Visual Working Memory in Older Adults With and Without Early-Onset Depression: An Ecological Momentary Assessment Study

Structured abstract

Objectives:

To use the dot memory task to examine the impact of early-onset depression (ED) on visual working memory (VWM) in later life when compared to those with no history of depression (CON). As an exploratory aim, we examined VWM in remitted vs non-remitted depression when compared to those with no history of depression.

Design, Setting, Participants, and Measurements:

79 participants aged 55-79 underwent lab-based neuropsychological testing. They were then provided with a study-supplied smartphone and were prompted to complete ecological momentary assessments over 14 consecutive days, receiving 3 separate notifications a day. Prompts consisted of a brief survey regarding recent activities, affect ratings, and use of emotion regulation strategies, followed by 3-5 brief cognitive tasks. VWM was measured with a dot memory task. Both response time and error distance were calculated.

Results:

Mean response rate was 70%. The ED group had significantly slower response times at baseline than the CON group (18.9% slower, 95% CI 2.7–37.6%, p=0.023). Across the study period, the ED group's response time decreased faster relative to CONs (−0.8% per day, 95% CI −1.6 to +0.0%, p=0.057), an effect driven by the actively depressed (AD) subgroup, whose response time decreased significantly faster (AD vs CON: −1.3% per day, 95% CI −2.3 to −0.4%, p=0.007).). There was no significant difference in either the level or the rate of change of error distance between CON and ED participants. Within-person variability (standard deviation) in response time did not differ by group or change over time; within-person variability in error distance declined over the study period in CONs but declined significantly less in the ED group (ED × time: 0.02 units/day, 95% CI 0.001–0.039, p=0.037).

Conclusions:

Our findings suggest that the VWM domain may be a malleable target for state-improvements in cognitive functioning in depressed patients.

Introduction

Depression is common in older adults, with a global prevalence of up to 13% (1). Individuals with late-life depression (LLD) often have impaired day-to-day functioning, and worse health outcomes (2). Those with LLD demonstrate cognitive decrements across all domains relative to their never-depressed peers, most pronounced in attention, executive functioning, and processing speed (3). Neuroimaging has found disproportionate disruptions in fronto-subcortical network systems (4). Although cognitive changes are not always present in LLD, its presence often signifies a more severe pathology, with increased impairments in day-to-day functioning, treatment resistance, and higher risk of relapse (5-7). We also know that those with both LLD and MCI are more likely to progress a major neurocognitive disorder, particularly vascular dementia (8).

Recent studies have identified additional LLD cognitive subtypes, including a primarily amnestic subtype (9). The ability to subtype LLD cognitive profiles has the potential to enable identification of groups at increased risk of LLD recurrence, or developing dementia, as well as identify tailored treatment approaches.

There is evidence that VWM is more sensitive to changes in processing speed in older adults and LLD patients than non-visual tasks, as they often require more processing speed and executive functioning demands (10,11). Additionally, research has shown that VWM task performance is among the strongest predictors of LLD symptom severity (12). However, it has also been found that after controlling for deficits in working memory and processing speed more broadly, depression no longer predicts performance on visual-based measures (13,14). VWM and other visual-based tasks have also historically been more criticized in comparison to their non-visual counterparts and are thus not as heavily relied on in general clinical practice (15). Despite this, there is the above-mentioned emergent evidence of VWM as a particularly sensitive domain-specific deficit in LLD. As such, it is important to further study VWM to establish a more complete cognitive profile of LLD, with the aim of establishing more accurate diagnosis, prognosis, and treatment via potentially underutilized VWM tasks.

Though lab-based cognitive testing is comprehensive and well validated, it does have limitations. It can be time and labor intensive and challenging to access, both because of a limited number of testing sites and challenges with mobility and transport. For these reasons, lab-based cognitive testing is not typically repeated over days/weeks. Subjects’ performance can fluctuate day to day and morning to evening, particularly in older adults (16,17). Single or infrequently repeated formal cognitive assessments make it hard to determine the impact of these fluctuations on an individual’s performance. Historically, there were few ways to mitigate these limitations. However, with the advent of portable electronic devices, namely the smartphone, it has become possible to take cognitive testing out of the lab and into the subject’s home environment using ecological momentary assessment (EMA).

Though not as comprehensive as formal lab-based cognitive testing, EMA can provide real-time assessments several times a day, over multiple days in the patient’s home environment (18). EMA has been established as a validated method of measuring VWM in healthy controls (19-23). There are several studies looking specifically at the use of EMA to measure VWM in older adults, both cognitively unimpaired and those with MCI/mild dementia (24-27). Nicosia et al assessed VWM using the GRIDS task in 290 older adults at risk of Alzheimer’s disease. Participants completed ≥1 cycle of brief cognitive assessments up to 4 times a day over 7 consecutive days. They found that EMA testing correlated well with conventional lab-based cognitive assessments and their adherence rate was high at 80.42% (25). This was reflected in the findings of Thompson et al who also used the GRIDS task to measure VWM in 52 cognitively unimpaired older adults. They performed 3 brief cognitive assessments daily for 8 consecutive days and had an average completion rate of 93%. This was maintained across the 8-day testing period with a completion rate of 95% on day 2 and 88% on day 8. Across all 24 sessions within-person average score reliability ranged from 0.89-0.97 (26).

Though there have been studies using EMA to assess VWM in individuals with bipolar disorder, schizophrenia and ADHD, there are no known EMA studies looking at the impact of depression on VWM in any age group (28-30). Our study aimed to use the dot memory task (DMT) to examine the impact of early-onset depression on VWM in later life when compared to those with no history of depression. As an exploratory aim, we examined VWM in remitted depression (REM) vs those who were actively depressed (AD) when compared to participants with no history of depression.

Methods

Participants

79 participants were recruited from Long Island, NY through advertisements and referrals from geriatric psychiatry clinics. The target population was English speaking adults, age 55-79, with and without a history of depression. An interview was conducted to categorize participants as either controls with no history of depression (CON), or as having had ≥1 episode of depression prior to the age of 35 (ED). Within the ED group, individuals were categorized as REM or AD. To be included, individuals were required to have been on the same medication in the 30 days preceding enrolment. Individuals who had severe medical illness, neurological conditions, active substance use, or MRI ineligibility were excluded. To be eligible, participants were required to have English fluency, reading ability, and sufficient sensory capabilities to follow instructions.

Procedure

Participants provided written consent prior to enrollment. Before starting the EMA protocol, they underwent a full diagnostic assessment, neuropsychological test battery, and neuroimaging visit. EMA assessments occurred over 14 consecutive days through a study-supplied smartphone. All participants were tasked with completing assessments three times per day, receiving notifications and reminders to maximize adherence. Notifications arrived at pseudo-random time intervals. The first was within one hour of a predetermined start time that the participant endorsed as “morning” (7:00am-10:00am). Subsequent notifications came in the afternoon and evening, approximately four hours after the previous time point. This resulted in 42 total sessions. Questions were answered on the Mobile Monitoring of Cognitive Change application (“M2C2”) (23). Prompts consisted of a brief survey regarding recent activities, affect ratings, and use of emotion regulation strategies, followed by several brief cognitive tasks. Upon device delivery, each participant was educated as to device operation and completion of EMA procedures. Compensation for participation was $65 granted that their overall response rate was ≥90%. This ≥90% compensation threshold was chosen with the aim of incentivizing participants to achieve a high response rate. The Stony Brook University Institutional Review Board approved the study.

Measures

Both clinical and cognitive measures were utilized throughout the study. The Structured Clinical Interview for DSM-5 Disorders – Research Version (SCID-5-RV) was used to determine current/past history of depression and other comorbidities. This was conducted by a master’s-level research clinician trained by the Principal Investigator, serving as a licensed, supervising psychologist.

The task of interest for the present study was the dot memory task (DMT), one of the available, validated, measures in the broader M2C2 battery. The DMT consisted of three stages: a learning task, a distraction task, and a memory task. First, participants were presented with a 5x5 grid of boxes for three seconds with three boxes containing a red dot. Second, a distraction task was presented where participants were asked to identify all the letter Fs on a screen containing Es and Fs. Third, participants were asked to touch the three boxes that had contained the red dot from the first stage of the task. Performance was assessed along two dimensions: accuracy, quantified as the error distance between the response and the target (Euclidean distance), and processing speed, quantified as response time in milliseconds. Each time the DMT was assigned, participants were prompted to complete it twice. For each dimension we derived two summary metrics across all test blocks: the median, capturing the participant's typical performance, and the standard deviation, capturing the within-subject stability of that performance. The DMT was initially designed as part of the Effects of Stress on Cognitive Aging, Physiology, and Emotion (ESCAPE) project, with Sliwinski et al using the DMT as an ambulatory assessment of VWM, reasoning it was a good measure of working memory as it includes a distraction task. In their study of 219 adults practice effects were not significant, with performance consistent across their 14-day study period. The DMT also correlated well with in-lab assessments of the same domain (23,31).

Statistical analysis

Descriptive visualization.

Prior to formal modeling, we examined the longitudinal structure of the data graphically. For each participant and study day (Day 0–13), we summarized each EMA metric, the median and standard deviation of response time (processing speed) and the median and standard deviation of error distance. We computed the daily mean of each metric by averaging over the three within-day assessments (morning, noon, and afternoon). The computed quantities were displayed as boxplots, with one box per day for each of the three groups (CON, REM, AD) from Day 0 through Day 13, to compare visually the group-level trajectories and their between-subject dispersion for measures of interest. We also visualized the distribution of these four outcome variables through histograms. Inspection of these histograms showed that the two response-time outcomes (response time and its standard deviation) were strongly right-skewed, whereas the two error-distance outcomes were approximately symmetric; we therefore analyzed the response-time outcomes after a natural-logarithm transformation (reporting their effects as percentage changes) and analyzed the error-distance outcomes on their original scale. We did not exclude any observations as outliers; using the within-assessment summary together with the log transformation limited the influence of a small number of extreme response times arising from interrupted assessments. The raw and log-transformed distributions of all four outcomes are shown in Supplementary Figure S1.

Primary longitudinal models.

To investigate depression as a predictor of visual memory performance and its evolution over time, we conducted two sets of regression analyses. In the first analysis, we assumed a linear time trend and fitted a linear mixed-effects model (LMM) including subject group (CON, REM, AD), time (days), and their interaction, with a subject-specific random intercept to account for within-subject correlation. In the second analysis, we relaxed the linearity assumption by using a generalized additive model (GAM) with group-specific smooth functions of time to accommodate nonlinear trajectories, again including a subject-specific random intercept. Each group-specific smooth was represented by a penalized thin-plate regression spline (the default spline basis in R package mgcv), and the smoothing parameters were estimated by restricted maximum likelihood (REML) so that each trajectory's effective wiggliness was determined from the data. The participant-specific random intercept was implemented through a Gaussian random-effect basis. For each smooth we report its approximate F-test p-value, and the parametric coefficients for group and covariates with 95% confidence intervals; fitted group trajectories are displayed with pointwise 95% confidence bands. All models were adjusted for age, sex, years of education, and baseline Montgomery-Åsberg Depression Rating Scale (MADRS) score. The primary longitudinal models were fitted using the lme4 and lmerTest packages for the linear mixed-effects models and the mgcv package for the generalized additive models. All models were fit to momentary, session-level data (one observation per completed EMA administration of the dot memory task), with the subject-specific random intercept accounting for the clustering of repeated assessments within participants; daily summaries were used only for the descriptive visualizations described above. In the GAM, group was additionally included as a parametric main effect alongside the group-specific smooths, because the by-group smooths are mean-centered and therefore capture the shape of each group's trajectory rather than overall between-group differences in level.

Time-of-day effects.

Because three assessments were collected per day, we evaluated whether performance varied systematically with time of day, and whether any such variation evolved over the study or differed by group. Rather than restricting this question to a single day, we entered time of day (morning [reference], noon, afternoon) as an additional covariate in the longitudinal models and assessed its contribution through nested model comparison. Beginning from the primary specification, we sequentially added the time-of-day main effect, its two-way interactions with day and group, and the full three-way time-of-day × day × group interaction, comparing successive models by likelihood-ratio tests (models fitted by maximum likelihood) and the Akaike information criterion (AIC). This comparison was conducted primarily within the LMM framework.

Model diagnostics.

Following each regression analysis, we performed standard diagnostic checks for outliers and for violations of the regression assumptions using the R package DHARMa. DHARMa (Diagnostics for hierarchical regression models) implements a simulation-based residual framework for generalized linear mixed models: it repeatedly simulates responses from the fitted model and, for each observation, computes a scaled residual defined as the position of the observed value within its simulated predictive distribution. Under a correctly specified model these scaled residuals are uniformly distributed, regardless of the model family, which makes the diagnostics interpretable and comparable across the LMM and GAM. We used the uniformity (QQ) plots with associated tests and residual-versus-predictor plots to evaluate model adequacy. Diagnostic plots (uniform QQ plots and residual-versus-predicted plots) for the four primary models are provided in Supplementary Figure S2.

Missing data.

We did not impute missing values; instead, the LMM and GAM are likelihood-based and use all available observations, which yields valid inference under a missing-at-random (MAR) assumption, that is, missingness may depend on observed data (including measured covariates and previously observed responses) but not on the unobserved value itself. We therefore proceed under MAR; we note that if missingness depended on contemporaneous unobserved performance (i.e., missing not at random), estimates could be biased, and this is a limitation inherent to observational EMA designs.

The two patient groups (REM and AD) are relatively small and have previously been combined to improve statistical power; we separated them here to allow a more granular examination of potential group differences. Comparisons involving these subgroups should therefore be interpreted as exploratory. All analyses were performed in R (version 4.5.1).

Results

Baseline characteristics

79 participants were included in the study (see Table 1 for baseline characteristics). 48 were CON and 31 had a history of ED. Of those with a history of ED, 18 were AD. CON, REM and AD had mean ages of 65.7, 62.5 and 64.4 respectively. Most ED participants reported >1 lifetime episode of depression.

Table 1.

Baseline characteristics of participants

Characteristic	Control (n=48)	Remitted (n=13)	Actively depressed (n=18)	p-value
Age, mean (SD)	65.7 (7.10)	62.5 (6.78)	64.4 (5.28)	0.31
Female, n (%)	26 (54.2%)	7 (53.8%)	8 (44.4%)	0.81
Race, n (%)
White	44 (91.7%)	11 (84.6%)	18 (100%)	0.19
Black/African	1 (2.1%)	2 (4.2%)	1 (2.1%)
American
Other	2 (4.2%)	0 (0%)	0 (0%)
Missing	1 (2.1%)	2 (15.4%)	0 (0%)
Education, mean years (SD)	17.6 (8.72)	15.9 (2.63)	15.9 (2.81)	0.66
MADRS, mean (SD)	1.44 (1.67)	2.62 (2.10)	11.3 (6.16)	<0.001

Abbreviations: Montgomery-Asberg Depression Rating Scale (MADRS). p-values are omnibus tests across the three groups: oneway ANOVA for age, education, and MADRS; Fisher's exact test for sex and race. Bold indicates p<0.05.

Participant adherence

The daily mean number of observations across all 79 subjects was 1.51 (SD 1.3) of 3 total possible assessments. The mean non-completion rate was 30% (SD 27%), with most subjects missing <20% of assessments (figure 1). 55.7% of all participants completed ≥1 assessment on each day of the 14-day study period (64.6% of CON subjects, 53.8% of REM and 33.3% of AD) (supplemental table S1).

Figure 1.

Number of observations per subject across study period

Each box represents the number of assessments a single participant completed in one day, with each horizontal line of boxes representing a single participant's assessments across the 14-day study period. Figure 1a represents controls, 1b represents remitted participants, and 1c represents actively depressed participants

Response time

The ED group had a slower overall mean response time than CONs. When separating the ED group into AD and REM participants, the CON group still had the fastest overall response time, and AD the slowest. None of these trends met statistical significance (Figure 2a).

Figure 2. Response time, error distance, and their within-person variability by depression status.

Each panel shows the distribution, across control, remitted, and actively depressed participants, of a per-participant summary averaged over all assessments: (a) response time, (b) error distance, (c) the within-person standard deviation (SD) of response time, and (d) the within-person SD of error distance. The response-time panels (a, c) use a logarithmic y-axis.

For each subject, we calculated daily mean response time. When comparing the ED to the CON group, the CON group was the fastest across the study period. When separating the ED group into AD and REM participants and comparing them to the CON group, we found a trend for decreasing response time over the first ~5 days of the study period in all three groups. Response times then generally plateaued. Across the study period, the CON group remained the fastest and AD patients the slowest. Figure 3a shows the distribution of the mean of each group over the 14-day follow-up period.

Figure 3. Daily response time, error distance, and their within-person variability by depression status.

Each panel shows the distribution of daily per-participant means across the 14-day study period for control, remitted, and actively depressed participants: (a) response time, (b) error distance, (c) the within-person standard deviation (SD) of response time, and (d) the within-person SD of error distance. The response-time panels (a, c) use a logarithmic y-axis.

We formally evaluated time-of-day effects by adding time of day, and its interactions with day and group, to the longitudinal models using all available days. For response time, none of these additions improved model fit (likelihood-ratio p=0.76, 0.98, and 0.67 for the time-of-day main effect, its two-way interactions, and the three-way interaction, respectively; the Akaike information criterion increased at each step), indicating no systematic variation in response time by time of day, and no evidence that any such variation differed by group or changed over the study period. Supplementary Figure S3 shows within-day variations in response time between all three groups at day 0 and across the 14-day study period.

The results of our linear mixed effects model, controlled for age, sex, education, and MADRS score, are shown in Table 2; the response-time outcomes were log-transformed and their effects are expressed as percentage changes. When comparing the ED to the CON group at baseline, the ED group was significantly slower than the CON group (18.9% slower, 95% CI 2.7–37.6%, p=0.023). In the three-group model, both the AD and REM subgroups were numerically slower than CONs at baseline, but neither difference reached statistical significance (AD vs CON: 16.4% slower, 95% CI −6.4 to 44.6%, p=0.175; REM vs CON: 16.3% slower, 95% CI −1.7 to 37.6%, p=0.081).

Table 2.

Regression analysis adjusting for potential confounders - Linear mixed models

Predictor	Response time (% change)†			Error distance (units)			SD of response time (% change)†			SD of error distance (units)
	Estimate	95% CI	p	Estimate	95% CI	p	Estimate	95% CI	p	Estimate	95% CI	p
Age (years)	+1.9%	+1.0 to +2.7%	0.007 *	0.027	−0.001 to 0.054	0.062	+1.5%	−0.2 to +3.2%	0.093	0.001	−0.010 to 0.011	0.910
Sex (male)	−4.9%	−14.7 to +5.9%	0.36	−0.260	−0.616 to 0.096	0.157	−8.4%	−25.9 to +13.4%	0.43	−0.007	−0.142 to 0.128	0.917
Education (years)	−2.0%	−4.3 to +0.4%	0.10	−0.036	−0.114 to 0.043	0.377	−3.8%	−8.2 to +0.8%	0.11	−0.014	−0.043 to 0.016	0.369
MADRS	+0.6%	−1.0 to +2.3%	0.46	0.015	−0.040 to 0.069	0.603	+1.1%	−2.3 to +4.6%	0.54	0.015	−0.007 to 0.037	0.179
Group difference at baseline: REM vs CON	+16.3%	−1.7 to +37.6%	0.081	0.282	−0.256 to 0.821	0.307	+19.2%	−18.1 to +73.6%	0.36	−0.107	−0.359 to 0.145	0.407
Group difference at baseline: AD vs CON	+16.4%	−6.4 to +44.6%	0.18	0.156	−0.550 to 0.863	0.666	+7.1%	−32.2 to +69.4%	0.77	−0.192	−0.492 to 0.109	0.214
Rate of change over time (CON, reference)	−0.8%	−1.3 to −0.4%	<0.001 *	−0.049	−0.062 to −0.035	<0.001 *	+1.3%	−0.3 to +2.9%	0.11	−0.019	−0.031 to −0.008	0.001 *
Group × time interaction: REM vs CON	−0.1%	−1.1 to +1.0%	0.91	0.025	−0.005 to 0.055	0.106	−1.7%	−5.1 to +1.7%	0.32	0.021	−0.004 to 0.047	0.101
Group × time	−1.3%	−2.3 to −0.4%	0.007 *	−0.004	−0.03	0.778	−0.7%	−3.8 to +2.5%	0.66	0.019	−0.00	0.110
interaction AD vs CON					1 to 0.024						4 to 0.042

Abbreviations: confidence interval (CI); CON, control; REM, remitted; AD, actively depressed; MADRS, Montgomery-Asberg Depression Rating Scale.

^†Response-time outcomes were natural-log transformed; estimates are back-transformed and expressed as the percentage change in response time per unit increase in the predictor (per day for time, per group contrast for group terms). Error-distance outcomes are on their original scale. CON is the reference group; baseline group rows are group main effects and Group × time rows are the group-by-time interaction terms.

^*p<0.05.

When comparing the rate of change in response time across the study period, the ED group's response time decreased faster relative to CONs, a difference that was marginal at the combined-group level (−0.8% per day, 95% CI −1.6 to +0.0%, p=0.057) and was driven by the AD subgroup: AD response time decreased significantly faster than that of CONs (−1.3% per day, 95% CI −2.3 to −0.4%, p=0.007), whereas the REM subgroup did not differ from CONs (−0.1% per day, 95% CI −1.1 to +1.0%, p=0.880). Note, response time in the CON (reference) group decreased significantly over time (−0.8% per day, 95% CI −1.3 to −0.4%, p<0.001). Of the factors adjusted for, only age was associated with response time, with each additional year of age associated with a 1.9% slower response time (95% CI 1.0-2.7%, p<0.001). Within-person variability in response time (Figures 2c and 3c) did not differ significantly between groups at baseline and did not change significantly over the study period in any group (Table 2).

The results of our non-linear mixed effects model, controlled for age, sex, education, and MADRS, are displayed in Table 3. The time trajectory of response time was significant in all three groups (CON p<0.001; REM p=0.023; AD p<0.001) and in the combined ED group (p<0.001), with the largest change occurring in the AD subgroup. Neither the REM nor the AD group differed significantly from CON in overall response-time level (REM 15.7% slower, p=0.064; AD 6.7% slower, p=0.544). In this model, only increasing age was significantly associated with a slower response time. The null results for within-person variability in response time remain unchanged when GAMs were used to fit the data (Figure 4c and Table 3)

Table 3.

Regression analysis adjusting for potential confounders - Non-linear (generalized additive) models

Predictor	Response time (% change)†			Error distance (units)			SD of response time (% change)†			SD of error distance (units)
	Estimate	95% CI	p	Estimate	95% CI	p	Estimate	95% CI	p	Estimate	95% CI	p
Age (years)	+1.9%	+1.0 to +2.8%	<0.001 *	0.027	−0.001 to 0.054	0.058	+1.5%	−0.2 to +3.2%	0.084	0.001	−0.010 to 0.011	0.910
Sex (male)	−5.1%	−14.8 to +5.8%	0.347	−0.261	−0.616 to 0.095	0.152	−8.4%	−26.0 to +13.3%	0.419	−0.007	−0.142 to 0.128	0.917
Education (years)	−2.0%	−4.3 to +0.4%	0.102	−0.036	−0.114 to 0.043	0.375	−3.7%	−8.1 to +0.9%	0.111	−0.014	−0.043 to 0.016	0.365
MADRS	+0.6%	−1.0 to +2.3%	0.453	0.015	−0.040 to 0.070	0.591	+1.1%	−2.3 to +4.6%	0.532	0.015	−0.007 to 0.037	0.176
Group difference (overall level): REM vs CON	+15.7%	−0.9 to +34.9%	0.064	0.434	−0.070 to 0.939	0.092	+6.7%	−21.3 to +44.7%	0.68	0.028	−0.165 to 0.222	0.77
Group difference (overall level): AD vs CON	+6.7%	−13.5 to +31.6%	0.544	0.127	−0.561 to 0.815	0.72	+2.3%	−32.6 to +55.3%	0.92	−0.072	−0.339 to 0.195	0.60
Time trajectory: CON			<0.001 *			<0.001 *			0.115			0.001 *
Time trajectory: REM			0.023 *			0.071			0.163			0.854
Time trajectory: AD			<0.001 *			<0.001 *			0.703			0.984

Abbreviations: CON, control; REM, remitted; AD, actively depressed; MADRS, Montgomery-Asberg Depression Rating Scale.

^†Response-time outcomes were natural-log transformed; parametric estimates are back-transformed to percentage change. Each “Time trajectory” row reports the approximate F-test p-value for the corresponding group-specific penalized thin-plate smooth of study day; a significant value indicates a non-flat trajectory over time. Models included a subject-specific (factor) random intercept via a Gaussian random-effect basis, with smoothing parameters estimated by REML.

^*p<0.05.

Figure 4. Fitted curves showing response time, error distance, and their within-person variability by depression status.

Fitted generalized additive model curves, with pointwise 95% confidence bands, by depression status across the study period for (a) response time, (b) error distance, (c) the within-person standard deviation (SD) of response time, and (d) the within-person SD of error distance. The response-time panels (a, c) are shown on the original millisecond scale after back-transformation from the log scale.

This non-linear change is reflected in Figure 4a, showing fitted curves of response time by depression status across the study period. At the start of follow-up, both the REM and AD groups had slower response times than CONs. Response times in the CON and AD groups decreased over the first several days; the AD group, which began among the slowest, decreased most steeply and continued to improve, overtaking the CON group in the latter part of the study to become the fastest group by the end of follow-up. The REM group remained the slowest throughout, decreasing until approximately the middle of the study period before rising again toward the end.

Error distance

The CON group had the lowest overall mean error distance, but this did not meet statistical significance. This was also the case when separating AD and REM sub-groups and comparing them to the CON group (Figure 2b).

For each subject, we calculated the daily mean error distance. There was a trend for decreasing error distance over time. Across the study period the CON group had a lower error distance when compared to the ED group, though this did not reach statistical significance. This was also the case when separating AD and REM sub-groups and comparing them to the CON group (Figure 3b).

We formally evaluated time-of-day effects for error distance by adding time of day, and its interactions with day and group, to the longitudinal model using all available days. None of these additions improved model fit (likelihood-ratio p=0.40, 0.17, and 0.11 for the time-of-day main effect, its two-way interactions, and the three-way interaction, respectively; the Akaike information criterion favored the model without time of day), indicating no systematic within-day variation in accuracy. Supplementary Figure S4 shows within-day variations in error distance between all three groups at day 0 and across the 14-day study period

The results of our linear mixed effects model are shown in table 2. At baseline, we found no significant difference in error distance between groups. When comparing the change in error distance across the study period, we found that the error distance in CONs decreased significantly (−0.05, 95% CI −0.06–−0.04, p=<0.001). We found no significant difference in the ED group over time when compared to CONs. This was also the case when separating AD and REM sub-groups and comparing them to the CON group. No covariates were associated with a significant difference in error distance.

The results of our non-linear mixed effects model are seen in table 3. Neither any covariate nor any overall between-group difference in error-distance level reached significance (REM vs CON p=0.092; AD vs CON p=0.72). The time trajectory of error distance changed significantly in the CON group (p<0.001) and in the combined ED group (p<0.001); when separating the subgroups, the AD trajectory was significant (p<0.001) while the REM trajectory showed a non-significant trend (p=0.071). Within-person variability in error distance (Figures 2d, 3d and 4d) decreased significantly over the study period in the CON group (in both the linear and non-linear models), with the combined ED group showing significantly less of this reduction in the linear model (ED × time interaction, p=0.037); the corresponding REM and AD interactions were in the same direction but did not reach significance (Tables 2 and 3).

This non-linear trend is reflected in Figure 4b, with fitted curves showing error distance across the study period. All groups had their highest error distance at day 0, and the REM group had the highest error distance throughout follow-up. Error distance in the CON group decreased steadily, reaching its lowest level toward the end of the study period. The REM group decreased until approximately day 8 and then rose again toward the end of follow-up, whereas error distance in the AD group decreased steadily across the study period. The differences between the three groups did not reach statistical significance.

Model diagnostics

Simulation-based residual diagnostics (Supplementary Figure S2) were generally satisfactory for all four primary models. The uniform quantile-quantile plots showed the scaled residuals closely following their expected distribution, with only minor departures (Kolmogorov-Smirnov D=0.05-0.09); although these departures were statistically significant given the large number of observations (n=2,317), their magnitude was small. There was no evidence of overdispersion (dispersion ratio approximately 1.02 in every model, all p>0.5), and the residual-versus-predicted plots showed no systematic trend or heteroscedasticity. The natural-log transformation markedly improved the residual distribution of the two response-time outcomes (Supplementary Figure S1); the largest residual departure was seen for the standard deviation of error distance (D=0.09), consistent with the bounded and mildly skewed nature of that outcome.

Discussion

Individuals with no history of depression were found to have quicker response times than those with a history of early-onset depression. This is consistent with well-established literature that has shown prolonged response times in individuals with depression compared to non-depressed individuals (32,33). Previous studies have also shown a similar discrepancy in response time on visual memory tasks as well as other cognitive domains (e.g. psychomotor, verbal working memory, attention) (34,35). Given this, deficits in response time (as a measure of processing speed) may reflect a broader pattern of cognitive dysfunction, rather than an isolated impairment. Alternatively, slower processing speed may directly impact other aspects of cognitive functioning as a side effect. This was first theorized by Salthouse, who hypothesized that slowed processing speed results in worsening cognitive performance as the “relevant operations cannot be successfully executed (limited time) and because the products of early processing may no longer be available when later processing is complete (simultaneity)” (36).

Despite an initial slower response time, the AD sub-group had more rapid improvement, overtaking both the CON and REM groups in the latter part of the study period and becoming the fastest group by the end of follow-up. There are several possible reasons for this. AD subjects may have been on treatment that improved their depressive symptoms over the study period, or engaging in the tasks themselves may have improved participant mood or activation energy. AD individuals may have a similar response time to REM or CONs on task repetition, with only initial performance being slower. This finding demonstrates an advantage of EMA studies, as this trend would not have been observed had we been performing this testing in a lab-based setting with limited repetition.

The initially slower mean response time of the AD group did not appear to influence the response accuracy compared to the non-depressed group. This contradicts prior literature which has consistently demonstrated that depressed individuals have higher error rates on tasks within several cognitive domains when compared to healthy controls. However, a meta-analysis found that certain tests (e.g., visuomotor speed, immediate and delayed visual memory) are not able to distinguish between depressed and non-depressed individuals (37). Some of these tests included in this meta-analysis had a visual component to them. As such, the results of the current study could reflect the visual domain of cognitive functioning being more resilient to the downstream cognitive effects of depression relative to others.

We found no significant fluctuations in either response time or error distance by time of day when these were modeled across all study days. An exploratory analysis restricted to the first study day had suggested better accuracy later in the day in the CON and AD groups, but this did not persist in the longitudinal models. Prior studies in this area have been mixed. Our finding goes against earlier non-EMA studies, which found a trend for faster processing speeds in the morning in older adults, though this difference has been found to disappear with task repetition (38,39). An earlier EMA study on older adults by Wilks et al. found no significant effect on time of day on performance of the GRIDS test (a measure of spatial working memory) across a 7-day study period (40).

We had a good overall mean response rate of ~70%. This is slightly lower than earlier EMA studies in older adults measuring VWM, with mean adherence rates of 81-93% (24-26). This may be explained by our study population including participants with AD. This group was the least likely to complete ≥1 assessment a day at a rate of only 33%, vs 64.6% in CONs. This difference may have been a result of a lack of motivation and/or interest to complete the EMA tasks, both common symptoms of depression (41).

There are limitations to our findings. Though the DMT has been found to correlate well with lab-based cognitive testing, the two assessments can’t be directly compared as, by their nature, the testing environments are not the same (i.e. ecological vs lab-based) (23). We did not control for participant’s technological literacy, which could have affected task performance, however there is no reason to expect that one group would have been more technologically literate than another. In addition, each participant received education to ensure that they were able to operate their device after receiving it, and complete tasks appropriately. In terms of adherence, our mean non-completion rate was 30%, while participants were only compensated by reaching a completion rate of ≥90%. If a participant had low initial adherence, they may have been less incentivized to continue to complete assessments if they knew they would fall below this ≥90% threshold. Further thought needs to be given as to how we should be financially compensating participants in these studies to avoid this potential disincentive. We did not control for the presence of apathy, which could have affected participant task completion rate and task engagement. To try to mitigate this, we excluded those with amnestic MCI to try to prevent the inclusion of those in the early stages of a neurodegenerative condition, a population at increased risk of apathy without depression (42). We also did not directly assess effort or task engagement, which could have affected both outcome measures. However, there is evidence to suggest when comparing adults with and without depression, motivation (i.e. boredom, fatigue, or a lack of concentration) does not significantly impact neurocognitive task performance, but rather it is an individual’s feeling of wellbeing during assessment (i.e. negative feeling around the test material or situation) that affects performance (43). Future studies should be developed to include a task engagement metric that is integrated into each test. In addition, our EMA protocol did not capture contextual information such as the participant's location (home vs. away), social setting (alone vs. with others), or interruptions and distractions during assessments; because such factors are known to influence momentary cognitive performance, we were unable to adjust for them, and future EMA protocols in this population should collect them.

Cognitive deficits associated with depression are both state- and trait-dependent, with improvements observed over this 14-day study in all groups possibly only reflecting a state-improvement due to practice effects (i.e., task familiarity because of repeated task exposure), which we did not adjust for in our analysis. Though the depressed group “caught up” to the non-depressed group in terms of reaction time over repeated assessments, cognitive improvements in depression treatment are known to be inconsistently maintained. Future research should be done to explore if these results are maintained after a period without constant task repetition, or on introducing novel VWM task at the end of the study to determine if improvements are transferable (44). Though participants were required to have been on the same medication regimen in the 30-days prior to study enrollment, we did not control for which psychiatric and non-psychiatric medications participants were taking, nor did we control for any medication changes in the period between enrollment and the EMA start date, or during the 14-day study period. As actively depressed participants were, by definition, more symptomatic, they may have been more likely to have a change in medication than other groups, though we do not have the data to assess this. As we did not control what treatment participants were receiving at the time of the study, it is possible that the improvements shown by the depressed group were secondary to concurrent pharmacological/non-pharmacological treatment. Participants may also have been on medication that could have negatively impacted both their reaction time and accuracy, for example any sedating medication such as Benadryl. Our sample size included only 13 REM and 18 AD participants, making it harder to draw firm conclusions from the results of our depression status sub-analysis. Future research should embed the current protocol into a larger randomized controlled treatment trial to determine what aspects of treatment may lead to improvements in specific cognitive domains.

While VWM functioning has previously been studied in several psychiatric populations, it has not been thoroughly explored in older adults with depression. The current study found that individuals with a history of early-onset depression had slower initial reaction times on a VWM task compared to individuals without a history of depression. AD individuals initially had the slowest reaction times but improved more rapidly and had the fastest reaction times by the end the study period. Our study findings suggest that the VWM domain may be a malleable target for state-improvements in cognitive functioning in depressed patients. As such, this study helps further identify cognitive domains that can effectively be targeted for improving cognitive function in depressed patients, despite the increased likelihood of cognitive remission post-treatment.

Highlights.

• What is the impact of early-onset depression on visual working memory in later life when compared to those with no history of depression?

• When completing the dot memory task as part of an ecological momentary assessment, those with a history of depression had a significantly slower response time at baseline than controls with no history of depression. Across the 14-day study period, response times of those with a history of depression decreased faster relative to controls, with this rate of change driven by actively depressed participants.

• The visual working memory domain may be a malleable target for improvements in cognitive functioning in actively depressed patients.

Data Sharing Statement:

The data is not publicly available but can be requested from the corresponding author.

Data Statement

These data have been presented at the 2025 Annual Meeting of the Gerontological Society of America, and the 2026 Annual Meeting of the American Association of Geriatric Psychiatry. Data from this study analyzing performance in other cognitive tasks (the Symbol Match/Search Task, and the Go/No-Go Task) has been published in the Journal of Affective Disorders.