Attenuated pupillary response during visual search in preclinical Alzheimer's disease

Abstract

INTRODUCTION

The locus coeruleus is among the earliest brain regions affected by tau pathology in Alzheimer's disease (AD), but the functional impact is difficult to measure in vivo. Task‐evoked pupil dilation provides an index of locus coeruleus–norepinephrine function that might be altered in cognitively normal older adults with underlying AD pathology.

METHODS

Cognitively normal older adults identified as AD biomarker positive (N = 25) or negative (N = 36) based on plasma phosphorylated tau (p‐tau) levels completed a conjunctive visual search task that manipulated attentional load by varying set size. Pupil dilation responses during the task were analyzed using mixed‐effects models and time‐resolved regression.

RESULTS

Despite comparable accuracy and reaction times, biomarker‐positive adults showed reduced load‐dependent modulation of pupil dilation during target‐present trials. Weaker modulation was associated with higher p‐tau levels and poorer executive and memory function.

DISCUSSION

Attenuated task‐evoked modulation of pupil dilation during visual search reveals locus coeruleus–norepinephrine dysfunction in preclinical AD.

Highlights

• Pupil dilation during visual conjunctive search showed sensitivity to cognitive load and target detection suggesting locus coeruleus (LC)‐related neuromodulatory engagement.

• Attentional load‐dependent pupil modulation was attenuated in Alzheimer's disease (AD) biomarker‐positive older adults despite intact behavioral performance.

• Attenuation of load‐dependent pupil modulation emerged during the peak arousal window consistent with reduced phasic LC–norepinephrine responsivity.

• Greater attenuation of load‐dependent pupil modulation was associated with higher plasma phosphorylated tau217 and poorer inhibitory control and visual memory.

• Task‐evoked pupillometry may be a sensitive behavioral marker of preclinical AD‐related neuromodulatory dysfunction.

1. BACKGROUND

The noradrenergic locus coeruleus (LC) neuromodulatory system is increasingly recognized as central to the earliest biological and behavioral manifestations of Alzheimer's disease (AD). Hyperphosphorylated tau (p‐tau) accumulates within LC neurons before measurable cortical tau or amyloid deposition, and decades before AD symptom onset. ¹ , ² , ³ Findings from post mortem studies, ¹ , ² , ⁴ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ tau positron emission tomography (PET) ⁷ , ⁹ , ¹² , ¹³ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ and neuromelanin‐sensitive magnetic resonance imaging (MRI) ⁷ , ⁸ , ⁹ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁹ , ²⁰ , ²¹ , ²² converge on the LC as the earliest site of AD pathology, possibly due to its high metabolic burden, extensive axonal arborization, and persistent activity across sleep–wake cycles. Understanding changes in the functional integrity of the LC in the earliest stages of AD is therefore critical for elucidating early pathogenesis and development of reliable physiological markers of the disease.

Neuromodulation by the noradrenergic LC system is essential for regulation of neural gain that sharpens attention, arousal, the stress response, and flexible, adaptive behavior ²³ , ²⁴ , ²⁵ , ²⁶ , ²⁷ , ²⁸ Its early involvement in AD suggests that sensitive measures of change in LC activity may signal emerging AD pathology before usual clinical symptoms appear. This may manifest as subtle alterations in neural activity and attentional function in cognitively normal older adults who have preclinical AD. If so, characterizing these alterations could provide information on how neuromodulatory dysfunction contributes to early, subtle cognitive and behavioral decline prior to overt cognitive impairment.

While direct recording of LC activity is difficult, pupil dilation is a well‐established and easily obtained measure of LC function that has been used to study its role in a wide range of cognitive tasks. Task‐evoked pupil dilation has been shown to closely track phasic LC activity, ²⁹ , ³⁰ , ³¹ , ³² reflecting changes in cognitive effort, uncertainty, feature‐integration demands, and decision‐related arousal. ³³ , ³⁴ , ³⁵ , ³⁶ , ³⁷ , ³⁸ , ³⁹ , ⁴⁰ , ⁴¹ , ⁴² , ⁴³ , ⁴⁴ , ⁴⁵ , ⁴⁶ , ⁴⁷ , ⁴⁸ , ⁴⁹ A growing body of evidence suggests that pupil‐based measures indexing LC function are also sensitive to AD risk. ⁴² , ⁵⁰ , ⁵¹ , ⁵² , ⁵³ , ⁵⁴ , ⁵⁵ , ⁵⁶ , ⁵⁷ Various studies have shown alterations in tonic pupil size and abnormal patterns of task‐evoked pupil dilation in individuals with mild cognitive impairment, ⁴² , ⁵¹ , ⁵³ , ⁵⁴ , ⁵⁶ apolipoprotein E (APOE) ε4 carriers, ⁵⁰ and older adults with higher AD polygenic risk scores. ⁵² Results have varied with some studies showing exaggerated pupil responses that are often interpreted as an indication of LC‐mediated compensatory hyperactivation, ⁴⁶ , ⁵¹ , ⁵⁸ and others showing reduced responses that may be attributable to inefficient neuromodulatory signaling. ⁴² , ⁵¹ , ⁵⁵ , ⁵⁸ These divergent outcomes may vary as a function of task demands, analytic approaches, and disease stage.

Conjunctive visual search tasks offer a powerful paradigm for examining pupil‐linked changes in LC dynamics in individuals with preclinical AD. In these tasks, individuals must detect a target by integrating multiple sensory features that are shared by the target and competing distractor objects. Cognitive demands of conjunctive visual search include sustained and selective attention, feature binding, distractor suppression, and rapid decision making. Prior work has shown that pupil dilation increases with distractor set size and perceptual load, ³³ , ³⁸ , ³⁹ , ⁴¹ , ⁴³ , ⁴⁵ reflecting greater demands on perceptual processing and attentional selection, and is reliably larger under conditions that require active, goal‐directed target discrimination compared to passive viewing or task‐irrelevant cuing conditions. ³¹ , ³⁶ , ³⁷ , ³⁸ , ⁴⁴ , ⁴⁹ , ⁵³ These findings are consistent with heightened LC‐mediated arousal during evidence accumulation and target detection. ²³ , ²⁴ , ²⁵ , ³¹ , ⁴⁷

The rapid rise and sustained peak of the pupil response during visual search also aligns with known LC phasic dynamics, ²³ , ²⁷ , ²⁹ , ³⁰ , ³¹ , ³² providing an opportunity to examine how AD‐related pathology may alter both the magnitude and temporal evolution of the LC response. Models propose that early tau pathology reduces phasic LC responsivity, diminishing the precision with which arousal scales with cognitive load. ²⁶ , ⁵⁹ Thus, impaired load‐dependent modulation of pupil dilation can possibly serve as an early functional signature of LC dysfunction even when overt cognitive performance remains intact.

The present study examined this possibility by assessing pupil dynamics during conjunctive visual search in cognitively normal older adults classified as either biomarker positive or biomarker negative using plasma p‐tau assays. ⁶⁰ , ⁶¹ , ⁶² , ⁶³ Based on previous studies, we predicted preserved behavioral performance and robust scaling of pupil dilation with set size and target presence across both groups. We also predicted that biomarker‐positive individuals would exhibit altered load‐dependent pupil modulation, particularly during target‐present trials that require strong phasic LC engagement. Such findings would support the influence of AD‐related pathology on neuromodulatory function and further clarify the functional significance of altered pupillary responses in the preclinical phase of AD.

2. METHODS

2.1. Participants

Participants (N = 61) were recruited from the longitudinal observational cohort of the University of California, San Diego (UCSD) Shiley‐Marcos Alzheimer's Disease Research Center (ADRC). All individuals were classified as cognitively normal at the time of testing, based on comprehensive clinical, neurological, and neuropsychological evaluations conducted by the ADRC. Corrected visual acuity was 20/40 or better for all participants. Individuals with medical, neurological, or psychiatric conditions that could affect cognitive functioning (e.g., stroke, major vascular disease, major depression) were excluded from participation.

Participants were classified as AD biomarker positive (p‐tau+) based on plasma level of phosphorylated tau at either threonine 181 (p‐tau181) or threonine 217 (p‐tau217). Levels of p‐tau were quantified by the National Centralized Repository for Alzheimer's Disease and Related Dementia's (NCRAD) Biomarker Assay Laboratory, ⁶⁴ which transitioned from providing results from p‐tau181 to p‐tau217 during participant recruitment into this investigation. Plasma p‐tau181 was measured using the Quanterix Simoa HDx pTau181 assay following the manufacturer's instructions. Plasma p‐tau181 concentrations of ≥ 4.09 pg/mL were classified as AD biomarker positive. This cut point was previously validated using amyloid PET data for ADRC participants available from the National Alzheimer's Coordinating Center (NACC). ⁶⁵ Plasma p‐tau217 was measured using the ALZpath assay run on the Quanterix HDx platform. Plasma p‐tau217 concentrations of ≥ 0.46 pg/mL were classified as AD biomarker positive (Jeff Dage, NCRAD, personal communication, 2024). This cut score was validated using amyloid PET from NACC, and it is similar to published proposed cutoffs using the ALZpath pTau217 assay in other cohorts. ⁶² The p‐tau level closest in time to pupillometry assessment was selected for use in the current analyses; when both assays were available within 12 months of behavioral assessment, classification was based on p‐tau217. On average, plasma sampling was within 2.5 months (standard deviation [SD] = 4.1) of behavioral assessment.

Demographic and biomarker characteristics are summarized in Table 1. The p‐tau– and p‐tau+ groups were comparable in age, education, APOE ε4 carrier status, and depressive symptoms. A higher proportion of women were present in the p‐tau– group (63.9%) than in the p‐tau+ group (32.0%; p = 0.014). As expected, plasma p‐tau217 and p‐tau181 levels differed significantly between groups (both p < 0.001), reflecting the biomarker thresholds used for classification.

TABLE 1.

Participant demographics and biomarker characteristics.

	p‐tau− (n = 36) mean (SD) / %	p‐tau+ (n = 25) mean (SD) / %	95% CI difference	p value
Age (years)	75.9 (6.1)	77.8 (5.0)	[−4.9,1.1]	0.203
Sex (% women)	63.9	32.0	—	0.014*
Education (years)	17.3 (2.4)	17.5 (2.3)	[−1.4, 1.0]	0.760
Race (% White)	97.2	96.0		0.792
Ethnicity (% Hispanic)	11.1	8.0		0.688
% Glasses	66.6	56.0	—	0.398
GDS score (/30)	0.39 (0.77)	0.68 (1.60)	[‐0.91, 0.32]	0.347
% APOE ε4 carrier	37.1	56.0	—	0.148
p‐tau217 (pg/mL) (n = 35)	0.279 (0.081)	0.772 (0.272)	[−0.621, −0.365]	<0.001*
p‐tau181 (pg/mL) (n = 39)	2.835 (1.322)	5.261 (1.482)	[−3.338, −1.514]	<0.001*
Interval between testing and blood draw (month)	2.34 (3.89)	2.79 (4.38)	[−2.58, 1.68]	0.675

Note: Group classification was based primarily on plasma p‐tau217 concentrations measured within 12 months of testing (n = 35). When p‐tau217 data were unavailable, classification was determined using p‐tau181 levels obtained within the same interval (n = 26). Participants were designated p‐tau+ if pTau181 > 4.09 pg/mL or p‐tau217 > 0.46 pg/mL. Reported n values reflect the number of participants with available biomarker data. Continuous variables are presented as mean ± SD and 95% confidence interval for the between‐group difference; categorical variables are expressed as percentages. Group comparisons were conducted using independent‐samples t tests (continuous) and χ² tests (categorical). Bolded values indicate statistically significant p values.

Abbreviations: APOE, apolipoprotein E; CI, confidence interval; GDS, Geriatric Depression Scale; p‐tau, phosphorylated tau; SD, standard deviation.

Neuropsychological performance is presented in Table 2. Global cognitive status (Montreal Cognitive Assessment) did not differ between groups. Several domain‐specific measures showed statistically reliable group differences. Relative to the p‐tau– group, the p‐tau+ group exhibited lower performance on California Verbal Learning Test‐II (CVLT‐II) long delay recall, category fluency, and the Delis–Kaplan Executive Function System (D‐KEFS) Color–Word interference test, as well as slower completion times on the Trail Making Test Part B (all ps < 0.05). These differences appear to reflect subtle changes within the normal range rather than frank cognitive dysfunction, although it is notable that group differences emerged on measures indexing episodic memory and executive control, domains widely recognized as vulnerable in the earliest stages of AD. Other neuropsychological measures, including Wechsler Memory Scale‐Revised (WMS‐R) Visual Reproduction (VR) Immediate recall, and modified Wisconsin Card Sorting Test (mWCST) performance, did not differ significantly between groups.

TABLE 2.

Neuropsychological test performance of participants.

	p‐tau− (n = 36) mean (SD)	p‐tau+ (n = 25) mean (SD)	95% CI difference	p value
MoCA (/30)	26.08 (2.14)	25.72 (2.59)	[−0.85, 1.58]	0.552
CVLT‐II immediate recall	51.53 (11.20)	45.96 (11.49)	[−0.45, 11.58]	0.069
CVLT‐II short delay recall	10.62 (4.24)	9.32 (4.21)	[0.94, 3.53]	0.249
CVLT‐II long delay recall	11.88 (3.06)	9.52 (4.19)	[0.37, 4.36]	0.022*
WMS‐R VR copy	36.47 (2.74)	36.28 (2.61)	[−1.20, 1.58]	0.783
WMS‐R VR immediate	33.03 (3.85)	31.28 (5.46)	[−0.81, 4.31]	0.175
WMS‐R VR delay	24.08 (7.68)	18.84 (11.63)	[−0.13, 10.62]	0.055
mWCST categories	5.75 (0.73)	5.48 (1.16)	[−0.26, 0.80]	0.309
D‐KEFS CWIT	59.18 (8.99)	64.88 (11.55)	[−11.30, −0.11]	0.046*
Letter fluency	47.39 (11.85)	45.00 (12.72)	[−4.08, 8.86]	0.462
Category fluency	51.67 (10.86)	45.00 (9.79)	[1.12, 11.81]	0.019*
Trail Making: Part A	29.78 (7.30)	35.24 (12.98)	[−11.29, 0.36]	0.065
Trail Making: Part B	41.58 (19.43)	52.92 (24.32)	[−23.11, 0.44]	0.031*

Note: Continuous variables are presented as mean ± SD and 95% confidence interval for the between‐group difference. Group comparisons were conducted using independent‐samples t tests.

Abbreviations: CI, confidence interval; CVLT‐II, California Verbal Learning Test 2nd Edition CWIT, Color–Word Interference Test; D‐KEFS, Delis–Kaplan Executive Function System; mWCST, modified Wisconsin Card Sorting Test; MoCA, Montreal Cognitive Assessment; p‐tau, phosphorylated tau; SD, standard deviation; WMS‐R VR, Wechsler Memory Scale‐Revised Visual Reproduction. Bolded values indicate statistically significant p values.

RESEARCH IN CONTEXT

1. Systematic review: Alzheimer's disease (AD) pathology occurs early in the locus coeruleus (LC) but identifying its functional consequences in vivo remains challenging. The authors reviewed literature on the impact of AD pathology on LC‐mediated cognitive function using traditional sources (e.g., Pub Med, SCOPUS). Prior work shows that pupil dilation indexes LC–norepinephrine (NE) activity and that task‐evoked pupil dilation can be affected by cognitive load, aging, and neurodegenerative disease risk. No studies have examined LC–NE dependent pupil dynamics during a visual search task that manipulates cognitive load in cognitively normal older adults with preclinical AD.

2. Interpretation: Our findings indicate that cognitively normal older adults with preclinical AD exhibit attenuated load‐dependent pupil modulation despite intact behavioral performance, which suggests early disruption in LC–NE gain scaling (i.e., ability to adjust the sensitivity of neural responses to optimize attention). This early neuromodulatory vulnerability in preclinical AD is not captured by traditional behavioral measures.

3. Future directions: Future studies should examine whether pupillary responses related to LC–NE function predict longitudinal cognitive decline in those with preclinical AD, interact with other risk factors (e.g., apolipoprotein E ε4, vascular burden), or respond to pharmacologic or cognitive interventions aimed at bolstering noradrenergic function.

Written informed consent was obtained from all study participants in accordance with the Declaration of Helsinki. All procedures were approved by the UCSD Institutional Review Board.

2.2. Visual search task

Participants completed a conjunctive visual search task in which the target was defined by a unique combination of motion and luminance. The target was a black dot moving vertically (up–down), while distractors were white dots moving vertically and black dots moving horizontally (left–right). Correct detection therefore required binding of motion and luminance features (Figure 1).

FIGURE 1.

Schematic of the conjunctive visual search task. Each trial began with a fixation display, followed by either a target‐present (left) or target‐absent (right) search array. The target was a black dot moving vertically, while distractors were white vertically moving dots and black horizontally moving dots.

Each trial began with a yellow fixation cross (1 second), followed by the stimulus display. Stimuli appeared randomly in six possible positions spaced 60° apart along an imaginary circle centered on the fixation cross at a fixed eccentricity of 8.8°. Each dot subtended 2.1° of visual angle and moved at 4.22°/second. Displays contained one, three, or five stimuli, with the target present on 50% of trials. Participants pressed the spacebar as quickly as possible when the target was present and withheld responses when it was absent. The display remained visible until response or 3000 ms had elapsed. Participants completed eight practice trials (four target present, four target absent) with accuracy feedback, followed by 108 test trials (54 target present, 54 target absent) presented in a single randomized block during which no feedback was provided. Trial order and positions were randomized. Eye‐tracking data were recorded continuously to capture gaze position and pupil changes throughout the task.

2.3. Procedure

Testing took place individually in a quiet room at the UCSD ADRC. Instructions were presented on screen and read aloud by the examiner. Before the main conjunctive visual search task, participants completed a choice reaction time (CRT) task designed to assess basic response speed when discriminating between static versions of the moving stimuli used in the search task. In the CRT task, participants viewed displays containing either white static dots or black static dots and pressed the spacebar when the dots were black (target‐present trials) and withheld responses when they were white (target‐absent trials). The CRT task consisted of 32 trials (16 black dot, 16 white dot) presented in randomized order. Each trial contained six dots in the same six spatial positions used in the search task. Accuracy and reaction time (RT) were recorded on target‐present trials. After the CRT task and prior to commencement of the visual search task, a nine‐point eye‐tracker calibration and validation procedure was performed.

2.4. Apparatus and calibration

Eye‐tracking data were recorded using the Tobii Pro Fusion system (Tobii AB) at 250 Hz. Participants were seated in a dimly lit, windowless room. Ambient illumination was measured at eye level and maintained at a mean of 67.8 lux (SD = 7.8) across participants. With the display turned off, this value corresponds to an estimated luminance of ≈ 17 cd/m², based on a diffuse Lambertian assumption (L = E/π) and average wall reflectance of 0.8. Visual stimuli were displayed on a 21.5‐inch LCD monitor (1920 × 1080 pixels; 60 Hz refresh rate) positioned ≈ 57 cm from the participant.

A nine‐point calibration and validation procedure was completed before each block of trials, with acceptable calibration defined as mean gaze error < 0.5°. Stimulus presentation was controlled using E‐Prime 3.0 (Psychology Software Tools). Raw gaze and pupil data were exported for offline analysis in MATLAB R2024b (MathWorks) using custom scripts aligned with established pupillometry guidelines. ⁶⁶ , ⁶⁷ , ⁶⁸

2.5. Pupillometric data preprocessing

Analyses were based on data from the right eye. Blinks were detected using a velocity‐based algorithm. Missing samples were linearly interpolated, and adjacent gaps shorter than 100 ms were merged. Trials were segmented from fixation onset through 2500 ms after stimulus onset. Baseline correction was performed by computing the median pupil diameter over the final 400 ms of the pre‐stimulus fixation period and subtracting this value from the entire trial time series.

A total of 105 cognitively normal participants with biomarker data and eye‐tracking data completed testing, but 44 individuals were excluded because they did not contribute the minimum number of valid pupil data trials (i.e., five trials with < 50% missing data per condition) after artifact rejection. The final sample for analysis therefore included 61 participants. While participant exclusion rates of ≈ 20% to 30% are common in pupillometry studies with older adults due to insufficient usable pupil data, our higher exclusion rate likely reflects several procedural and design factors. Our study was partially performed under COVID‐era safety protocols, which required mask wearing during testing. Occasionally this interfered with eye tracking because the system sometimes captured the mask edge or transiently lost the eye signal. Additional factors that may have contributed to missing data included the absence of a headrest, fixation instructions that did not explicitly require strict gaze stabilization, and task timing that was not structured to provide explicit blink periods. Because trials proceeded continuously, participants were not instructed to minimize blinking during critical trial intervals. All of these factors can increase signal loss and blink‐related artifacts, thereby reducing the number of usable pupil trials.

To evaluate potential attrition bias, demographic and neuropsychological characteristics were compared between retained and excluded participants (Table S1 in supporting information). The groups did not differ on age, sex, global cognitive status, depressive symptoms, or neuropsychological performance (all ps > 0.05). Excluded participants had slightly fewer years of education (p = 0.036). Importantly, education was not associated with trial retention within the sample analyzed, and there was no evidence that exclusion selectively altered cognitive or biomarker group composition. These results suggest that exclusions were driven primarily by data quality considerations rather than systematic demographic or cognitive differences.

Task accuracy for the 61 included participants was near ceiling, with < 1% of trials classified as incorrect across participants and conditions (Table S2 in supporting information). Given the very small number of error trials, analyses of pupil dynamics were restricted to correct trials. Trials with > 50% missing pupil data were excluded. Across participants, 70.2% (SD = 16.6%) of trials per condition were retained for analysis, corresponding to a mean of 12.6 (SD = 3.0) valid trials per condition. Valid trials did not differ as a function of set size, trial type, or biomarker group (all ps > 0.05, Table S3 in supporting information). No evidence of systematic missingness was observed. Trial loss was driven primarily by blink‐related artifacts rather than condition‐ or group‐related factors.

2.6. Data analysis

Behavioral and pupillometric data were analyzed using distinct statistical approaches. Behavioral RT data were analyzed at the trial level using linear mixed effects models. Pupillometric data were analyzed using two complementary approaches: (1) conventional trial‐level analyses based on peak amplitude (maximum pupil size), half‐max latency (time at which amplitude reaches 50% of its maximum), and peak velocity (maximum positive slope during rising phase), all derived from fitted pupil traces; and (2) continuous time‐resolved regression analyses modeling pupil diameter at each time point. All analyses were performed in IBM SPSS 31.0 and MATLAB R2024b.

2.6.1. Behavioral analyses

Behavioral accuracy was compared across groups using Mann–Whitney U tests for both the CRT task and the conjunctive search task due to non‐normal distributional characteristics. CRT RTs were compared across groups using independent‐samples t tests. RTs in the search task (correct target‐present trials only) were analyzed with a linear mixed‐effects model (LMM) structure parallel to the structure used for the pupil outcomes (i.e., set size × group; covariates: age, sex, trial index).

2.6.2. Conventional trial‐level analyses

For each trial, baseline‐corrected pupil traces from 0 to 2500 ms post stimulus onset were fit using a second‐order polynomial (quadratic) function. This functional form was selected to provide a smooth, single‐peaked approximation of the pupil response while minimizing high‐frequency noise. From each fitted curve, three dependent measures were extracted: (1) peak amplitude, defined as the maximum of the quadratic function; (2) half‐max latency, defined as the time from stimulus onset at which the fitted function reached 50% of its peak value; and (3) peak dilation velocity, defined as the maximum positive first derivative of the fitted quadratic function prior to the peak amplitude (expressed in millimeters per second). This velocity measure captured the steepest rate of pupil dilation along the ascending limb of the response while excluding post peak slope fluctuations. Each measure was analyzed using LMMs with set size (1, 3, 5), trial type (target present, target absent), and group (p‐tau+, p‐tau–) as fixed effects, and all two‐ and three‐way interactions. Age, sex, and trial index (all z scored) were included as covariates. Baseline pupil size was not included as a covariate because pupil values were already baseline corrected.

Random intercepts and random slopes for set size were estimated at the participant level using restricted maximum likelihood (REML) with an unstructured (UN) covariance matrix. Model diagnostics included checks of convergence, variance components, and residual distributions. All models converged without warnings, variance estimates were within plausible ranges, and residual patterns did not indicate violations requiring transformation. Estimated marginal means (EMMs) and pairwise comparisons used least significant difference correction. Fixed effects were evaluated using F tests. Effect sizes are reported as approximate partial η ² values computed from the F statistics for each fixed effect. Because variance partitioning differs in mixed‐effects models, these values are descriptive rather than exact measures of explained variance.

To evaluate potential pupil foreshortening effects associated with gaze position, two trial‐level gaze metrics were computed across the 0 to 2500 ms pupil analysis window: (1) gaze eccentricity, defined as the mean Euclidean distance of gaze samples from screen center; and (2) gaze dispersion, defined as the spatial variability of gaze samples quantified as $\sqrt{\mathrm{var}(x)+\mathrm{var}(y)}$. These measures captured fixation deviation and gaze variability during the interval used for pupil analysis. Both gaze metrics were analyzed using LMMs with the same fixed effects, covariates, and random effects structure as the pupil models. To directly assess foreshortening confounds, the pupil LMMs were reestimated with trial‐level gaze eccentricity included as an additional covariate. Inclusion of eccentricity did not alter the pattern or significance of any pupil effects. Detailed results are provided in the supporting information.

Although peak amplitude, half‐max latency, and peak dilation velocity were analyzed as separate dependent variables, no additional family‐wise corrections were applied at the trial level. Interpretation emphasizes convergent patterns across measures and consistency with the time‐resolved analyses, reducing the likelihood that isolated effects reflect Type I error.

2.6.3. Continuous time–series regression analyses

To complement the trial‐level summary analyses, a two‐stage regression approach was used to model pupil diameter as a continuous time series across the 0 to 2500 ms interval. At the first (within‐subject) stage, pupil diameter (z scored, non‐baseline corrected) was regressed onto trial‐level predictors at each time point. For the full model, predictors included: trial type, set size, time‐on‐task (ToT, z score trial index), baseline pupil size (z scored), and the interaction terms trial type × set size, trial type × ToT, set size × ToT, and baseline pupil size × ToT. Baseline pupil size was included to account for trial‐to‐trial fluctuations in tonic arousal.

In addition, separate models were estimated for target‐present and target‐absent trials. The present‐only model included a z scored speed predictor (1/RT for correct trials) together with set size, ToT, baseline pupil size, and their interactions (speed × set size, speed × ToT, set size × ToT, baseline pupil size × ToT). The absent‐only model included set size, ToT, baseline pupil size, and the set size × ToT and baseline pupil size × ToT interactions. Trials with missing predictor values or missing pupil data were excluded on a per‐model basis. This first stage yielded a β(t) curve for each participant and predictor.

At the group‐level stage, participant‐level β(t) estimates were entered into a general linear model containing an intercept and group (p‐tau+ vs. p‐tau–), with age (z scored) and sex (coded 0/1) as covariates. Significance was assessed using cluster‐based permutation tests (Freedman–Lane method, 5000 permutations, family‐wise error controlled across time at p < 0.05). For each significant cluster we report its temporal extent, the peak t value, cluster mass, and mean β with a bootstrap confidence interval (CI).

3. RESULTS

3.1. Behavioral performance

Performance on the CRT task confirmed that groups were well matched on basic perceptual and motor speed. Reaction times did not differ significantly between groups (p‐tau−: M = 478 ms, SD = 129; p‐tau+: M = 491 ms, SD = 122), p = 0.70). Accuracy was also near ceiling and comparable across groups (p‐tau−: M = 98.8%, SD = 2.4; p‐tau+: M = 98.3%, SD = 5.2, p = 0.62).

In the conjunctive visual search task, accuracy was uniformly high across set sizes and did not differ by p‐tau group (all ps > 0.78; see Table S2). As expected, search RTs increased with set size (F[2, 122.278] = 241.513, p < 0.001, η ² p = 0.798), indicating a robust effect of attentional load on performance. There was no main effect of p‐tau group (F[1, 60.783] = 0.063, p = 0.803) and no significant set size × p‐tau group interaction (F[2, 122.169] = 1.547, p = 0.217). Among the covariates, age was a significant predictor (F[1, 55.843] = 4.385, p = 0.041, η ² p = 0.073) with greater age associated with slower response times. Sex was also significant (F[1, 52.410] = 10.043, p = 0.003, η ² p = 0.161) with men responding faster than women in the search task. Trial index was not significant (F[1, 2210.663] = 0.028, p = 0.867).

Overall, both groups demonstrated high accuracy and comparable search efficiency, indicating equivalent task engagement. Thus, any group differences in pupil dynamics are unlikely to be driven by gross differences in behavioral performance or basic processing speed.

3.2. Baseline‐corrected pupil time courses

Figure 2 shows baseline‐corrected pupil diameter as a function of stimulus time separately for target‐present and target‐absent trials, with set size represented across panels and biomarker groups shown with each panel. Across both groups, pupil diameter increased after display onset, reached a peak between ≈ 1000 and 1500 ms after stimulus presentation, and then gradually returned toward baseline. Pupil responses scaled positively with set size in all conditions and were markedly larger and more sustained on target‐present trials than on target‐absent trials.

FIGURE 2.

Baseline‐corrected pupil time courses by trial type, set size, and phosphorylated tau (p‐tau) group. Grand‐average baseline‐corrected pupil diameter is plotted as a function of stimulus time for each set size (1, 3, 5). Target‐present trials are shown in the top row and target‐absent trials in the bottom row. Within each panel, curves represent p‐tau− and p‐tau+ participants. Shaded regions indicate 95% confidence intervals. Across groups, pupil dilation increased systematically with set size and was larger and more sustained for target‐present than target‐absent trials. Relative to the p‐tau− group, the p‐tau+ group showed reduced load‐related modulation during target‐present trials, most evident during the late dilation interval (≈ 800–1800 ms), consistent with the time‐resolved regression results (Tables 4 and 5).

Visual inspection also revealed subtle but consistent group differences during target‐present trials. Whereas p‐tau− participants showed clear separation of the set‐size functions throughout the rising and peak phases of the response, the p‐tau+ group exhibited noticeably shallower load‐related divergence, particularly during the late peak interval (≈ 800–1800 ms). In contrast, group differences were minimal for target‐absent trials, with highly similar waveforms across all set sizes in both groups. These descriptive patterns are consistent with the mixed‐effects and time‐series regression analyses reported below, which indicate attenuated load‐related modulation of pupil dilation in biomarker‐positive individuals under target‐present conditions.

3.3. Trial‐level summary pupil analyses

3.3.1. Peak amplitude

The linear mixed‐effects model revealed a strong effect of set size on peak pupil dilation, with amplitudes increasing with higher loads (F[2, 129.390] = 150.180, p < 0.001, η ² p = 0.699). A robust main effect of trial type was also observed, with greater peak dilation on target‐present than target‐absent trials (F[1, 4504.151] = 839.563, p < 0.001, η ² p = 0.157). These effects were qualified by a significant set size × trial type interaction (F[2, 4504.084] = 13.188, p < 0.001, η ² p = 0.006), indicating that the effect of set size differed across trial types (Table 3). Peak amplitude increased with set size for both trial types, but the magnitude of the trial‐type difference varied across loads. This pattern is consistent with the separation of waveforms across set sizes and trial types shown in Figure 2.

TABLE 3.

Pupil eye metrics by trial type, set size, and p‐tau group.

	Trial type	Set size	p‐tau− (n = 36) mean (SE)	95% CI	p‐tau+ (n = 25) mean (SE)	95% CI
Peak amplitude (mm)	Present	1	0.223 (0.012)	[0.198, 0.247]	0.216 (0.015)	[0.187, 0.246]
		3	0.296 (0.014)	[0.267, 0.325]	0.268 (0.014)	[0.233, 0.303]
		5	0.332 (0.017)	[0.298, 0.366]	0.304 (0.021)	[0.262, 0.345]
	Absent	1	0.070 (0.012)	[0.045, 0.094]	0.087 (0.015)	[0.057, 0.117]
		3	0.194 (0.014)	[0.165, 0.223]	0.183 (0.017)	[0.148, 0.217]
		5	0.222 (0.017)	[0.188, 0.256]	0.203 (0.021)	[0.162, 0.244]
Half‐max latency (ms)	Present	1	630 (27)	[576, 684]	684 (33)	[619, 749]
		3	794 (30)	[736, 853]	877 (36)	[806, 950]
		5	874 (35)	[805, 943]	890 (42)	[807, 974]
	Absent	1	339 (28)	[283, 395]	345 (34)	[278, 413]
		3	565 (30)	[506, 624]	572 (36)	[501, 643]
		5	654 (34)	[585, 722]	694 (42)	[611, 777]
Peak dilation velocity (mm/s)	Present	1	3.02 (0.16)	[2.71, 3.34]	2.76 (0.19)	[2.38, 3.14]
		3	3.35 (0.17)	[3.01, 3.69]	3.36 (0.21)	[2.95, 3.78]
		5	3.71 (0.19)	[3.34, 4.08]	3.34 (0.23)	[2.89, 3.79]
	Absent	1	2.36 (0.16)	[2.05, 2.68]	2.21 (0.19)	[1.83, 2.59]
		3	3.15 (0.17)	[2.81, 3.49]	2.96 (0.21)	[2.55, 3.37]
		5	3.43 (0.18)	[3.06, 3.79]	3.33 (0.23)	[2.89, 3.78]

Note: Values reflect estimated marginal means derived from the linear mixed effects model including trial number, sex, and age as fixed covariates, with subject‐level random intercepts and a random slope for set size. Standard errors appear in parentheses. Confidence intervals represent model‐adjusted 95% CIs.

Abbreviations: CI, confidence interval; p‐tau, phosphorylated tau; SE, standard error.

The main effect of p‐tau group was not significant (F[1, 60.869] = 0.346, p = 0.558], and the set size × trial type × group interaction was also not significant (F[2, 4504.326] = 0.244, p = 0.78). A small but significant trial type × group interaction emerged (F[1, 4504.340] = 4.620, p = 0.032, η ² p = 0.001), reflecting reduced peak dilation on target‐present trials in the p‐tau+ group relative to p‐tau− group, while target‐absent trials were comparable across groups. The set size × group interaction was marginal (F[2, 129.351] = 2.717, p = 0.07, η ² p = 0.040).

Taken together, these results indicate that peak pupil dilation scaled robustly with both visual load and trial type in the sample as a whole. Group differences were minimal, although biomarker‐positive individuals exhibited an attenuation of dilation on target‐present trials.

3.3.2. Half‐max latency

Half‐max latency showed a broadly similar pattern to peak amplitude, with systematic modulation by both visual load and trial type. The mixed‐effects model revealed a strong main effect of set size, with pupil responses reaching 50% of peak later at higher loads (F[2, 121.071] = 93.087, p < 0.001, η ² p = 0.606). Latencies were also longer on target‐present than target‐absent trials (F[1, 4308.433] = 403.722, p < 0.001, η ² p = 0.086), indicating longer latencies on target‐present relative to target‐absent trials. These effects were qualified by a significant set size × trial type interaction (F[2, 4310.374] = 5.503, p = 0.004, η ² p = 0.003). EMMs (Table 3) indicated monotonic latency increases with set size for both trial types, with the present–absent difference largest at set size 1.

No main effect of p‐tau group (F[1, 59.033] = 0.829, p = 0.366), and neither the trial type × group interaction (F[1, 4308.389] = 1.583, p = 0.208) nor the set size × trial type × group interaction was significant (F[2, 4310.785] = 1.347, p = 0.260). The set size × group interaction was also not significant (F[2, 121.014] = 0.173, p = 0.841). EMMs indicated numerically longer latencies in the p‐tau+ group, but these differences were small and did not reach statistical significance.

3.3.3. Peak dilation velocity

Peak dilation velocity varied reliably with set size (F[2, 163.025] = 79.558, p < 0.001, η ² p = 0.494), with velocities increasing across loads (Table 3). A significant main effect of trial type was also observed (F[1, 4508.328] = 42.694, p < 0.001, η ²Pp = 0.009), reflecting slightly greater peak velocities on target‐present relative to target‐absent trials. These effects were qualified by a significant set size × trial type interaction (F[2, 4508.093] = 6.352, p = 0.002, η ² p = 0.003), indicating that the trial‐type difference was largest at lower set sizes. No significant effects involving p‐tau group were observed.

3.3.4. Gaze control analyses

To assess whether systematic differences in gaze behavior could contribute to the observed pupil effects, we conducted parallel mixed‐effects analyses on two trial‐level gaze metrics: gaze eccentricity (deviation of gaze from screen center) and gaze dispersion (spatial variability of gaze samples). Both models included fixed effects of set size, trial type, and p‐tau group, their interactions, covariates (trial index, age, sex), and the same random‐effects structure used in the pupil models (Supplementary Analyses S1 and S2, and Table S4 in supporting information).

Gaze eccentricity varied reliably as a function of task demands, showing significant effects of set size (F[2, 140.883] = 37.800, p < 0.001), trial type (F[1, 4526.565] = 13.066, p < 0.001], and a set size × trial type interaction (F[2, 4525.870] = 11.243, p < 0.001). Gaze dispersion likewise exhibited a strong main effect of set size (F[2, 124.532] = 81.114, p < 0.001) and a robust set size × trial type interaction (F[2, 4518.949] = 31.963, p < 0.001), indicating increased gaze variability with higher visual load. Critically, neither metric showed a main effect of p‐tau group nor any interactions involving group (all ps > 0.10), indicating comparable gaze behavior across biomarker groups.

To evaluate potential pupil foreshortening artifacts, all trial‐level models were re‐estimated with trial‐level gaze eccentricity included as an additional covariate. Inclusion of eccentricity did not alter the pattern or significance of any effects (Supplementary Analysis S3, and Tables S5–S7 in supporting information).

3.4. Continuous time–series regression analyses

3.4.1. Population‐level effects

Time‐resolved mixed‐effects regression revealed clear and consistent predictors of pupil dynamics across the 0–2500 ms interval (Table 4). Across all models, baseline pupil size showed a strong positive association across the entire analysis window (0–2496 ms, ps < 0.001). The magnitude of this effect was largest early in the epoch and decreased gradually over time, indicating that trials with larger initial pupil diameters were associated with systematically larger pupil values across the response period. A broad and sustained effect of set size was also evident with significant positive clusters beginning between 220 and 1100 ms and extending to the end of the epoch (ps < 0.001), confirming that pupil dilation scaled reliably with increasing visual load throughout the trial.

TABLE 4.

Population‐level time‐series regression models.

Model	Predictor	Temporal window (ms)	Peak t	Mean β [95% CI]	Cluster p value
Full	Baseline	0–2496	35.64	0.412 [0.343, 0.482]	<0.001
	Set size	572–2496	16.53	0.304 [0.254, 0.352]	<0.001
	Trial type	496–2496	16.32	0.305 [0.262, 0.347]	<0.001
	Time‐on‐task	1308–2496	−3.59	−0.077 [−0.137, −0.021]	0.022
	Set size × trial type	300–1596	−8.70	−0.113 [−0.144, −0.087]	<0.001
Present	Baseline	0–2496	36.37	0.449 [0.386, 0.511]	<0.001
	Set size	1112–2496	4.25	0.191 [0.093, 0.286]	<0.001
	Set size × time‐on‐task	692–948	3.22	0.130 [0.030, 0.232]	0.036
	Speed	1376–2496	−7.84	−0.329 [−0.461, −0.171]	0.002
	Speed × set size	1936–2496	−5.28	−0.178 [−0.253, −0.103]	0.003
	Time‐on‐task	1644–2076	−3.17	−0.107 [−0.195, −0.028]	0.048
	Baseline x time‐on‐task	852–988	3.13	0.104 [0.022, 0.199]	0.026
Absent	Baseline	0–2496	26.07	0.445 [0.359, 0.522]	<0.001
	Set size	220–2496	15.89	0.333 [0.266, 0.394]	<0.001

Note: β values reflect standardized regression coefficients from time‐resolved mixed‐effects models computed at each time point. Temporal windows represent cluster‐corrected intervals of significance using 5000 sign‐flip permutations with a cluster‐forming threshold of p < 0.05 (two tailed). Peak t corresponds to the time point within the cluster showing maximum absolute t value. CI = 95% bootstrap confidence interval around the window‐mean β. ToT = Time‐on‐Task (trial index, z scored). Speed = 1/reaction time of correct trials (z scored). Set size refers to visual load levels (1, 3, 5). Models include covariates for age and sex (both z scored). All reported clusters survived family‐wise error correction across time.

In the full model, trial type was associated with a long‐lasting positive cluster (496–2496 ms, p < 0.001), reflecting the larger and more sustained dilation on target‐present trials relative to target‐absent trials. A significant set size × trial type interaction (300–1596 ms, p < 0.001) showed that load‐related increases in dilation were amplified when a target was present. ToT contributed a small late negative cluster (1308–2496 ms, p = 0.022), consistent with a gradual reduction in pupil magnitude over the course of the session, a pattern commonly observed in sustained cognitive tasks and typically attributed to fatigue, habituation, or declining task engagement.

When analyses were restricted to target‐present trials, baseline and load effects persisted. Baseline pupil size produced a strong full‐epoch effect (0–2496 ms, p < 0.001). In addition, a significant baseline × ToT interaction emerged (852–988 ms, p = 0.026), indicating that early evoked pupil responses varied systematically as a function of session progression. Set size yielded a sustained late cluster (1112–2496 ms, p < 0.001). A modest early set size × ToT interaction (692–948 ms, p = 0.036) suggested that sensitivity to load declined slightly over time. Two predictors associated with individual differences in processing speed yielded reliable late negative clusters: a main effect of speed (1/RT; 1376–2496 ms, p = 0.002) and a speed × set size interaction (1936–2496 ms, p = 0.003). Trials with faster responses exhibited reduced late‐trial pupil diameter, particularly at higher loads. ToT also yielded a small standalone late negative cluster (1644–2076 ms, p = 0.048).

For target‐absent trials, the pattern was more selective. Baseline pupil size again produced a strong full‐epoch effect (0–2496 ms, p < 0.001), and set size yielded a robust sustained cluster (220–2496 ms, p < 0.001). No ToT or interaction effects survived cluster correction, indicating that in the absence of target detection demands, pupil responses tracked load but were comparatively insensitive to fatigue.

Together with Figure 2, these results indicate that the paradigm elicited robust and load‐sensitive pupil dynamics that were consistently captured by both conventional summary metrics and time‐resolved modeling approaches.

3.4.2. Group differences

Time‐resolved group‐level analyses revealed converging evidence that biomarker‐positive (p‐tau+) individuals showed reduced load‐related modulation of pupil dilation under conditions involving active target processing, whereas evoked dynamics were otherwise largely comparable across groups (Table 5). In the full model that included both present and absent trials, a significant set size × group cluster emerged between 752 and 1860 ms (cluster‐corrected p = 0.0038), peaking at 1392 ms (t = −4.02). Across this window, the p‐tau+ group exhibited attenuated scaling of pupil dilation with increasing load (mean β = −0.118; 95% CI [−0.185, −0.055]). This effect aligns with the period of maximal task‐evoked dilation, suggesting blunted load‐related modulation in biomarker‐positive individuals, and is consistent with the shallower set‐size separation in the p‐tau+ target‐present waveforms in Figure 2. A second, very small intercept × group cluster (356–364 ms, p = 0.044) had negligible magnitude and no clear physiological interpretation. No other group‐related effects survived cluster correction.

TABLE 5.

Group‐level time‐series regression models.

Model	Predictor	Window (ms)	Peak t	Mean β[95% CI]	Cluster p value
Full	Set size × group	752–1860	−4.02	−0.118 [−0.185, −0.055]	0.004
	Intercept x group	356–364	−2.40	≈ 0 [−2.46e−15, 6.24e−16]	0.044
Present	Set size × group	1016–2192	−4.73	−0.178 [−0.252, −0.108]	<0.001
Present (with speed)	No significant effects	—	—	—	—
Absent	No significant effects	—	—	—	—

Note: β values represent standardized coefficients at the group level. Temporal windows reflect cluster‐corrected significance using 5000‐permutation Freedman–Lane tests. CI = bootstrap percentile confidence interval.

In the target present‐only model that paralleled the full analysis, a robust set size × group effect emerged from 1016 to 2192 ms (p < 0.001, peak t = −4.73). As in the full model, p‐tau+ participants showed significantly attenuated load‐related modulation (mean β = −0.178, 95% CI [−0.252, −0.108]). Figure 3A illustrates the time‐resolved standardized regression coefficient β(t) for the set size × group interaction, with the significant interval highlighted in red. To summarize this effect at the individual level, β(t) values were averaged across the significant window to derive a subject‐level β _SetSize index. As shown in Figure 3B, the distribution of β _SetSize values reflects reduced load‐dependent modulation in the p‐tau+ group, mirroring the sustained negative cluster observed in the time‐series analysis. When speed‐related predictors (speed, speed × set size, and speed × ToT) were additionally included in the model, no significant group differences were detected, suggesting that group‐related changes in LC load responsivity may be reflected in individual differences in processing speed.

FIGURE 3.

Time‐resolved and subject‐level indices of the set size × phosphorylated tau (p‐tau) group effect (target‐present model). A, Standardized regression coefficients (β) for the set size × p‐tau group interaction are plotted as a function of time from stimulus onset. The green line shows the group‐level β(t) curve, and the shaded region represents the bootstrap 95% confidence interval. Time points within the cluster‐corrected significant window (1016–2192 ms) are highlighted in red, indicating a sustained negative group effect during this interval. Negative β values reflect attenuated load‐related modulation in the p‐tau+ group relative to the p‐tau− group. B, Raincloud plot showing the distribution of subject‐level β _SetSize values, computed as the mean set size slope across the significant interval. Half‐violins depict kernel density distributions, boxplots summarize the median and interquartile range, and points represent individual participants. Consistent with the time‐resolved analysis, p‐tau+ individuals exhibit reduced load‐dependent pupil modulation relative to p‐tau− individuals.

In contrast, no group differences were observed in the target absent‐only model. Neither main effects nor interactions involving p‐tau group produced significant temporal clusters. Thus, when no target was present and no response was required, p‐tau+ and p‐tau− participants showed comparable pupil dynamics, even though load effects were robust at the population level (bottom panels in Figure 2).

Taken together, these findings indicate that group differences in pupil dynamics are not global shifts in overall amplitude, but rather selective reductions in load‐dependent modulation that emerge during periods of active target processing and speeded responding.

3.5. Individual‐difference analyses

To examine whether individual variability in pupil‐based load modulation was related to biomarker status or cognitive performance, we derived a subject‐level index from the target‐present model. Set size was treated as a continuous predictor in the time‐series framework, yielding a β(t) function representing the slope of the load–pupil relationship. For each participant, β(t) values were averaged across the significant set size × group cluster (1016–2192 ms). The resulting β _SetSize index reflects the magnitude of linear load–related modulation within the interval of strongest group differentiation.

We then examined associations between β _SetSize , plasma biomarker concentrations (p‐tau217, p‐tau181), and performance on standardized neuropsychological tests commonly used in ADRC assessments. The cognitive domains assessed included verbal learning and memory (CVLT‐II), visual memory (WMS‐R VR), executive functions (mWCST; D‐KEFS Color–Word Interference Test [CWIT]; Verbal Fluency), and processing speed and attention (Trail Making Test, Parts A and B).

Higher plasma p‐tau217 concentrations were significantly associated with reduced load‐related pupil modulation (r[33] = −0.36, p = 0.034), consistent with attenuated LC–NE gain control in biomarker‐positive individuals (Figure 4A). In contrast, p‐tau181 showed no relationship with β _SetSize (r[37] = −0.17, p = 0.291; Figure 4B), reinforcing emerging evidence that p‐tau217 provides a more sensitive index of early tau pathology relevant to neuromodulatory system dysfunction. Across neuropsychological measures, weaker load‐related pupil modulation was associated with poorer delayed visual memory (WMS‐R VR Delay; r[59] = 0.28, p = 0.031; Figure 4C) and reduced inhibitory control (D‐KEFS Color–Word Stroop Inhibition; r[57] = −0.26, p = 0.050; Figure 4D). A marginal association was observed with processing speed (Trail Making Test, Part A; r[59] = −0.23, p = 0.080), such that slower performance corresponded to weaker modulation. Performance on tests of verbal fluency, verbal learning (CVLT‐II), and set shifting (mWCST) showed no significant associations with β _SetSize , suggesting that the strongest cognitive correlates arise within domains most dependent on rapid attentional control and visuospatial memory.

FIGURE 4.

Associations of load‐related pupil modulation with biomarker and cognitive measures. Scatterplots illustrate relationships between the subject‐level β _SetSize index, computed as the mean set size slope across the significant interaction window (1016–2192 ms), and (A) plasma p‐tau217 concentration, (B) plasma p‐tau181 concentration, (C) delayed visual memory recall performance (WMS‐R Visual Reproduction Delay total score), and (D) inhibitory control (D‐KEFS Color‐Word interference test, time [seconds] to completion). Points represent individual participants and are color coded by p‐tau group (blue = p‐tau−; red = p‐tau+). Solid black lines depict regression fits across all participants. Higher p‐tau217 concentrations were associated with reduced load‐related pupil modulation, whereas p‐tau181 showed no significant association. Weaker modulation was also related to poorer delayed visual memory (fewer items recalled) and reduced inhibitory control (longer completion time). D‐KEFS, Delis–Kaplan Executive Function System; p‐tau, phosphorylated tau; WMS‐R, Wechsler Memory Scale‐Revised.

To determine whether group‐related pupil effects could be explained by cognitive variability (e.g., driven by differences in cognitive effort), time‐series regression models were reestimated with WMS‐R VR Delay included as an additional predictor (Supplementary Analysis S4, and Tables S8 and S9 in supporting information). Inclusion of the cognitive covariate did not alter the primary task‐related pupil effects. Critically, the significant set size × group clusters observed in both the full and present models were preserved with highly similar temporal boundaries and effect magnitudes. In the present model including speed, in which no clusters reached significance in the original specification, adjustment for WMS‐R VR Delay resulted in the emergence of a reliable set size × group cluster. Adjustment for delayed memory performance increased sensitivity to load‐dependent group differences, suggesting that shared variance between cognition and pupil dynamics influenced detectability in the original model.

Additional clusters involving baseline × trial type × WMS‐R VR delay, trial type × WMS‐R VR delay, and speed × WMS‐R VR delay terms were also observed after covariate inclusion. These effects indicate that pupil dynamics are sensitive to individual differences in delayed memory ability, consistent with the known relationships among pupil diameter, cognitive effort, and task engagement. Importantly, these cognition‐related effects did not eliminate or attenuate the load‐dependent group effects that constitute the central findings of the study. Instead, the results indicate that cognitive variability and group‐related physiological effects explain partially independent components of pupil variance.

For completeness, analogous LMMs of the conventional pupil metrics (peak amplitude, half‐max latency, peak dilation velocity) were also re‐estimated with WMS‐R VR Delay included as a covariate (Supplementary Analysis S5, and Tables S10–S12 in supporting information). Consistent with the time series results, inclusion of the cognitive predictor did not modify the pattern of main effects or interactions, and no changes in statistical inference were observed.

Taken together, these findings indicate that pupil‐based load modulation is systematically related to both plasma tau levels and performance in specific cognitive domains. At the same time, the stability and enhanced detectability of set size × group effects after adjustment for delayed memory performance demonstrate that the observed group differences cannot be reduced to cognitive performance variability alone. Rather, the results support an interpretation in which pupil dynamics simultaneously reflect cognitive effort–related processes and task‐dependent neuromodulatory mechanisms.

4. DISCUSSION

The present study examined whether AD‐related pathology is associated with altered task‐evoked pupil dynamics in cognitively normal older adults performing a conjunctive visual search task. Although the overall shape of the pupil response when performing the task was similar between biomarker groups, there was a selective reduction in load‐dependent modulation among biomarker‐positive individuals. The p‐tau+ participants showed a diminished increase in pupil diameter as visual load increased and this attenuation was most pronounced during the mid‐to‐late portion of the dilation response, a time window corresponding to peak phasic LC‐mediated neuromodulatory engagement in primate and human studies. ²³ , ²⁹ , ³¹ , ³²

When analyses were restricted to target‐present trials and speed‐related predictors were removed, this same pattern emerged with even greater clarity. This may be because removal of speed‐related predictors frees variance that is mechanistically shared with neuromodulatory function because pupil‐linked arousal and response speed are both downstream consequences of LC–NE engagement (rather than speed‐related predicters representing an independent confound). Analyses limited to target‐absent trials showed no group differences in pupil dynamics, suggesting that p‐tau positivity is not associated with a generalized reduction in pupil responsivity. Thus, the observed attenuation appears most clearly in conditions that require binding of motion and luminance features, rapid monitoring of dynamic stimuli, and speeded decision making together with response preparation and execution, all of which can recruit pupil‐linked arousal systems. ²³ , ²⁶ , ²⁷ , ²⁸ , ⁵⁹ Because pupil dilation reflects the integrated output of decision, arousal, and action‐related processes, the present pattern is consistent with altered neuromodulatory regulation under response‐demanding conditions, rather than a generalized reduction in pupil responsivity.

Despite differences in pupil dynamics, the biomarker groups did not differ in behavioral performance. Participants responded with comparable speed on both simple CRT and visual search tasks and showed uniformly high accuracy across load levels. Sex emerged as a significant predictor of reaction time, with men responding faster than women across set sizes; however, inclusion of sex as a covariate in all models did not alter any pupillometric effects. The absence of behavioral differences supports the interpretation that the observed pupillary effects reflect underlying physiological or neuromodulatory variation rather than performance‐related factors. This interpretation, however, does not imply that cognitive processes are unrelated to pupil dynamics, only that they do not explain the group effects observed here.

A related question is whether group differences in pupil modulation could be attributed to cognitive variability. Because pupil diameter is sensitive to cognitive effort, one possibility is that biomarker‐positive individuals required greater effort to maintain equivalent performance. Multiple analyses argue against this account. Inclusion of delayed visual memory performance as a covariate did not eliminate or attenuate the primary set size × group effects, and cluster timing and magnitude were similar across model specifications. Notably, adjusting for memory performance increased sensitivity to the load‐dependent group pupil effects in the target‐present model that initially failed to reach cluster‐corrected significance. These results indicate that although pupil dynamics did covary with individual differences in memory ability, this cognitive variability does not account for the biomarker‐related modulation effects.

Control analyses addressing gaze behavior similarly suggest that group differences in pupil dynamics do not reflect gaze‐related artifacts or differences in visual exploration strategies. Pupil measurements can be influenced by geometric foreshortening artifacts associated with gaze deviation from screen center. However, gaze eccentricity and dispersion did not differ between groups, and inclusion of trial‐level gaze eccentricity as a covariate did not alter any pupil effects.

The attenuated task‐evoked pupil dynamics in p‐tau+ cognitively normal older adults are consistent with several converging lines of evidence of early involvement of the LC–NE neuromodulatory system in AD. Post mortem and in vivo studies have consistently shown that the LC is among the earliest sites of tau deposition in AD, with such pathology appearing decades before clinical symptoms. ¹ , ² , ³ Functional imaging studies suggest that LC integrity supports attentional flexibility, adaptive gain control, and resistance to distraction, particularly in challenging environments. ³¹ , ⁵⁸ , ⁶⁹ , ⁷⁰ Pupillometry studies in memory and attention paradigms have reported altered pupil responses in individuals with mild cognitive impairment ⁴² , ⁵¹ , ⁵³ , ⁵⁴ , ⁵⁶ or higher AD polygenic risk scores, ⁵² although the direction of effects has varied. The present findings extend this literature by demonstrating disrupted scaling of phasic pupil responses with cognitive load in preclinical AD.

Several LC–NE mechanisms may account for this pattern. Phasic LC responses facilitate transient, task‐locked adjustments in processing, whereas adaptive gain regulates the sensitivity of cortical representations, and arousal reflects broader state‐level activation. Although interdependent, these processes serve distinct functional roles. Attenuated load‐dependent scaling may therefore reflect reduced precision or dynamic range of LC‐mediated modulation rather than a global reduction in pupil reactivity. This interpretation is consistent with models proposing that early tau‐related changes alter LC functional dynamics, potentially weakening coupling between environmental demands and neuromodulatory regulation. ²³ , ²⁴ , ²⁶ , ²⁷

Exploratory individual‐differences analyses further support this neuromodulatory interpretation. Reduced load‐related modulation was associated with higher plasma p‐tau217 levels, a biomarker increasingly recognized as one of the most sensitive indicators of early AD‐related processes. ⁶⁰ , ⁶¹ , ⁶² , ⁶³ The absence of a similar relationship with plasma p‐tau181 levels is notable, but consistent with accumulating evidence that p‐tau217 more precisely captures earlier AD‐specific changes. ⁶¹ , ⁶³ Weaker pupil modulation was also associated with lower inhibitory control and delayed visual memory performance, indicating that pupil dynamics captured behaviorally meaningful variance. Critically, group pupil effects persisted after adjustment for cognition, suggesting partially dissociable physiological and cognitive contributions. These associations should be interpreted cautiously but are consistent with the view that pupil‐linked effects reflect variability in neuromodulatory regulation rather than serving as direct proxies for cognitive capacity.

Recent multimodal imaging studies provide useful context for interpreting the relationships between plasma tau markers and LC‐related measures. One investigation ²² reported that LC integrity was associated with plasma p‐tau217 levels in symptomatic AD participants but not healthy controls, suggesting that LC–p‐tau coupling may vary with disease stage or pathological burden. Another study ⁸ found that higher plasma p‐tau markers were linked to lower integrity of dorso‐rostral LC (with p‐tau231 showing the earliest and most robust association) in a largely cognitively unimpaired lifespan cohort, converging with autopsy evidence of preferential rostral LC vulnerability. Together, these findings indicate that plasma tau biomarkers track with LC system integrity across the AD continuum, although the strength and detectability of these relationships likely vary as a function of pathological stage and measurement sensitivity.

The present findings are consistent with this emerging framework while extending it to functional dynamics. While prior studies have examined the relationship between AD and structural indices of LC integrity, the current study focuses on task‐evoked pupil modulation, a proxy of LC‐mediated neuromodulatory function. Although structural integrity, regional vulnerability, and functional responsivity are not perfectly coupled, particularly in the preclinical phase of AD, it is not surprising that plasma p‐tau217 relates to both LC structural measures and the pupil‐linked functional effects we observed, with variability across studies reflecting differences in measurement modality, sample composition, and disease stage. From this perspective, the biomarker–pupillometry associations we observed complement LC imaging findings and support the interpretation that early AD produces detectable neuromodulatory system changes even when overt cognition remains preserved.

The conjunctive visual search paradigm elicited a robust pattern of load‐ and target‐dependent pupil modulation that closely aligns with prior work. ³³ , ³⁶ , ³⁷ , ³⁸ , ³⁹ , ⁴¹ , ⁴³ , ⁴⁴ , ⁴⁵ , ⁴⁹ Across participants, peak pupil amplitude increased systematically with set size, and half‐max latency shifted later in time. Time‐resolved regression models confirmed large and sustained influences of baseline pupil size, visual load, and target detection. These findings indicate that the task produced well‐characterized pupil signatures of cognitive demand and effectively engaged phasic arousal mechanisms associated with LC activity. The emergence of group differences under conditions requiring feature binding, attentional selection, and speeded decision making suggests that such paradigms may be particularly well suited for probing neuromodulatory dynamics relevant to early AD‐related processes.

Several limitations should be acknowledged. First, although we observed reliable group differences, the sample size was modest and participants were predominantly White, non‐Hispanic, and college educated. Replication in larger and more diverse cohorts is needed to establish the robustness and generalizability of the effects. Second, plasma p‐tau217 and p‐tau181 are sensitive indices of AD‐related biological change, but only indirectly related to tau pathology within the LC. Thus, biomarker–pupillometry associations we observed may reflect downstream neuromodulatory consequences of early AD pathophysiology rather than direct evidence of LC tau deposition. An additional consideration is that p‐tau217 measurements were not available for all participants. Although group classification and primary analyses were not dependent on complete p‐tau217 data, missing p‐tau217 values may have reduced statistical power and limited the precision of the individual‐differences estimates. Future studies with complete biomarker panels across all participants will be important for confirming the stability and strength of these associations. Third, although the pupillary response in p‐tau+ and p‐tau− older adults differed during a conjunctive visual search task that requires feature binding and rapid attentional selection, it is unknown if similar group differences would appear in other tasks that may place different demands on LC–NE neuromodulation such as those that emphasize uncertainty, reward processing, or sustained effort. A broader task battery would help delineate the boundaries of pupil‐based neuromodulatory alterations in preclinical AD. Fourth, task‐evoked pupil dilation represents the integrated output of multiple physiological processes (e.g., sympathetic activation, parasympathetic withdrawal, cortical–subcortical interaction) and does not provide a specific direct measurement of LC activity. Complementary neuroimaging approaches, including neuromelanin‐sensitive MRI or LC–functional MRI, will be needed to more precisely link pupil signatures to specific neural mechanisms. Finally, the cross‐sectional design of our study precludes conclusions regarding temporal progression. Longitudinal follow‐up will be essential to determine whether pupil‐based signatures track disease progression or identify individuals at heightened risk of AD dementia.

In summary, our findings demonstrate that cognitively normal older adults with elevated plasma p‐tau exhibit subtle reductions in load‐dependent pupil modulation during a conjunctive visual search task. These findings are consistent with the hypothesis that early AD is associated with alterations in neuromodulatory dynamics. More broadly, the results highlight the utility of task‐evoked pupillometry as a tool for investigating LC‐related function and its association with biological changes in preclinical AD.

CONFLICT OF INTEREST STATEMENT

Elena K. Festa, William C. Heindel, Camille A. Marangi, Britney Escobedo, Jingjing Zou, and Diane M. Jacobs report no conflicts of interest. David P. Salmon has been a paid consultant for Aptinyx and Biogen. Douglas R. Galasko reports grant support (to UCSD) from the Michale J. Fox Foundation; consulting fees from Eisai, GE Healthcare, Roche Diagnostics, and Cognition Therapeutics; DSMB membership for Artery Therapeutics; and Advisory Board membership for Actinogen. Author disclosures are available in the supporting information.

CONSENT STATEMENT

Written informed consent was obtained from all study participants in accordance with the Declaration of Helsinki. All procedures were approved by the UCSD Institutional Review Board.

Supporting information

Supporting Information

Supporting Information