroeta, I., et al. (2019a). The sant pau initiative on doi: 10.1046/j.1460-9568.2003.02570.x neurodegeneration (SPIN) cohort: a data set for biomarker discovery and Gorno-Tempini, M. L., Hillis, A. E., Weintraub, S., Kertesz, A., Mendez, M., Cappa, validation in neurodegenerative disorders. Alzheimer’s Dement. 5, 597–609. S. F., et al. (2011). Classification of primary progressive aphasia and its variants. doi: 10.1016/j.trci.2019.09.005 Neurology 76, 1006–1014. doi: 10.1212/WNL.0b013e31821103e6 Alcolea, D., Pegueroles, J., Muñoz, L., Camacho, V., López-Mora, D., Fernández- Graf, A. B. A., and Borer, S. (2001). “Normalization in support vector machines,” in León, A., et al. (2019b). Agreement of amyloid PET and CSF biomarkers for Pattern Recognit, eds B. Radig, S. Florczyk (Berlin: Springer Berlin Heidelberg), alzheimer’s disease on lumipulse. Ann. Clin. Transl. Neurol. 6, 1815–1824. 277–82. doi: 10.1007/3-540-45404-7_37 doi: 10.1002/acn3.50873 Hernández, E., Hernández, S., Molina, D., Acebrón, R., and Garcia Cena, C. E. Boxer, A. L., Garbutt, S., Rankin, K. P., Hellmuth, J., Neuhaus, J., Miller, B. L., et al. (2018). OSCANN: Technical characterization of a novel gaze tracking analyzer. (2006). Medial versus lateral frontal lobe contributions to voluntary saccade Sensors 18, 1–16. doi: 10.3390/s18020522 control as revealed by the study of patients with frontal lobe degeneration. J. Heuer, H. W., Mirsky, J. B., Kong, E. L., Dickerson, B. C., Miller, B. L., Kramer, J. Neurosci. 26, 6354–6363. doi: 10.1523/JNEUROSCI.0549-06.2006 H., et al. (2013). Antisaccade task reflects cortical involvement in mild cognitive Boxer, A. L., Garbutt, S., Seeley, W. W., Jafari, A., Heuer, H. W., Mirsky, J., impairment. Neurology 81, 1235–1243. doi: 10.1212/WNL.0b013e3182a6cbfe et al. (2012). Saccade abnormalities in autopsy-confirmed frontotemporal Illán-Gala, I., Montal, V., Borrego-Écija, S., Vilaplana, E., Pegueroles, J., Alcolea, lobar degeneration and alzheimer disease. Arch. Neurol. 69, 509–517. D., et al. (2019). Cortical microstructure in the behavioural variant of doi: 10.1001/archneurol.2011.1021 frontotemporal dementia: looking beyond atrophy. Brain 142, 1121–1133. Brown, M. R. G., DeSouza, J. F. X., Goltz, H. C., Ford, K., Menon, R. S., Goodale, doi: 10.1093/brain/awz031 M. A., et al. (2004). Comparison of memory- and visually guided saccades using Jamadar, S. D., Fielding, J., and Egan, G. F. (2013). Quantitative meta-analysis of event-related fMRI. J. Neurophysiol. 91, 873–889. doi: 10.1152/jn.00382.2003 fMRI and PET studies reveals consistent activation in fronto-striatal-parietal Burrell, J. R., Hornberger, M., Carpenter, R. H. S., Kiernan, M. C., and Hodges, J. regions and cerebellum during antisaccades and prosaccades. Front. Psychol. R. (2012). Saccadic abnormalities in frontotemporal dementia. Neurology 78, 4:749. doi: 10.3389/fpsyg.2013.00749 1816–1823. doi: 10.1212/WNL.0b013e318258f75c Kahana Levy, N., Lavidor, M., and Vakil, E. (2018). Prosaccade and antisaccade Coiner, B., Pan, H., Bennett, M. L., Bodien, Y. G., Iyer, S., O’Neil- paradigms in persons with alzheimer’s disease: a meta-analytic review. Pirozzi, T. M., et al. (2019). Functional neuroanatomy of the human eye Neuropsychol. Rev. 28, 16–31. doi: 10.1007/s11065-017-9362-4 movement network: a review and atlas. Brain Struct. Funct. 224, 2603–2617. Kaufman, L. D., Pratt, J., Levine, B., and Black, S. E. (2010). Antisaccades: a probe doi: 10.1007/s00429-019-01932-7 into the dorsolateral prefrontal cortex in alzheimer’s disease. a critical review. J. Coppe, S., Orban de Xivry, J. J., Yuksel, D., Ivanoiu, A., and Lefevre, P. (2012). Alzheimers Dis. 19, 781–793. doi: 10.3233/JAD-2010-1275 Dramatic impairment of prediction due to frontal lobe degeneration. J. Kim, S. J., Cho, K. J., and Oh, S. (2017). Development of machine Neurophysiol. 108, 2957–2966. doi: 10.1152/jn.00582.2012 learning models for diagnosis of glaucoma. PLoS ONE 12:e0177726. Corey, K. M., Kashyap, S., Lorenzi, E., Lagoo-Deenadayalan, S. A., Heller, K., doi: 10.1371/journal.pone.0177726 Whalen, K., et al. (2018). Development and validation of machine learning Krauzlis, R. J. (2004). Recasting the smooth pursuit eye movement system. J. models to identify high-risk surgical patients using automatically curated Neurophysiol. 91, 591–603. doi: 10.1152/jn.00801.2003 electronic health record data (Pythia): a retrospective, single-site study. PLoS Krishnan, A., Williams, L. J., McIntosh, A. R., and Abdi, H. (2011). Partial least Med. 15:e1002701. doi: 10.1371/journal.pmed.1002701 squares (PLS) methods for neuroimaging: a tutorial and review. Neuroimage De Fauw, J., Ledsam, J. R., Romera-Paredes, B., Nikolov, S., Tomasev, N., Blackwell, 56, 455–475. doi: 10.1016/j.neuroimage.2010.07.034 S., et al. (2018). Clinically applicable deep learning for diagnosis and referral in McDowell, J. E., Dyckman, K. A., Austin, B. P., and Clementz, B. A. retinal disease. Nat. Med. 24, 1342–1350. doi: 10.1038/s41591-018-0107-6 (2008). Neurophysiology and neuroanatomy of reflexive and volitional Douglass, A., Walterfang, M., Velakoulis, D., and Abel, L. (2018). Behavioral saccades: evidence from studies of humans. Brain Cogn. 68, 255–270. variant frontotemporal dementia performance on a range of saccadic tasks. J. doi: 10.1016/j.bandc.2008.08.016 Alzheimers Dis. 65, 231–242. doi: 10.3233/JAD-170797 McKhann, G. M., Knopman, D. S., Chertkow, H., Hyman, B. T., Jack, C. R. Elaziz, M. A., Hosny, K. M., Salah, A., Darwish, M. M., Lu, S., and Sahlol, J., Kawas, C. H., et al. (2011). The diagnosis of dementia due to alzheimer’s A. T. (2020). New machine learning method for image-based diagnosis of disease: recommendations from the national institute on aging-alzheimer’s COVID-19. PLoS ONE 15:e0235187. doi: 10.1371/journal.pone.0235187 association workgroups on diagnostic guidelines for alzheimer’s disease. Ettinger, U., Ffytche, D. H., Kumari, V., Kathmann, N., Reuter, B., Zelaya, F., Alzheimers Dement. 7, 263–269. doi: 10.1016/j.jalz.2011.03.005 et al. (2008). Decomposing the neural correlates of antisaccade eye movements Meyniel, C., Rivaud-Péchoux, S., Damier, P., and Gaymard, B. (2005). Saccade using event-related FMRI. Cereb. Cortex 18, 1148–1159. doi: 10.1093/cercor/ impairments in patients with fronto-temporal dementia. J. Neurol. Neurosurg. bhm147 Psychiatry 76, 1581–1584. doi: 10.1136/jnnp.2004.060392 Folstein, M. F., Folstein, S. E., and McHugh, P. R. (1975). “Mini-mental state”. A Mirsky, J. B., Heuer, H. W., Jafari, A., Kramer, J. H., Schenk, A. K., Viskontas, practical method for grading the cognitive state of patients for the clinician. J. I. V., et al. (2011). Anti-saccade performance predicts executive function Psychiatr. Res. 12, 189–98. doi: 10.1016/0022-3956(75)90026-6 and brain structure in normal elders. Cogn. Behav. Neurol. 24, 50–58. Fontanella, S., Frainay, C., Murray, C. S., Simpson, A., and Custovic, doi: 10.1097/WNN.0b013e318223f6c6 A. (2018). Machine learning to identify pairwise interactions between Molitor, R. J., Ko, P. C., and Ally, B. A. (2015). Eye movements in alzheimer’s specific IgE antibodies and their association with asthma: a cross-sectional disease. J. Alzheimers Dis. 44, 1–12. doi: 10.3233/JAD-141173 analysis within a population-based birth cohort. PLoS Med. 15:e1002691. Munoz, D. P., and Everling, S. (2004). Look away: the anti-saccade task and doi: 10.1371/journal.pmed.1002691 the voluntary control of eye movement. Nat. Rev. Neurosci. 5, 218–228. Ford, K. A., Goltz, H. C., Brown, M. R., and Everling, S. (2005). Neural processes doi: 10.1038/nrn1345 associated with antisaccade task performance investigated with event-related Pa, J., Dutt, S., Mirsky, J. B., Heuer, H. W., Keselman, P., Kong, E., et al. (2014). FMRI. J. Neurophysiol. 94, 429–440. doi: 10.1152/jn.00471.2004 The functional oculomotor network and saccadic cognitive control in healthy Garbutt, S., Matlin, A., Hellmuth, J., Schenk, A. K., Johnson, J. K., elders. Neuroimage 95, 61–68. doi: 10.1016/j.neuroimage.2014.03.051 Rosen, H., et al. (2008). Oculomotor function in frontotemporal lobar Pavisic, I. M., Firth, N. C., Parsons, S., Rego, D. M., Shakespeare, T. J., Yong, degeneration, related disorders and alzheimer’s disease. Brain 131, 1268–1281. K. X. X., et al. (2017). Eyetracking metrics in young onset alzheimer’s doi: 10.1093/brain/awn047 disease: a window into cognitive visual functions. Front. Neurol. 8:377. García Cena, C. E., Andrés, D. G., and Valdeolivas, I. P. (2020). Measurement and doi: 10.3389/fneur.2017.00377 analysis of eye movements performance to predict healthy brain aging. IEEE Pena-Casanova, J., Gramunt-Fombuena, N., Quinones-Ubeda, S., Sanchez- Access 8, 87201–87213. doi: 10.1109/ACCESS.2020.2992254 Benavides, G., Aguilar, M., Badenes, D., et al. (2009a). Spanish multicenter Frontiers in Aging Neuroscience | www.frontiersin.org 19 February 2021 | Volume 12 | Article 603790 Lage et al. Oculomotor Behavior in Dementia normative studies (NEURONORMA project): norms for the rey-osterrieth Reisberg, B., Ferris, S. H., de Leon, M. J., and Crook, T. (1982). The global complex figure (copy and memory), and free and cued selective reminding test. deterioration scale for assessment of primary degenerative dementia. Am. J. Arch. Clin. Neuropsychol. 24, 371–393. doi: 10.1093/arclin/acp041 Psychiatry 139, 1136–1139. doi: 10.1176/ajp.139.9.1136 Pena-Casanova, J., Quinones-Ubeda, S., Quintana-Aparicio, M., Aguilar, M., Shakespeare, T. J., Kaski, D., Yong, K. X. X., Paterson, R. W., Slattery, C. F., Badenes, D., Molinuevo, J. L., et al. (2009c). Spanish multicenter normative Ryan, N. S., et al. (2015). Abnormalities of fixation, saccade and pursuit studies (NEURONORMA project): norms for verbal span, visuospatial span, in posterior cortical atrophy. Brain 138, 1976–1991. doi: 10.1093/brain/ letter and number sequencing, trail making test, and symbol digit modalities awv103 test. Arch. Clin. Neuropsychol. 24, 321–341. doi: 10.1093/arclin/acp038 Stemmer, A., Shadmi, R., Bregman-Amitai, O., Chettrit, D., Blagev, D., Orlovsky, Pena-Casanova, J., Quintana-Aparicio, M., Quinones-Ubeda, S., Aguilar, M., M., et al. (2020). Using machine learning algorithms to review computed Molinuevo, J. L., Serradell, M., et al. (2009b). Spanish multicenter normative tomography scans and assess risk for cardiovascular disease: retrospective studies (NEURONORMA project): norms for the visual object and space analysis from the national lung screening trial (NLST). PLoS ONE 15:e0236021. perception battery-abbreviated, and judgment of line orientation. Arch. Clin. doi: 10.1371/journal.pone.0236021 Neuropsychol. 24, 355–370. doi: 10.1093/arclin/acp040 Thier, P., and Ilg, U. J. (2005). The neural basis of smooth-pursuit eye movements. Pena-Casanova, J. (2005). “Test normalidad, semiología y patología. Imitación Curr. Opin. Neurobiol. 15, 645–652. doi: 10.1016/j.conb.2005.10.013 de posturas”, in Test Barcelona Revisado, ed J. Pena-Casanova (Barcelona: Viscaino, M., Maass, J. C., Delano, P. H., Torrente, M., Stott, C., and Normalidad, semiología y patología neuropsicológicas. Masson), 159–160. Auat Cheein, F. (2020). Computer-aided diagnosis of external and middle Petit, L., and Haxby, J. V. (1999). Functional anatomy of pursuit eye ear conditions: a machine learning approach. PLoS ONE 15:e0229226. movements in humans as revealed by fMRI. J. Neurophysiol. 82, 463–471. doi: 10.1371/journal.pone.0229226 doi: 10.1152/jn.1999.82.1.463 Pierrot-Deseilligny, C., Müri, R. M., Ploner, C. J., Gaymard, B., Demeret, S., and Conflict of Interest: The authors declare that the research was conducted in the Rivaud-Pechoux, S. (2003). Decisional role of the dorsolateral prefrontal cortex absence of any commercial or financial relationships that could be construed as a in ocular motor behaviour. Brain 126, 1460–1473. doi: 10.1093/brain/awg148 potential conflict of interest. Pierrot-Deseilligny, C., Rivaud, S., Gaymard, B., and Agid, Y. (1991). Cortical control of reflexive visually-guided saccades. Brain 114, 1473–1485. Copyright © 2021 Lage, López-García, Bejanin, Kazimierczak, Aracil-Bolaños, doi: 10.1093/brain/114.3.1473 Calvo-Córdoba, Pozueta, García-Martínez, Fernández-Rodríguez, Bravo-González, Rajpurkar, P., Irvin, J., Ball, R. L., Zhu, K., Yang, B., Mehta, H., et al. (2018). Jiménez-Bonilla, Banzo, Irure-Ventura, Pegueroles, Illán-Gala, Fortea, Rodríguez- Deep learning for chest radiograph diagnosis: a retrospective comparison of Rodríguez, Lleó-Bisa, García-Cena and Sánchez-Juan. This is an open-access article the CheXNeXt algorithm to practicing radiologists. PLoS Med. 15:e1002686. distributed under the terms of the Creative Commons Attribution License (CC BY). doi: 10.1371/journal.pmed.1002686 The use, distribution or reproduction in other forums is permitted, provided the Rascovsky, K., Hodges, J. R., Knopman, D., Mendez, M. F., Kramer, J. H., original author(s) and the copyright owner(s) are credited and that the original Neuhaus, J., et al. (2011). Sensitivity of revised diagnostic criteria for publication in this journal is cited, in accordance with accepted academic practice. the behavioural variant of frontotemporal dementia. Brain 134, 2456–2477. No use, distribution or reproduction is permitted which does not comply with these doi: 10.1093/brain/awr179 terms. Frontiers in Aging Neuroscience | www.frontiersin.org 20 February 2021 | Volume 12 | Article 603790