Higher Amyloid and Tau Burden Is Associated With Faster Decline on a Digital Cognitive Test

ABSTRACT

Objective

A 2‐min digital clock‐drawing test (DCTclock) captures more granular features of the clock‐drawing process than the pencil‐and‐paper clock‐drawing test, revealing more subtle deficits at the preclinical stage of Alzheimer's disease (AD). A previous cross‐sectional study demonstrated that worse DCTclock performance was associated with higher Aβ and tau burden in older cognitively normal (CN) participants. This study investigates whether longitudinal changes in DCTclock performance are associated with amyloid‐β (Aβ) and tau burden in preclinical AD.

Methods

A total of 219 CN participants completed baseline and follow‐up DCTclock assessments, baseline Aβ ([¹¹C]PiB) and tau ([¹⁸F]Flortaucipir) PET imaging. Global Aβ and regional tau burden were estimated. Linear mixed models examined associations between longitudinal DCTclock and (1) Aβ, (2) tau, and (3) Aβ and tau burden, adjusted for age, sex, and education. Cognitive domain‐specific performance and fine‐grained features of DCTclock were analyzed.

Results

Elevated baseline Aβ or tau was most strongly associated with accelerated decline in DCTclock performance, particularly in the Information Processing cognitive domain, with stronger associations noted for tau burden. The associations were driven by pen‐stroke latency‐related features. Participants without elevated Aβ or tau burden demonstrated improved performance in these latency features, suggesting practice effects.

Interpretation

Longitudinal declines in DCTclock performance, especially in Information Processing involving speed and executive function, were linked to early Aβ and tau burden in preclinical AD. These findings highlight the potential of digital cognitive assessment tools for tracking disease progression and assessing therapeutic efficacy in clinical trials.

1. Introduction

Alzheimer's disease (AD) is a multifaceted progressive neurodegenerative disorder characterized by the accumulation of tau tangles and amyloid‐β (Aβ) plaques, both of which begin many years before noticeable cognitive deficits are detectable with traditional neuropsychological assessments [1, 2, 3]. Recent advancements in digital cognitive assessments have introduced novel methods for detecting cognitive deficits. One such approach utilizes a digital pen device, which has been shown to correlate with baseline levels of Aβ and tau pathologies as quantified by positron emission tomography (PET) in older cognitively normal (CN) individuals [4]. This study extends prior findings by examining whether longitudinal cognitive performance, assessed via digital pen technology, tracks early disease progression in relation to baseline levels of pathologies.

Developing tools that are scalable, accessible, and easy to administer could broaden their use across underserved communities. Such tools, sensitive to early cognitive decline related to underlying pathophysiology, could offer a complementary framework for comprehensive preclinical detection and monitoring of AD. Traditional paper‐and‐pencil neuropsychological assessments, often designed for impaired populations, showed limited cross‐sectional associations with neuroimaging biomarkers in preclinical AD, though stronger correlations have been observed longitudinally [1, 5, 6, 7, 8, 9]. Digital cognitive assessments, particularly a digitized version of the clock drawing test (DCTclock), have shown promise in detecting AD‐related cognitive deficits related to the underlying pathophysiology [10]. Using a digitizing pen and machine learning algorithms, DCTclock captures precise time‐stamped pen movements, providing a summary score targeting multi‐domain cognitive performance, including Drawing Efficiency, Information Processing, Spatial Reasoning, and Simple Motor function [11, 12, 13]. These scores include unique features reflecting hesitation and speed during the drawing process, many of which are unnoticeable in pencil‐and‐paper clock drawing tests. This test is an FDA‐listed Class II medical device to aid in the assessment of a patient's level of cognitive impairment, regardless of etiology.

Research on DCTclock has primarily focused on its ability to detect cognitive impairment in ad [ 14, 15, 16] and other neurodegenerative disorders [17, 18] or its associations with other neuropsychological assessments [19, 20]. Limited studies have explored its relationship with underlying pathophysiology. Recent cross‐sectional studies have reported that worse DCTclock performance was associated with MRI‐based brain atrophy in individuals with mild cognitive impairment (MCI) [21] and with higher global Aβ and entorhinal tau burden in older CN individuals [4]. These associations were stronger than those observed using traditional tests, such as the extended Preclinical Alzheimer Cognitive Composite (PACC5) [22]. The DCTclock's ability to detect subtle executive function deficits, such as reasoning and problem‐solving [23], may make it a more sensitive tool for capturing early AD‐related changes. However, the longitudinal performance of DCTclock in detecting progressive cognitive decline, particularly in relation to Aβ and tau burden, has yet to be fully established. Previous studies have focused on memory decline in CN individuals with elevated Aβ [6, 7, 8].

This study extends previous cross‐sectional findings by evaluating whether longitudinal DCTclock performance tracks cognitive decline in preclinical AD in comparison to conventional neuropsychological assessments. Specifically, we hypothesized that baseline elevation of Aβ and tau burden would predict longitudinal DCTclock changes. These results may support the use of DCTclock as a biomarker for monitoring disease progression and evaluating therapeutic efficacy in preclinical AD.

2. Methods

2.1. Study Participants

This longitudinal study involved 219 cognitively normal (CN) participants recruited from the Harvard Aging Brain Study (HABS). All participants had a Clinical Dementia Rating (CDR) score of 0 at baseline, indicating no clinical signs of dementia [24]. Informed consent was obtained, and the study was approved by the Mass General Brigham Human Research Committee.

2.2. Conventional Neuropsychological Assessments

Participants underwent standard neuropsychological assessments, including PACC5 [22]. The PACC5 included Logical Memory Delayed Recall (LMDR), the Free and Cued Selective Reminding Test (FCSRT), the Mini‐Mental State Examination (MMSE), the Digit Symbol Substitution Test (DSST), and Category Fluency (CAT). Cognitive composites for memory, executive functioning, and processing speed were also computed using previously described methods in a subset of participants (n = 181) [25]. Briefly, the memory composite included FCSRT‐Free Recall, the 6‐trial Selective Reminding Test Delayed Recall, and the Logical Memory–Delayed Recall; the executive function composite included Trail‐Making Test B, Phonemic Fluency, and Letter‐Number Sequencing; and the processing speed composite included Trail‐Making Test A and DSST.

2.3. Digital Clock Drawing Test (DCTclock)

Participants underwent DCTclock assessments (www.linushealth.com) at baseline and annually over 6.5 years (average follow‐up interval: 14.5 ± 6.4 months). The DCTclock uses a digital pen (Linus Health) to record precise time‐stamped spatial coordinates every 12 ms during clock drawing tasks [4, 11]. Participants were instructed to “draw the face of a clock, put in all numbers, and set the hands to 10 after 11” (command version) and to copy a pre‐drawn clock (copy version). A total of 44 DCTclock features were calculated, reflecting both the process and outcome of the drawing task, such as stroke speed, latency, and spatial organization. These features were grouped into four cognitive domains: Drawing Efficiency, Information Processing, Spatial Reasoning, and Simple Motor, and subscores for each domain were calculated. The combined scores from the command and copy tasks were used to generate a DCTclock summary score, which ranged from 0 to 100, with higher scores indicating better performance (Table 1).

TABLE 1.

Basis of DCTclock features selected as most clinically relevant for AD diagnosis in the identification of DCTclock summary scores [4, 11].

Cognitive domains	DCTclock features
Drawing efficiency	Stroke count conformity (deviation from the expected number of pen strokes)
	Total time
	Ink length
	Drawing size
	Drawing process efficiency (combines ink length and total time)
	Noise (overwriting or non‐standard pen strokes)
Information processing	Percent think time (time spent holding the pen but not actively drawing)
	Average latency (average duration of the delays between each pen stroke)
	Latency variability (variability in the delays)
	Relative long latency (combined duration of delays that are long compared to the participant's average performance)
	Long latency count (total number of delays that are longer than the participant's average performance)
	Longest latency
Simple motor	Percent ink time (time spent actively drawing)
	Average speed
	Max speed
	Initiation speed
	Termination speed
	Oscillatory motion (deviation from smooth motion)
Spatial reasoning	Clockface circularity
	Component placement (spatial relationship of clock components)
	Vertical spatial placement
	Horizontal spatial placement

2.4. Neuroimaging

Participants underwent amyloid‐β (Aβ) PET imaging using [¹¹C]Pittsburgh compound B (PiB, n = 219) and tau PET imaging using [¹⁸F]Flortaucipir (FTP, n = 162) on a Siemens ECAT HR+ scanner and structural MRI (n = 219). All PET data were collected and processed according to prior protocols [3]. Briefly, reconstructed and motion‐corrected PET images were coregistered to the MRI images. Regional parcellations were generated using FreeSurfer (version 6) using the MRI images and brought into the PET space. Global Aβ burden was quantified as distribution volume ratios (DVR, Logan graphical method) across neocortical regions, including the frontal, lateral temporal, parietal, and retrosplenial regions, using cerebellar gray matter as the reference region. Binary Aβ status was determined using a DVR cutoff of 1.185, as previously described.

Regional tau burden quantified as standardized uptake value ratios (SUVR, 75–105 min post‐injection) in early‐tau deposition regions, including the entorhinal cortex and inferior temporal cortex, using the cerebellar gray matter as the reference region. Partial volume correction was applied to the regional tau SUVR values.

2.5. Statistical Analyses

2.5.1. Baseline Analyses

Analysis of variance (ANOVA) and Cohen's d were used to examine the differences in demographic variables, PACC5 scores, cognitive composite scores, DCTclock summary scores, and domain‐specific subscores, and regional tau burden between Aβ+ and Aβ− participants. Regression analyses examined associations between baseline DCTclock summary scores and domain‐specific subscores with PACC5, cognitive composite scores, and Aβ and tau burden.

2.5.2. Longitudinal Analyses

Linear mixed‐effect (LME) models were used to evaluate the associations between longitudinal DCTclock performance and baseline Aβ and tau burden, adjusted for age, sex, education, and handedness:

$$\begin{array}{c}\text{DCTclock}~\mathrm{age}+\mathrm{sex}+\text{education}+\text{handedness}+\mathrm{PET}+\text{inter‐test time}\hfill \\ +\mathrm{PET}\times \left(\text{inter‐test time}\right)+\mathrm{age}\times \left(\text{inter‐test time}\right),\text{random}=~1\hfill \\ +\text{inter‐test time}\mid \text{subject}\hfill \end{array}$$

where inter‐test time is the time difference between the DCTclock baseline and follow‐up visits.

Separate LME models examined associations between longitudinal DCTclock and (1) global Aβ burden, (2) regional tau burden, and (3) global Aβ and regional tau burden. Separate models first examined DCTclock summary scores and domain‐specific subscores. Exploratory analyses then identified specific DCTclock features within the domains that were most strongly associated with PET imaging outcomes to minimize type 1 errors. LME models examined associations between longitudinal changes of DCTclock and conventional neuropsychological assessments to evaluate the added value of DCTclock for capturing cognitive decline. Significance was assessed at p = 0.05, and no multiple comparison correction was performed.

3. Results

3.1. Demographic Characteristics

Demographic information for the 219 CN participants who underwent baseline and follow‐up DCTclock assessments is summarized in Table 2. A total of 62 participants (28%) were Aβ+. No significant differences in age, sex, years of education, handedness, or average follow‐up period were observed between the Aβ+ and Aβ− groups. The Aβ+ group showed slightly lower PACC5 scores compared to the Aβ− group (β = −0.05, SE = 0.03, p = 0.12). Additionally, 181 participants underwent neuropsychological assessments for cognitive composites of executive function, memory, and processing speed, with no significant differences between the Aβ+ and Aβ− groups in these composite scores. As expected, the Aβ+ group exhibited significantly higher entorhinal and inferior temporal tau burden than the Aβ− group (β = 0.17, SE = 0.03 for entorhinal; β = 0.16, SE = 0.03 for inferior temporal, p < 1e⁻⁷). Out of the 219 CN participants, 34 individuals (15.5%; 17 Aβ+ and 17 Aβ−) progressed to a CDR of 0.05 over the follow‐up interval. Individuals who progressed to a CDR of 0.05 showed significantly higher baseline global Aβ burden (β = 0.12, SE = 0.04, p < 0.001) and entorhinal and inferior temporal tau burden (β = 0.20, SE = 0.05, p < 0.001 for entorhinal; β = 0.06, SE = 0.03, p = 0.03 for inferior temporal) compared to individuals who stayed at CDR of 0.

TABLE 2.

Characteristics for n = 219 cognitively normal participants with longitudinal DCTclock.

	CN Aβ− (n = 157)	CN Aβ+ (n = 62)
Sex (% female)	57%	60%
Age, years	78.1 ± 8.6 (45 to 93)	79.1 ± 8.0 (52 to 96)
Education, years	16.3 ± 2.9 (8 to 21)	16.5 ± 2.8 (8 to 21)
Handedness (% right‐handed)	81%	85%
PACC5	0.13 ± 0.76 (−2.4 to 1.9)	−0.17 ± 0.82 (−2.6 to 1.8)
Executive composite	0.17 ± 0.77 (−1.6 to 2.1), n = 134	0.22 ± 0.88 (−2.1 to 1.9), n = 47
Processing speed composite	−0.09 ± 1.00 (−2.8 to 2.7), n = 134	−0.22 ± 0.99 (−2.0 to 2.4), n = 47
Memory composite	0.46 ± 0.82 (−1.5 to 2.3), n = 134	0.20 ± 0.97 (−1.6 to 2.4), n = 47
CDR = 0.05 at follow‐up	n = 17 (10.8%)	n = 17 (27.4%)
Baseline DCTclock summary scores
Summary score (0–100)	72.7 ± 22.6 (5 to 100)	65.2 ± 22.3 (2 to 98)
Baseline DCTclock domain‐specific subscores (z‐scores)
Drawing efficiency (e.g., total time, drawing size)	−0.02 ± 1.12 (−5.0 to 1.7)	−0.04 ± 1.17 (−4.6 to 1.8)
Information processing (e.g., percent thinking time, latency)	−0.01 ± 0.98 (−4.2 to 1.5)	−0.03 ± 1.04 (−3.6 to 1.7)
Simple motor (e.g., oscillatory motion, speed)	−0.12 ± 1.15 (−5.0 to 1.8)	−0.25 ± 1.11 (−3.6 to 1.5)
Spatial reasoning (e.g., clockface circularity, clock component placement)	−0.19 ± 1.33 (−4.5 to 1.2)	−0.47 ± 1.40 (−4.7 to 1.3)
DCTclock average longitudinal followup duration	15.4 ± 7.7 (5.5 to 71.7) months ≥ 1 follow‐up: n = 157 ≥ 2 follow‐up: n = 104 ≥ 3 follow‐up: n = 72 ≥ 4 follow‐up: n = 47 ≥ 5 follow‐up: n = 24	13.9 ± 5.5 (5.7 to 30.4) months ≥ 1 follow‐up: n = 62 ≥ 2 follow‐up: n = 42 ≥ 3 follow‐up: n = 23 ≥ 4 follow‐up: n = 14 ≥ 5 follow‐up: n = 4
PET imaging outcomes
[11C]PiB (threshold = 1.19)
Number of subjects	n = 157	n = 62
Time differences from baseline DCTclock	15.5 ± 21.3 (−29 to 103) months	12.7 ± 16.5 (−26 to 76) months
DVR in FLR composite	1.08 ± 0.05 (0.97 to 1.18)	1.50 ± 0.22 (1.19 to 2.20)
[18F]Flortaucipir
Number of subjects	n = 120	n = 42
Time differences from baseline DCTclock	13.0 ± 17.1 (−28 to 78) months	12.8 ± 10.1 (−14 to 36) months
Entorhinal SUVR75‐105 (PVC)	1.32 ± 0.24 (0.42 to 2.00)	1.59 ± 0.40 (1.04 to 2.47)
Inferior Temporal SUVR75‐105 (PVC)	1.46 ± 0.15 (1.11 to 1.84)	1.60 ± 0.22 (1.17 to 2.25)

Note: Continuous values are expressed as mean ± standard deviation (min–max).

Abbreviations: Aβ, amyloid‐β; CDR, clinical dementia rating; CN, cognitively normal; DCTclock, digital clock drawing test; DVR, distribution volume ratio (using Logan graphical method and cerebellar gray matter reference region); FLR, frontal, lateral temporal and retrosplenial composite region for estimating global Aβ burden; PACC, preclinical Alzheimer's cognitive composite; PiB, [¹¹C]Pittsburgh‐Compound‐B; PVC, partial volume correction using geometric transfer matrix algorithm; SUVR, standardized uptake value ratio (using cerebellar gray matter reference region).

3.2. Baseline Associations Between DCTclock, Conventional Neuropsychological Assessments and AD Biomarkers

At baseline, worse DCTclock summary scores were significantly associated with older age (β = −0.87, SE = 0.2, p < 0.001, Figure 1A) and fewer years of education (β = 1.4, SE = 0.5, p = 0.006), but not with sex or handedness. The DCTclock summary scores were significantly correlated with conventional neuropsychological assessment scores, including PACC5 (β = 10.7, SE = 2.0, p < 0.001) and composite scores of executive function (β = 9.1, SE = 2.0, p < 0.001) and processing speed (β = 10.7, SE = 1.6, p < 0.001), but not memory. In relation to AD biomarkers, worse DCTclock summary scores were weakly associated with higher global Aβ burden (β = −12.1, SE = 6.5, p = 0.065, Figure 2A) and significantly differentiated Aβ+ from Aβ− groups (β = −0.09, SE = 0.03, p = 0.004, Cohen's d = 0.47), outperforming both PACC5 and cognitive composites. Worse DCTclock summary scores were significantly associated with higher tau burden in the entorhinal cortex (β = −11.8, SE = 4.8, p = 0.014, Figure 3A), with similar associations with entorhinal tau levels observed with PACC5 and memory composite scores. Complete cross‐sectional results, including all cognitive domain‐specific subscores for DCTclock and comparisons with conventional neuropsychological assessments, are provided in the Supporting Information.

FIGURE 1.

Associations between DCTclock performance and age at baseline (A) and longitudinally (B). At baseline, in CN individuals, older age was associated with worse DCTclock summary scores, but not with the subscores in the command versions of the four DCTclock cognitive domains. Longitudinally, older age was significantly associated with accelerated decline in DCTclock summary scores and in the command versions of Drawing Efficiency, Spatial Reasoning, and Simple Motor subscores, but not in the Information Processing subscores. In the scatter plots between baseline DCTclock performance and age, the blue lines indicate the best‐fitted linear regression line, and the shaded gray area represents the 95% confidence interval. Longitudinal DCTclock performance was represented by the marginal effects of linear mixed effect models for illustrative individuals with ages of 67 years (young), 78 years (average), and 89 years (old) in our cohort. The shaded area represents the 95% confidence interval. linear mixed effect model coefficients for the age × time term were displayed with the corresponding standard errors and p‐values.

FIGURE 2.

Associations between DCTclock performance and global Aβ burden at baseline (A) and longitudinally (B). At baseline, in CN individuals, higher Aβ burden showed weak associations with worse DCTclock summary scores (p = 0.065), but not with the command versions of the four DCTclock subscores. Longitudinally, higher baseline Aβ burden was significantly associated with accelerated decline in DCTclock summary scores and in the command version of Drawing Efficiency, Information Processing, and Spatial Reasoning subscores, but not in the Simple Motor subscores. In the scatter plots between baseline DCTclock performance and PiB DVR, the blue lines indicate the best‐fitted linear regression line, and the shaded gray area represents the 95% confidence interval. Longitudinal DCTclock performance was represented by the estimated marginal effects from linear mixed effect models at varying Aβ levels (PiB DVR ranged between 1.0 and 1.8). The shaded area represents the 95% confidence interval. PiB DVR distributions: Majority of participants showed PiB DVR values between 1.0 and 1.2 (low Aβ), with some participants showing varying levels of higher Aβ burden: 26 in (1.2, 1.4], 19 in (1.4, 1.6], and 12 above 1.6. linear mixed effect model coefficients for the Aβ × time term were displayed with the corresponding standard errors and p‐values.

FIGURE 3.

Associations between DCTclock performance and entorhinal tau burden at baseline (A) and longitudinally (B). At baseline, in CN individuals, higher entorhinal tau burden was significantly associated with worse DCTclock summary scores (p = 0.014) but not with the command versions of the four subscores. Longitudinally, in CN individuals, higher entorhinal tau burden was significantly associated with accelerated decline in the command version of the Information Processing subscores but not in the DCTclock summary scores nor in the other subscores. In the scatter plots between baseline DCTclock performance and FTP SUVR, the blue lines indicate the best‐fitted linear regression line, and the shaded gray area represents the 95% confidence interval. Longitudinal DCTclock performance was represented by the estimated marginal effects from linear mixed effect models at varying tau levels (FTP SUVR ranged between 1.2 and 2.0). The shaded area represents the 95% confidence interval. FTP SUVR distribution: Participants showed varying levels of FTP SUVR values: 24 in (1.2–1.4], 81 in (1.4–1.6], 39 in (1.6, 1.8], 18 above 1.8. linear mixed effect model coefficients for the tau × time term were displayed with the corresponding standard errors and p‐values.

3.3. Longitudinal Changes in DCTclock Summary Scores in Relation to AD Biomarkers

3.3.1. Longitudinal Decline With Age

The DCTclock summary scores decreased over time (β = −0.11, SE = 0.03, p < 0.001), with accelerated decline associated with older age (β = −0.01, SE = 0.003, p < 0.001, Figure 1B). For example, a 78‐year‐old right‐handed male with 16 years of education is estimated to have a baseline DCTclock summary score of 67 and to decline at a rate of −0.11/month or −1.3/year. An additional −0.1/year decline is observed with every additional year in age: an 89‐year‐old with the same demographic profile would be expected to decline at a rate of −2.6/year, while a 67‐year‐old would show minimal change, and a 56‐year‐old would show modest improvement at a rate of +1.3/year. LME results evaluating the longitudinal performance of DCTclock are summarized in Table 3. Individual DCTclock trajectories are illustrated in Figure S1.

TABLE 3.

Linear mixed effect model estimates for longitudinal DCTclock summary scores and the command versions of the domain‐specific subscores.

Predictors	Estimates [standard errors]
DCTclock variables	Summary scores	Drawing efficiency	Information processing	Spatial reasoning	Simple motor
Clinical variables Model 1—DCTclock ~ time + age + sex + education + handedness + age × time
Intercept	67 [3.0]***	N.S.	N.S.	−0.52 [0.17]**	−0.40 [0.16]*
Age	−0.70 [0.15]***	N.S.	N.S.	N.S.	N.S.
Sex	N.S.	N.S.	N.S.	N.S.	N.S.
Education	1.31 [0.42]**	0.06 [0.02]**	0.04 [0.02]*	0.07 [0.02]*	N.S.
Handedness	N.S.	N.S.	N.S.	N.S.	N.S.
Time (months)	−0.11 [0.03]**	N.S.	N.S.	N.S.	−0.006 [0.001]**
Time × Age	−0.01 [0.003]***	−7.8e−4 [2.3e−4]**	N.S.	−7.9e−4 [3.2e−4]*	−5.1e−4 [2.1e−4]*
Global Aβ burden—[11C]PiB Model 2—DCTclock ~ time + age + sex + education + handedness + PiB FLR + age × time + PiB FLR × time
Baseline PiB FLR	−14 [6.4]*	N.S.	N.S.	N.S.	N.S.
Time (months)	−0.13 [0.03]***	−0.003 [0.001]*	N.S.	N.S.	−0.007 [0.001]***
Time × PiB	−0.28 [0.13]*	−0.02 [0.007]*	−0.02 [0.007]**	−0.02 [0.01]*	N.S.
Time × Age	−0.01 [0.004]**	−6.8e−4 [2.4e−4]**	N.S.	N.S.	−5.7e−4 [2.2e−4]**
Entorhinal Tau burden—[18F]Flortaucipir (FTP) Model 3—DCTclock ~ time + age + sex + education + handedness + FTP Entorhinal + age × time + FTP Entorhinal × time
Baseline FTP Entorhinal	−11.3 [4.6]*	N.S.	N.S.	N.S.	N.S.
Time (months)	−0.12 [0.03]***	N.S.	N.S.	N.S.	−0.006 [0.001]**
Time × FTP Entorhinal	N.S.	N.S.	−0.02 [0.006]**	N.S.	N.S.
Time × Age	−0.01 [0.004]**	−6.8e−4 [2.5e−4]**	N.S.	N.S.	−5.8e−4 [2.2e−4]**

Note: Linear mixed models (LME) with random slopes and intercepts were used to assess DCTclock longitudinal performance with respect to clinical variables and PET‐based measures of pathophysiological burden, as DCTclock ~ time + age + sex + education + handedness + age × time + PET × time, random = ~1 + time|subject, where time is time since DCTclock baseline. Age, years of education, and PET outcomes were centered around group‐means. LME coefficients were displayed with the estimated standard errors in brackets.

Abbreviations: DCTclock, digital clock drawing test; FLR, frontal, lateral temporal and retrosplenial composite region for estimating global Aβ burden; N.S., non‐significant at p = 0.05; PiB, [¹¹C]Pittsburgh‐Compound‐B.

^*** p < 0.001.

^** p < 0.01.

^* p < 0.05.

3.3.2. Longitudinal Decline With Aβ Burden

Incorporating baseline global Aβ burden, DCTclock summary scores decreased over time (β = −0.13, SE = 0.03, p < 0.001), with greater Aβ burden associated with accelerated decline (β = −0.28, SE = 0.13, p = 0.03, Figure 2B). For example, a PiB DVR of 1.19 in the above 78‐year‐old participant corresponds to an estimated decline of −1.6/year, while a DVR of 1.7 accelerates the decline to −3.3/year. In contrast, a DVR of 1.0 would be associated with a more modest decline of −0.9/year.

3.3.3. Longitudinal Decline With Tau Burden

Incorporating baseline tau burden, DCTclock summary scores decreased over time (entorhinal: β = −0.12, SE = 0.03, Figure 3B; inferior temporal: β = −0.13, SE = 0.03, p < 0.001, Figure S2). However, no significant tau burden and time interaction was observed, suggesting baseline tau burden did not predict accelerated decline in DCTclock summary scores.

3.4. Longitudinal Changes in DCTclock Cognitive Domain‐Specific Subscores in Relation to AD Biomarkers

3.4.1. Longitudinal Decline With Age

Older age was associated with accelerated decline across most DCTclock subscores in both the command and copy versions, including Drawing Efficiency, Information Processing, Spatial Reasoning, and Simple Motor subscores (Figure 1B). The only exception was the command version of the Information Processing subscores, which showed no significant age‐related decline. Younger CN individuals (age ~65 years) exhibited improved performance over time in Drawing Efficiency, Spatial Reasoning, and Simple Motor domains. These associations were driven by the total time to complete the task, pen speed (average, max, initiation, and termination), component placement, and latency (particularly in the copy version). Detailed statistical outcomes are reported in Table 3.

3.4.2. Longitudinal Decline With Aβ Burden

Higher baseline global Aβ burden was associated with accelerated decline in the command versions of Drawing Efficiency, Information Processing, and Spatial Reasoning subscores (Figure 2B), but not in the Simple Motor domain or the copy versions. Each unit increase in PiB DVR from 1.19 was associated with an additional estimated decline of −0.02/month or −0.24/year in these domain‐specific subscores (z‐scores). For example, the same 78‐year‐old male with a PiB DVR of 1.0, 1.19, and 1.7 would show a longitudinal change in the command version of the Information Processing subscores at a rate of +0.05/year, 0/year, and −0.12/year, respectively. The associations were driven by latency‐related features (e.g., average latency, latency variability, relative long latency, and longest latency; Information Processing domain), total time (Drawing Efficiency domain), and component placement (Spatial Reasoning domain).

3.4.3. Longitudinal Decline With Tau Burden

Elevated tau burden in the entorhinal (β = −0.02, SE = 0.006, p = 0.008, Figure 3B) and inferior temporal cortex (β = −0.03, SE = 0.009, p = 0.001, Figure S2) was associated with accelerated decline in the command version of the Information Processing subscores. Each unit increase in entorhinal FTP SUVR from 1.2 was associated with an additional decline of −0.02/month or −0.24/year. The example 78‐year‐old male with SUVR values of 0.9, 1.2, and 1.5 would show changes in the command version of Information Processing subscores at a rate of +0.07/year, 0/year, and −0.07/year, respectively. These associations were driven by the latency features within the Information Processing domain. No significant tau associations were observed in the copy versions.

3.4.4. Longitudinal Decline With Aβ Burden and Tau Burden

Evaluating both global Aβ and regional tau burden, only tau burden remained significantly associated with accelerated decline in the command version of the Information Processing subscores (β = −0.03, SE = 0.01, p = 0.001, Figure S3). These associations, again, were driven by the latency features.

3.5. Comparison Between Longitudinal Changes in DCTclock and Conventional Neuropsychological Assessments

3.5.1. Longitudinal Associations

Changes in DCTclock summary scores were significantly associated with changes in PACC5 (β = 8.6, SE = 1.1, p < 1e⁻¹³), cognitive composites of memory (β = 3.4, SE = 0.1, p = 0.002), executive function (β = 8.4, SE = 1.3, p < 0.001), and processing speed (β = 10.3, SE = 1.1, p < 0.001). Importantly, DCTclock summary scores exhibited additional decline even after accounting for changes in PACC5 (β = −0.08, SE = 0.03, p = 0.006) and cognitive composite scores (β = −0.14 − −0.11, p < 0.001 for memory and executive function; β = −0.06, SE = 0.03, p = 0.03 for processing speed), suggesting that DCTclock captures unique variance in cognitive decline.

3.5.2. Longitudinal Associations With Aβ Burden

Considering baseline global Aβ burden, DCTclock summary scores did not capture unique variance in cognitive decline beyond changes in PACC5 and cognitive composites of memory, executive function, and processing speed. However, the command Information Processing subscores from DCTclock did capture unique variance (β = −0.02, SE = 0.01, p = 0.02) after accounting for the cognitive composites.

3.5.3. Longitudinal Associations With Tau Burden

Higher baseline entorhinal tau burden was consistently associated with accelerated decline in the command version of the Information Processing subscores, even after accounting for changes in PACC5 (β = −0.02, SE = 0.01, p = 0.01) and cognitive composites of memory, executive function, and processing speed (β = −0.03, p: 0.002–0.005).

4. Discussion

This study demonstrates that quick and accessible digital cognitive assessments, such as the digital clock drawing test, can capture subtle cognitive deficits and longitudinal decline in CN older individuals with elevated AD‐related neuropathology (i.e., Aβ and tau), even before clinical diagnoses of dementia. Specifically, CN individuals with elevated Aβ and tau burden exhibited accelerated decline in the command Information Processing domain of DCTclock, which may reflect early changes in executive function and processing speed. These findings support the use of digital cognitive tools for monitoring preclinical AD progression and potentially for assessing therapeutic efficacy in clinical trials.

The clock drawing test has long been used for cognitive assessments [26, 27, 28, 29], but the digital adaption has enhanced its sensitivity by capturing nuanced temporal and graphomotor features in real time. We found significant associations between DCTclock performance and PET measures of Aβ and tau burden in older CN individuals. These results, obtained with a larger sample size, confirm our prior findings [4] and reinforce the potential role of digital cognitive assessment for early disease detection [4, 16, 30, 31, 32, 33]. More importantly, this study highlights that DCTclock is not only useful for cross‐sectional detection of cognitive deficits but also for tracking longitudinal cognitive decline over time.

A key observation in our study was that the association between baseline Aβ and tau levels and cognitive decline on DCTclock over four years in preclinical AD was primarily driven by latency features in the command condition. Increased latency, which manifests as more pauses and hesitation despite familiarity with the task, may indicate greater cognitive effort and inefficiency. This greater cognitive effort may represent an early sign of emerging cognitive impairment in preclinical AD, especially in executive function and processing speed. Most prior DCTclock studies have focused on its diagnostic ability at a single time point, with limited exploration of longitudinal performance changes. While cross‐sectional studies have emphasized graphomotor features, such as completion time and misplacement errors, as key markers of cognitive impairment [4, 30, 34], our findings suggest that latency features may be more sensitive to early AD progression.

This increased latency is unlikely attributed to normal aging alone. We found no significant longitudinal associations between DCTclock latency features and age, suggesting that these changes reflect AD‐related neuropathology rather than typical age‐related decline. Given that executive function and processing speed decline with age in the general population [35, 36], our findings highlight the potential of digital tools to distinguish AD‐related cognitive decline from aging. Future studies, including younger CN individuals, are needed to further clarify these differences.

A notable difference between our current findings and prior cross‐sectional DCTclock study is that the command version, rather than the copy version, demonstrated stronger sensitivity to Aβ‐ and tau‐related cognitive decline in older CN individuals [4]. The command version of the clock drawing test places greater cognitive demands on executive function and semantic memory, requiring individuals to retrieve and generate clock features without external guidance, in contrast to the copy version, which relies more on visuospatial processing and perceptual abilities [26, 28, 37]. This suggests cognitive tests optimized for detecting cross‐sectional associations with AD pathology may differ from those most sensitive to tracking longitudinal cognitive decline, with the command version being more sensitive to subtle, progressive cognitive decline over time.

Some individuals with low AD pathological burden exhibited improved DCTclock performance over time. These improvements may partly reflect test familiarity; other factors such as cognitive reserve, neuroplasticity, and lifestyle factors—including physical activity, reduced stress, improved sleep, and diet—may also contribute to enhanced executive function and processing speed [38, 39, 40]. Compensatory mechanisms may temporarily support cognitive function in those with low neuropathological burden. In contrast, these benefits may be less pronounced in individuals with high neuropathological burden, likely due to the dampening effects of pathology on compensatory processes.

Our results further suggest that Aβ and tau pathology contribute to cognitive decline in distinct ways. Elevated Aβ burden was linked to accelerated decline in multiple DCTclock features, including total completion time and component placement. In contrast, tau burden showed stronger associations with declines in latency features, reinforcing the notion that tau pathology is more closely linked to subsequent cognitive decline in preclinical AD. This aligns with the hypothesis that Aβ accumulation may initiate AD pathology, but tau deposition is a stronger driver of cognitive impairment [1, 2, 41, 42]. While tau pathology has traditionally been associated with memory deficits [43, 44], our findings indicate that digital cognitive tools may tap into executive aspects of memory retrieval and source recollection, particularly in the command condition. This supports the idea that AD‐related cognitive decline is multifaceted, affecting executive function, processing speed, and memory retrieval in distinct but interrelated ways.

Furthermore, our findings highlight the overlapping yet complementary contributions of DCTclock and conventional neuropsychological assessments, including the PACC5, in tracking early cognitive decline among CN older adults. Changes in DCTclock performance were strongly associated with changes in PACC5 and cognitive composite scores, confirming that these measures reflect shared variance in cognitive decline. More importantly, DCTclock also captured unique aspects of cognitive decline even after accounting for changes in PACC5 and cognitive composite, suggesting its sensitivity to cognitive processes not fully captured by conventional tests. While DCTclock did not add explanatory value beyond PACC5 in relation to Aβ burden, consistent with the design of PACC5 to detect cognitive decline in Aβ individuals [6, 22], it did capture unique variance associated with tau burden, particularly through the latency features. The additional tau‐related cognitive decline remained even after controlling for PACC5, underscoring DCTclock's sensitivity to tau‐related cognitive decline in preclinical AD. Collectively, these results support the utility of DCTclock as a fast, scalable, and complementary digital tool alongside conventional cognitive assessments.

Recent AD clinical trials have demonstrated promising results when intervening at early disease stages, highlighting the need for sensitive cognitive measures to track subtle cognitive decline before significant impairment occurs [45, 46]. As the field shifts towards a biological definition of AD, which prioritizes neuropathologic biomarkers over symptomatic diagnosis [2, 47], identifying sensitive cognitive markers that correlate with these biomarkers is critical. Digital cognitive tools offer a scalable, low‐burden method to enhance early detection, patient stratification, and outcome measurement in clinical trials targeting AD pathology. Interestingly, DCTclock exhibited weaker associations with education compared to the PACC5 and other conventional cognitive composites, which suggest that this tool may generalize well to diverse populations.

Despite the strengths of the study, several limitations should be acknowledged. First, this study's highly educated, predominantly white cohort may limit generalizability, as education can enhance cognitive test performance and reduce sensitivity to early AD‐related decline. However, DCTclock may improve detection in such populations by overcoming ceiling effects seen in traditional assessments. Future studies should examine more diverse groups to validate its broader applicability. Additionally, DCTclock requires a specialized digital pen and in‐person administration, limiting scalability. Linus Health has developed a version that uses an Apple iPad and Apple Pencil for drawing by eliminating the need for a digitizing pen, but further development is required to expand its scalability. Finally, establishing baseline norms and test–retest reliability in individuals without AD pathology will improve the interpretation of longitudinal changes and help differentiate early disease‐related decline from normal aging or test variability.

In summary, our results suggest that longitudinal decline in DCTclock performance is related to early Aβ and tau burden in older CN adults and captures unique variance in tau‐related cognitive decline compared to conventional neuropsychological assessments. The decline in longitudinal DCTclock performance, especially in latency features uniquely captured by digital tools, may reflect greater cognitive efforts in preclinical AD. Digital cognitive tools offer a non‐invasive, scalable, and cost‐effective method for assessing cognitive status, monitoring disease progression, and improving participant screening for clinical trials, in the context of progressive pathophysiological burden, particularly in underserved and remote populations.

Author Contributions

J.F.F., R.A.S., K.V.P., K.A.J., J.C.P., and D.M.R. contributed to the conception and design of the study. J.F.F., T.R., M.R.A., A.F.M., R.J.J., M.J.P., B.C.H., J.C.T., G.D.C.M., and E.T. contributed to the acquisition and analysis of data. J.F.F., T.R., D.P., R.D., and D.M.R. contributed to drafting the text or preparing the figures.

Conflicts of Interest

D.P. and R.D. are named inventors on the DCTclock (the digital cognitive tool used in this work) intellectual property patents and receive royalties for patents licensed by Linus Health through Lahey Hospital and the Massachusetts Institute of Technology, respectively.

Supporting information

Figure S1: Longitudinal DCTclock trajectories by baseline biomarker burden in CN individuals. Raw longitudinal DCTclock data were plotted by baseline global Aβ burden (A) and entorhinal tau burden (B), with color indicating baseline Aβ and tau values. Shading and color bars reflect continuous PET DVR (Aβ) and SUVR (tau) values at baseline.

Figure S2: Associations between DCTclock performance and global Aβ burden at baseline (A) and longitudinally (B). At baseline, in CN individuals, higher entorhinal tau burden was significantly associated with worse DCTclock summary scores (p = 0.014) but not with the command versions of the four subscores. Longitudinally, in CN individuals, higher entorhinal tau burden was significantly associated with accelerated decline in the command version of the Information Processing subscores but not in the DCTclock summary scores nor in the other subscores.

Figure S3: Impact of both Aβ and tau burden on longitudinal DCTclock performance. Tau burden was significantly associated with steeper decline in DCTclock performance over time (β (tau × time) = −0.03 [0.01], p = 0.001), whereas baseline Aβ burden demonstrated trend level association with decline DCTclock performance (β (Aβ × time) = −0.02 [0.01], p = 0.05).

Funding Statement

This work was funded by National Institute of Biomedical Imaging and Bioengineering grants P41EB01589 and P41EB030006; National Institute on Aging grants R00AG081457, P01AG036694, and R01AG046396.

Data Availability Statement

The data that support the findings of this study are available from the Harvard Aging Brain Study and Massachusetts General Hospital. Restrictions apply to the availability of these data, which were used under license for this study. Data are available [https://habs.mgh.harvard.edu/researchers/].