Cognitive correlates of antisaccade behaviour across multiple neurodegenerative diseases

Abstract Oculomotor tasks generate a potential wealth of behavioural biomarkers for neurodegenerative diseases. Overlap between oculomotor and disease-impaired circuitry reveals the location and severity of disease processes via saccade parameters measured from eye movement tasks such as prosaccade and antisaccade. Existing studies typically examine few saccade parameters in single diseases, using multiple separate neuropsychological test scores to relate oculomotor behaviour to cognition; however, this approach produces inconsistent, ungeneralizable results and fails to consider the cognitive heterogeneity of these diseases. Comprehensive cognitive assessment and direct inter-disease comparison are crucial to accurately reveal potential saccade biomarkers. We remediate these issues by characterizing 12 behavioural parameters, selected to robustly describe saccade behaviour, derived from an interleaved prosaccade and antisaccade task in a large cross-sectional data set comprising five disease cohorts (Alzheimer’s disease/mild cognitive impairment, amyotrophic lateral sclerosis, frontotemporal dementia, Parkinson’s disease, and cerebrovascular disease; n = 391, age 40–87) and healthy controls (n = 149, age 42–87). These participants additionally completed an extensive neuropsychological test battery. We further subdivided each cohort by diagnostic subgroup (for Alzheimer’s disease/mild cognitive impairment and frontotemporal dementia) or degree of cognitive impairment based on neuropsychological testing (all other cohorts). We sought to understand links between oculomotor parameters, their relationships to robust cognitive measures, and their alterations in disease. We performed a factor analysis evaluating interrelationships among the 12 oculomotor parameters and examined correlations of the four resultant factors to five neuropsychology-based cognitive domain scores. We then compared behaviour between the abovementioned disease subgroups and controls at the individual parameter level. We theorized that each underlying factor measured the integrity of a distinct task-relevant brain process. Notably, Factor 3 (voluntary saccade generation) and Factor 1 (task disengagements) significantly correlated with attention/working memory and executive function scores. Factor 3 also correlated with memory and visuospatial function scores. Factor 2 (pre-emptive global inhibition) correlated only with attention/working memory scores, and Factor 4 (saccade metrics) correlated with no cognitive domain scores. Impairment on several mostly antisaccade-related individual parameters scaled with cognitive impairment across disease cohorts, while few subgroups differed from controls on prosaccade parameters. The interleaved prosaccade and antisaccade task detects cognitive impairment, and subsets of parameters likely index disparate underlying processes related to different cognitive domains. This suggests that the task represents a sensitive paradigm that can simultaneously evaluate a variety of clinically relevant cognitive constructs in neurodegenerative and cerebrovascular diseases and could be developed into a screening tool applicable to multiple diagnoses.


Cohort diagnostic criteria
The specific diagnostic criteria used for each ONDRI cohort was as follows: the National Institute on Aging-Alzheimer's Association criteria for probable Alzheimer's disease dementia or amnestic MCI 1,2 ; possible, probable, or definite ALS based on El Escorial criteria 3 ; subtype criteria for behavioural variant frontotemporal dementia (bvFTD), 4 progressive supranuclear palsy (PSP), 5 corticobasal syndrome (CBS), 6 or primary progressive aphasia (PPA) semantic variant or nonfluent variant 7 ; or United Kingdom Brain Bank criteria for idiopathic Parkinson's disease. 8 Participants enrolled in the CVD cohort had MRI-or CT-confirmed ischemic stroke at least three months prior to recruitment, with or without cognitive impairment. All participants were required to have a Montreal Cognitive Assessment (MoCA) score ≥18 (≥14 for participants with atypical Alzheimer's disease or FTD) 9 ; note that the Parkinson's disease and CVD cohorts intentionally included participants both with and without cognitive impairment based on MoCA score.

Eye movement recording and task paradigm
Participants were seated in a dark room approximately 60cm away from a 17-inch computer monitor with 1280x1024 pixel resolution. Monocular eye position was tracked using an infrared video-based eye tracker (EyeLink 1000 Plus; SR Research Ltd., Ottawa, ON, Canada) at a sampling rate of 500 Hz. A nine-point array calibration was performed for each participant prior to task initiation whenever possible, with occasional use of a five-point array if the nine-point array proved unsuccessful.
All participants completed an interleaved pro-and anti-saccade task (IPAST) 10 consisting of two blocks of 120 trials ( Supplementary Fig. 2). Following an intertrial interval lasting 1000ms, each trial began with the appearance of a central fixation point (0.5⁰ diameter; 42 cd/m 2 ) lasting 1000ms and displayed on a black background (0.1 cd/m 2 ). The fixation point could be one of two luminance-matched colours that each indicated a different task instruction (green = prosaccade, red = antisaccade). The fixation point then disappeared for a gap period of 200ms, during which the screen remained empty, followed by the appearance of a peripheral stimulus (0.5⁰ diameter; 42 cd/m 2 ) at 10⁰ horizontally to either the left or right of the fixation position. On prosaccade trials, participants were instructed to make a saccade to the stimulus location as quickly as possible; on antisaccade trials, they were instructed not to look at the stimulus and instead to look in the opposite direction from where it appeared. Saccades in the opposite direction from the stimulus during prosaccade trials (which occur infrequently) and saccades towards the stimulus during antisaccade trials (which occur frequently) were considered direction errors.
Supplementary Figure 2. Interleaved pro-and anti-saccade task (IPAST) visual display and saccade latency classification. X-axis shows time in milliseconds relative to peripheral stimulus appearance. Stimuli displayed on the screen are indicated in the top rows: fixation point (-1200 ms to -200 ms; green=prosaccade trial; red=antisaccade trial), gap period without visual stimulus (-200 ms to 0 ms), and peripheral stimulus (0 ms onward; green arrow indicates correct prosaccade, red arrow indicates correct antisaccade). Latency windows for saccade classification are indicated by grey boxes and corresponding labels in the bottom row: task disengagements (occur -1100 ms to -111 ms), anticipatory saccades (occur -110 ms to 89 ms), express latency saccades (occur 90 ms-139 ms), and regular latency saccades (occur 140 ms-800 ms). Viable saccades encompass both express and regular latency saccades.

Cognitive domain scores
As described elsewhere, 11,12 cognitive domain scores were calculated for all ONDRI participants using 23 test scores from the neuropsychology battery 13 for each of five cognitive domains: attention/working memory, executive function, language, memory, and visuospatial function. Standardized residuals were computed for each raw test score using a linear model with age, sex, and years of education as independent variables. Test version was also included where applicable (Symbol Digit Modality Test written vs oral version; Delis-Kaplan Executive Function System -Verbal Fluency standard vs alternate version). Timed scores were inverted so that higher scores signified better performance. Residuals for each test were grouped according to cognitive domain based on established neuropsychological convention 14,15 and consensus among ONDRI neuropsychologists (Supplementary Table 1); residuals were then averaged within each domain to produce a single score per participant per domain. Note that although most participants completed the Brief Visuospatial Memory Test -Revised (BVMT-R), including timed recall trials that can be used to measure non-verbal memory, it was not used in the calculation of memory scores as it was not completed by the ALS cohort. This ensured that memory scores were calculated identically across all cohorts; note therefore that the memory score represents verbal memory only.

Cognitive status
Cognitive status was used to subdivide ALS, Parkinson's disease, and CVD into one of four categories: cognitively normal, mild cognitive impairment (MCI), dementia, or other (individuals who did not clearly meet criteria for another category). This classification was based on three metrics: 1) cognitive performance on the neuropsychology test battery; 2) subjective cognitive decline (self-or study partner-report) as assessed by the Short Informant Questionnaire on Cognitive Decline in the Elderly (Short IQ-CODE) 16 ; and 3) activities of daily living (ADLs) assessed using an instrumental ADLs questionnaire adapted from Lawton et al. 17 and completed by participants' study partners.
Tests comprising the neuropsychology battery were divided into five cognitive domains (attention/working memory, executive function, language, memory, visuospatial function) as shown in Supplementary Table 1. Each individual test score was considered abnormal if the participant scored ≥1.5SD below age-and sex-corrected norms. In the ALS and CVD cohorts, the overall neuropsychology battery was considered impaired if the participant had at least 2 abnormal scores within a single domain. In the Parkinson's disease cohort, the requirement for multiple domains was dropped and the overall neuropsychology battery was considered impaired if the participant had at least 2 abnormal scores regardless of cognitive domain, in accordance with established criteria. 15 Subjective cognitive decline was considered present if participants or their study partners reported cognitive decline (responding either "a bit worse" or "much worse") on at least one of the sixteen everyday cognitive tasks probed by Short IQ-CODE (e.g. "Compared to 10 years ago, how are you/the participant at recalling conversations a few days later?"). Subjective cognitive decline was not considered present if all responses reported no change or improvement.
ADLs were assessed using a questionnaire on which participants' study partners were asked to rate the participants' capacity for independent function in each of eight types of daily tasks (using the telephone, doing laundry, shopping, using transportation, preparing food, managing medication, home maintenance, and finances). ADLs were considered intact or minimally impacted if study partners provided only "A" or "B" responses (corresponding to full independence or limited support; e.g., "does personal laundry completely" and "launders small items, rinses socks, etc." respectively) with no "C", "D", or "E" responses (corresponding to increasing degrees of required support from others).
Participants were considered cognitively normal if they demonstrated normal cognition on the test battery and had intact ADLs, regardless of subjective cognitive decline. Participants who demonstrated cognitive impairment on the test battery and reported a subjective decline in cognition, but whose ADLs remained intact or minimally impacted, were classified as MCI; those who showed cognitive impairment on the test battery across multiple cognitive domains (abnormal scores in at least two different domains), endorsed subjective cognitive decline, and had impaired ADLs were considered to have dementia. Participants who did not meet criteria for any of these categories were classified as other. In the Parkinson's disease cohort, cognitive impairment was considered present on the neuropsychology battery if the participant scored ≥1.5SD below education-and/or age-corrected norms on at least two measures, according to the recommendations of the Movement Disorder Society Task Force for classifying MCI in Parkinson's disease. 15 In ALS and CVD, cognitive impairment was considered present if the participant scored ≥1.5SD below norms on at least two measures within the same cognitive domain (Supplementary Table 1), as in previous work 18 . Additionally, impairment on any of multiple interrelated measures were considered only one impairment within a domain. Note that although the BVMT-R recall trials were not used in the calculation of memory domain scores, it was included where possible (i.e. in the Parkinson's disease and CVD cohorts) in evaluation of overall cognitive status.
Ultimately, this classification resulted in four subgroups within each of the ALS, Parkinson's disease, and CVD cohorts: cognitively normal (CN), MCI, dementia (D), and other. However, due to the small size of the ALS-MCI and ALS-D subgroups, we combined these into a single subgroup called ALS-CI (cognitively impaired). We note that this may have excluded some ALS participants who were classified as Other because they demonstrated impaired iADLs, which could have been due to disease-related physical rather than cognitive impairment, despite their normal cognitive performance on the neuropsychology battery.
Missing neuropsychology data was imputed with the worst possible score if missing because the participant was too impaired to complete the task (41 scores total across all tests and participants), or with the variable mean if missing for non-disease-related reasons (133 scores total across all tests and participants).
Although different classification methods were used to subdivide each cohort by cognitive status, we only performed statistical comparisons for each subgroup to the control cohort and to the other subgroups within the same cohort. No subgroups generated using different classification methods were directly compared.

Saccade classification
Saccade data was preprocessed 19 and saccades were then categorized by an auto-marking script written in MATLAB (The MathWorks, Inc., Natick, MA, USA). We excluded trials with poor data quality or significant behavioural aberrations as described in the main text. Saccades were subsequently classified based on saccadic reaction time (SRT) (Supplementary Fig. 2) (time between peripheral stimulus appearance and saccade initiation).
A minimum of 90ms is required for a visual signal to propagate through the oculomotor system and trigger a saccade. 20 Therefore, any saccades occurring 90-800ms post-stimulus appearance were considered viable task-relevant responses made with perception of the stimulus location. Viable saccades were further subdivided based on latency and correctness. Saccades occurring 90ms to 139ms after stimulus appearance were classified as express latency saccades, and those occurring 140ms to 800ms after stimulus appearance were classified as regular latency saccades. Saccades made in a direction consistent with task instruction on a given trial were considered correct; those made in the opposite direction were considered direction errors. Only the first saccade initiated was considered for analysis; any subsequent corrections were not analyzed. Saccades with reaction time >800ms occur rarely and were therefore considered outliers and excluded from analysis.
Due to the abovementioned 90ms delay, the participant perceives the screen to be empty from 110ms before to 89ms after stimulus appearance, so any saccades initiated during this period were equally likely to be correct or incorrect and were considered indicative of guessing behaviour. These saccades were classified as anticipatory saccades.
Saccades were considered task disengagements if the participant looked away from the fixation point while they perceived it on the screen (1110ms before to 111ms before stimulus appearance).
Reaction times of initial prosaccades and antisaccades were also measured, as well as peak velocity (degrees/s), and amplitude (distance between the starting point of the initial saccade and its endpoint in degrees) for initial viable correct prosaccades only.

Saccade parameter calculation
We focused on 12 specific parameters that captured behaviour at different times across the tasks. Also see Supplementary Table 2.
To characterize behaviour early in the task, we identified trials in which participants looked away from the fixation point and never returned, as a measure of task disengagement. The percentage of prosaccade task disengagements was determined by calculating the percentage of prosaccade task disengagements out of all prosaccade trials completed excluding trials with technical issues (e.g., lost tracking, bad calibration). Percentage of antisaccade task disengagements was determined identically but using antisaccade trials only.
To quantify and characterize guessing behaviour occurring during the brief interval following fixation disappearance and prior to stimulus presentation, when the participant perceives a blank screen, we measured the percentage of anticipatory saccades. The percentage of anticipatory prosaccades was determined by calculating the percentage of anticipatory prosaccades out of all prosaccade trials completed excluding trials with technical issues; percentage of anticipatory antisaccades was determined identically using antisaccade trials instead.
Percentage of express latency correct prosaccades was determined by calculating the percentage of express latency correct prosaccades out of all viable prosaccades only. This parameter was chosen to understand participants' prosaccade behaviour during the early portion of the stimulus presentation epoch and because it depends on excitability levels in oculomotor circuits at the time of stimulus appearance. 21 Percentages of express latency antisaccade errors and regular latency antisaccade errors were determined by calculating the percentage of express latency antisaccade errors and regular latency antisaccade errors respectively only out of all viable antisaccade trials. By separating these parameters, we were able to characterize error behaviours occurring at both short and longer latencies during the stimulus presentation epoch.
Mean prosaccade SRT and antisaccade SRT for each participant were determined based only on viable correct prosaccades or antisaccades respectively. Participants were also required to have at least five viable correct prosaccades or antisaccades for this measure to be calculated to ensure representative values; if they did not meet this criterion, the parameter was not computed for that participant. Mean prosaccade velocity and mean prosaccade amplitude were also calculated based only on viable correct prosaccades and required five viable correct prosaccades. We did not quantify the velocity and amplitude of antisaccades because there was considerable variability in antisaccade endpoints among participants, who were not instructed to make accurate antisaccades but only to look to the side opposite the stimulus. These parameters enabled us to characterize the overall speed of participants' task-appropriate prosaccade and antisaccade responses.
VOT was calculated by first generating a curve for each participant displaying the cumulative percentage of correct antisaccades and a second curve for percentage of antisaccade direction errors, then subtracting the error curve from the correct curve. This produces a characteristic curve whose nadir indicates the time during antisaccade trials after which voluntary processes begin to outcompete automated visually driven signals to produce correct antisaccades. 10,22 VOT was determined as the time in milliseconds after stimulus appearance at which this nadir occurred. It was restricted from 90-300ms following stimulus appearance and required at least 10 antisaccades at any latency. This parameter was selected to understand when during the trial participants' voluntary motor programs were able to outcompete automated ones.

Quality control and group analysis exclusion criteria
Of the 520 recruited ONDRI participants, 485 successfully completed eye tracking assessment ( Supplementary Fig. 3A, B). Additional quality control criteria were then applied across all groups to remove participants with low-quality eye tracking data from all subsequent group and factor analyses. All participants were required to have completed at least 120 total IPAST trials ( Supplementary Fig. 3C), of which at least 80% were required to have appropriate calibration and no loss of tracking ( Supplementary Fig. 3D). A minimum of five viable prosaccades and five viable antisaccades were also required to mitigate the effects of potentially unrepresentative or noisy parameter values derived from a very small number of instances-for example, an antisaccade SRT value for a participant who made only one viable correct antisaccade. These criteria eliminated 35 ONDRI participants and no control participants, resulting in 450 ONDRI and 149 control participants ( Supplementary Fig. 3E).
We included all ONDRI participants in the factor analysis, regardless of subtype or cognitive status, but required that they have no missing data. Practically, this excluded only participants who made too few viable correct antisaccades to accurately estimate their antisaccade SRT and resulted in 391 ONDRI participants included in the final factor analysis ( Supplementary Fig. 3F). Control participants were excluded since they did not complete neuropsychological assessment and could not be included in subsequent analyses investigating relationships between factor scores and neuropsychology data. Demographic characteristics of participants included in factor analysis are displayed in Table 1 and Supplementary Fig. 1.
We removed very small and/or poorly defined subgroups (atypical Alzheimer's disease, progressive nonfluent aphasia, semantic dementia, corticobasal syndrome, participants with cognitive status "other", and one participant with both bvFTD and PSP) from parameter-level analysis. This resulted in 344 ONDRI and 149 control participants included.

Supplementary Figure 3. Number of ONDRI participants by cohort and subgroup remaining after each exclusion criterion. (A) All participants initially recruited by ONDRI; (B)
All participants who successfully completed eye tracking; (C) All participants with at least 120 completed trials; (D) All participants with at least 80% of trials having appropriate tracking (i.e. no tracking loss or inappropriate calibration); (E) All participants who made at least 5 viable prosaccades and 5 viable antisaccades; (F) All participants included in factor analysis; (G) All participants included in parameter-level analysis. MCI: mild cognitive impairment. AD: Alzheimer's disease. ALS: amyotrophic lateral sclerosis. bvFTD: behavioural variant frontotemporal dementia. PSP: progressive supranuclear palsy. CBS: corticobasal syndrome. PNFA: progressive nonfluent aphasia. SD: semantic dementia. PD: Parkinson's disease. CVD: cerebrovascular disease. CN: cognitively normal. CI: cognitively impaired. D: dementia. Other: cognitive status unclear.

Supplementary Results
Factor analysis structure matrix Supplementary

Cumulative distributions of saccades
Cumulative distributions of saccadic reaction times were constructed from the raw data (i.e. uncorrected for age and sex) to summarize the behaviour of subgroups within each cohort ( Supplementary Fig. 4) and qualitatively evaluated for differences between groups. Visual inspection indicated that subgroups largely did not differ in behaviour on prosaccade trials within or across cohorts. On antisaccade trials, however, more cognitively impaired subgroups displayed worse performance relative to less cognitively impaired subgroups, as indicated by consistently shallower slopes in the correct curve (i.e., slower correct antisaccade reaction time) and overall lower correct and higher error curves (i.e., higher proportion of antisaccade direction errors). Disease subgroups, particularly those with cognitive impairment, displayed generally worse performance than controls. We additionally computed cumulative distributions for the control data based on a median split by age ( Supplementary Fig. 4F), which indicated that even the older group of controls demonstrated better antisaccade performance than patient groups.

Additional parameter-level results
Anticipatory saccades occur at latencies when the participant perceives a blank screen and are equally likely to occur in either direction; they are therefore considered indicative of impulsive guessing behaviour. The percentage of anticipatory prosaccades was significantly different across groups (H(12)=32.94, P<0.001) with the MCI (z=3.28, P=0.021) and CVD-MCI (z=3.53, P=0.0087) groups making significantly more anticipatory prosaccades than controls (Supplementary Fig. 5A). A Kruskal-Wallis test indicated significant between-group differences in anticipatory antisaccades (H(12)=24.68, P=0.016); however, no post hoc tests were significant (Supplementary Fig. 5B).
We tested for between-group differences in express latency correct prosaccades. Prosaccades at this latency typically reflect the integrity of visuomotor circuitry that generates a visually-driven saccade. There were no differences between groups in express latency correct prosaccades (H(12)=17.92, P=0.12) (Supplementary Fig. 5E).

Disease-specific findings
Many existing studies have examined various antisaccade parameters in a variety of neurodegenerative and cerebrovascular conditions. We situate the results from this study in context of prior disease-specific analyses.

Alzheimer's disease and mild cognitive impairment
Most studies have focused on Alzheimer's disease alone, with a minority also including an MCI comparison group and a smaller minority examining MCI only. The most reported finding is increased error rates in disease groups relative to controls, which we replicated in Alzheimer's disease [23][24][25][26][27][28][29][30][31][32][33][34][35][36][37] and MCI. 32,33,35,38 However, we did not corroborate the finding in some studies that Alzheimer's disease participants generated more errors than MCI participants 39,40 or that MCI participants did not differ from controls. 39,41 Together, these results suggest that antisaccade error rate robustly detects Alzheimer's disease/MCI but does not differentiate between prodromal and full-blown states; clinical overlap between and heterogeneity within these subgroups may complicate efforts to develop behavioural biomarkers that separate them by appraising underlying neural dysfunction. Some recent work involving both amnestic and nonamnestic MCI participants indicates that amnestic MCI participants have similar error rates to Alzheimer's disease patients and make more errors than nonamnestic MCI participants 36,42 ; in line with this, our study included only amnestic MCI and found no differences between them and Alzheimer's disease participants. Future work should further probe this distinction.

Amyotrophic lateral sclerosis
Neither cognitively normal nor cognitively impaired ALS participants in our study displayed any differences from controls. This is particularly in opposition to previous studies of ALS that have typically found increased antisaccade errors in ALS. However, studies with larger sample sizes have more consistently found higher error rates in ALS 51-54 than those with small sample sizes. 55,56 The lack of significance in this study may therefore be due to the comparatively small size of our ALS subgroups. We also did not replicate a few previous findings of slowed antisaccade reaction time, 51-53 although this is not universal. 57 Clinical heterogeneity may also play a role but was considered outside the scope of this study.

Frontotemporal dementia
We did not divide the FTD cohort into cognition-based subgroups, but observed results in line with previous work finding increased antisaccade errors relative to controls in PSP. 28,29,[58][59][60][61] Our results regarding saccadic reaction time in PSP were in line with studies indicating slowed antisaccade rection time 60,61 although some studies found no difference 28,62 ; and no difference in prosaccade reaction time 28,29,58-62 with one exception. 63 We also replicated a very consistent finding of reduced saccade amplitude, but our results did not support reports of reduced saccade velocity, 28,29,59,62,64,65 although this may relate to statistical power issues as the median saccade velocity appeared much lower in PSP.
We did not observe an increase in errors in the bvFTD subgroup, in opposition to existing literature indicating more errors in bvFTD alone or a pooled FTD group relative to controls. [27][28][29]60,66 We found no differences between bvFTD and controls in prosaccade and antisaccade reaction time, saccade amplitude, or saccade velocity.

Parkinson's disease
Many studies have described antisaccade behaviour in Parkinson's disease. Most of these find increased error rates in Parkinson's disease compared to controls 61,67-78 which has also been confirmed by meta-analysis. 79 However, our finding that only PD-D had increased regular latency errors relative to controls, PD-CN, and PD-MCI indicates this relationship may be related to cognitive impairment, similar to Mosimann et al. 26 Medication effects 80,81 and disease severity 82 may also modulate error rates, which we did not control for, although note that participants classified as PD-D are likely to have longer disease course and may therefore have more severe disease.
Antisaccade reaction time is sometimes reported as being slowed in Parkinson's disease. A recent meta-analysis indicated increased antisaccade reaction time relative to controls, 79 which our results support. We also corroborate suggestions that cognitive impairment may modulate this relationship. 26 However, we did not find any slowing of prosaccade reaction time in Parkinson's disease, nor did we replicate findings that cognitive impairment may slow this metric further. 26,83 Studies reporting express latency prosaccades typically find increases in Parkinson's disease, 69,77,84 which we did not replicate. However, there are some indications that dopaminergic medication may slow prosaccades 80,85 ; note that ONDRI Parkinson's disease participants were all medicated. Finally, our study is in line with previous work showing that saccade amplitude is reduced in Parkinson's disease.

Cerebrovascular disease
Although only CVD-MCI demonstrated significantly more regular latency antisaccade errors than controls, there was a general qualitative trend towards increased errors with increased cognitive impairment. This result is novel due to the relative dearth of cohort studies of antisaccades in CVD. Existing cohort studies 86,87 and studies of localized ischemic lesions 88,89 collectively suggest that stroke location and size may be paramount in determining pro-and antisaccade behaviour in this group.

Voluntary override time and overall cognitive impairment
We additionally report the novel parameter VOT, which denotes the time taken for voluntary processes to begin overcoming automated processes during antisaccades. 10 As with regular latency antisaccade errors, cognitively impaired subgroups generally demonstrated poorer performance. In general, therefore, cognitive impairment probably slows initiation of voluntary processes and generates inhibitory control deficits.

Limitations & future directions
Due to the comprehensive nature of the study, we assessed several cohorts and subgroups of varying size, unequal variance, and clinical heterogeneity across numerous behavioural parameters. To offset the large number of between-group comparisons required by this design and limit type 1 errors, we employed Kruskal-Wallis tests, which offer less statistical power than typical parametric ANOVAs on normally distributed variables, and post hoc Dunn tests with a relatively conservative Holm-Bonferroni correction. Additionally, some cohorts and component subgroups (e.g. ALS, FTD) contained few participants. Therefore, the possibility of type II errors is nontrivial. Our findings, particularly those on infrequently reported IPAST parameters, should therefore be validated by cohort-or subgroup-specific studies in future.
Secondly, we did not attempt to control for clinical variables such as disease severity, medication type and dosage, or other disease-related parameters. However, most patients were taking medication and there is evidence that drugs used to manage neurodegenerative disease symptoms, especially dopaminergic anti-Parkinsonian medications, may affect oculomotor behaviour. 80,85 Exploration of the oculomotor effects of medication type, dosage, and other pertinent clinical parameters should be a target of future study.
This study is among a very small number to report pro-and antisaccade behaviour in CVD or related diseases. Behaviour in this cohort is difficult to typify at the group level; as demonstrated by lesion studies, stroke location determines the resultant behavioural deficits. We considered examination of IPAST behaviour according to stroke location to be beyond the scope of this study. Accordingly, the CVD cohort is best considered an illustration of our findings about cognitive impairment rather than as a uniform disease cohort. Since we did not exclude those with infarcts in visual or oculomotor pathways, it is also possible that some members of this cohort performed poorly due to insufficient visual acuity or oculomotor ability. However, CVD collectively did not display differences from controls on prosaccade measures, indicating that visuomotor circuitry was intact for most participants. Stroke location should be considered in future investigation of this cohort.
Finally, although only baseline (first visit) data was reported in the current study, ONDRI participants completed the same assessments at follow-up visits for up to three years. Evaluation of cognitive decline over the time course of the study, and its relationship to IPAST parameters, should be completed as a supplement to the results described here and to understand how oculomotor behaviour may change with disease progression.