Association of Digital Clock Drawing With PET Amyloid and Tau Pathology in Normal Older Adults

Abstract

Objective

To determine whether a digital clock-drawing test, DCTclock, improves upon standard cognitive assessments for discriminating diagnostic groups and for detecting biomarker evidence of amyloid and tau pathology in clinically normal older adults (CN).

Methods

Participants from the Harvard Aging Brain Study and the PET laboratory at Massachusetts General Hospital were recruited to undergo the DCTclock, standard neuropsychological assessments including the Preclinical Alzheimer Cognitive Composite (PACC), and amyloid/tau PET imaging. Receiver operating curve analyses were used to assess diagnostic and biomarker discriminability. Logistic regression and partial correlations were used to assess DCTclock performance in relation to PACC and PET biomarkers.

Results

A total of 300 participants were studied. Among the 264 CN participants, 143 had amyloid and tau PET imaging (Clinical Dementia Rating [CDR] 0, Mini-Mental State Examination [MMSE] 28.9 ± 1.2). An additional 36 participants with a diagnosis of mild cognitive impairment or early Alzheimer dementia (CDR 0.5, MMSE 25.2 ± 3.9) were added to assess diagnostic discriminability. DCTclock showed excellent discrimination between diagnostic groups (area under the receiver operating characteristic curve 0.86). Among CN participants with biomarkers, the DCTclock summary score and spatial reasoning subscores were associated with greater amyloid and tau burden and showed better discrimination (Cohen d = 0.76) between Aβ± groups than the PACC (d = 0.30).

Conclusion

DCTclock discriminates between diagnostic groups and improves upon traditional cognitive tests for detecting biomarkers of amyloid and tau pathology in CN older adults. The validation of such digitized measures has the potential of providing an efficient tool for detecting early cognitive changes along the AD trajectory.

Classification of Evidence

This study provides Class II evidence that DCTclock results were associated with amyloid and tau burden in CN older adults.

Alzheimer disease (AD) is a continuum, with the pathophysiologic brain changes of β-amyloid (Aβ) plaques, neurofibrillary tau tangles, and subsequent neurodegeneration occurring 15–20 years before the clinical stage of AD dementia.¹ Identifying individuals at risk for AD will be critical as prevention trials or potential treatments become available. Amyloid and tau PET biomarkers of AD pathology can detect evidence of preclinical AD in asymptomatic individuals, but they are costly and inaccessible in the clinical setting.^2,3 Furthermore, traditional cognitive screening tests were not designed to correlate with these early biomarker abnormalities.⁴ We explored whether a digital cognitive test, designed to elicit multiple performance features not available in traditional cognitive formats, could be useful for this purpose.

For nearly a century, clock-drawing tests have been used to assess the mental status of patients and is one of the most widely used screening tests for cognitive impairment in AD.^5,6 This simple test taps into a wide range of cognitive domains including executive functioning, visuospatial abilities, and semantic memory and has shown high sensitivity for discriminating cognitive decline from normal performance in the clinical setting.⁵

We used the digital clock-drawing test (DCTclock), obtained from Digital Cognition Technologies, to explore whether this digitized measurement approach would be clinically useful for discriminating cognitively normal (CN) from diagnostic groups. We also hypothesized that digital clock performance scores would have a greater association with PET amyloid and tau deposition in CN older adults than traditional paper and pencil cognitive tests.

Methods

Study Participants

This cross-sectional study was conducted at Massachusetts General Hospital (MGH) using protocols and informed consent procedures approved by the Partners Human Research Committee. In total, 300 participants were recruited from the Harvard Aging Brain Study (HABS) and separate imaging protocols from the MGH PET laboratory. Clinically normal participants (CN) (n = 264) were included if they had a Clinical Dementia Rating (CDR) score of 0, DCTclock, and standard neuropsychological tests including the Primary Alzheimer Cognitive Composite (PACC). A total of 143 CN participants underwent PET imaging for AD biomarkers on the same PET tracers (described below). We also included 36 participants referred to the MGH PET laboratory who were diagnosed by neurology specialists at the Memory Disorders Unit of MGH with mild cognitive impairment (MCI) (n = 29) (CDR 0.57 ± 0.18) or early AD (n = 7) (CDR 0.62 ± 0.22). These 36 participants were combined to form a single diagnostic group because of their clinical equivalence on CDR. The 36 MCI and AD participants also had DCTclock and standard neuropsychological tests including the PACC. They were added to the sample for the purpose of exploring the diagnostic discriminability of the DCTclock but were not included in the PET analyses because they had PET imaging on different PET tracers. None of the participants included in the study had a history of substance abuse, psychiatric disorders, or other neurologic diseases.

All 300 participants completed the copy and command versions of the DCTclock using a commercially available digital pen and specialized paper with a faint grid pattern. All participants completed additional cognitive testing using standard neuropsychological tests. We computed the PACC,⁷ a well-validated multidomain composite score currently serving as the primary outcome measure in a large AD secondary prevention trial.^8,9 The PACC in this study includes Logical Memory Delayed Recall (LMDR),¹⁰ the Free and Cued Selective Reminding Test (FCSRT),¹¹ the Mini-Mental State Examination (MMSE),¹² and the Digit Symbol Substitution Test (DSST).¹³ Cognitive composites for the domains of memory, executive functioning, and processing speed were computed from tests administered in the HABS according to previously described methods.¹⁴ These comprised a memory composite, which included the FCSRT–Free Recall, the 6-trial Selective Reminding Test Delayed Recall,¹⁵ and the Logical Memory–Delayed Recall; an executive functioning composite that included Trail-Making Test B,¹⁶ Phonemic Fluency,¹⁷ and Letter-Number Sequencing¹⁸; and a processing speed composite comprising Trail-Making Test A¹⁶ and DSST.¹³

Standard Protocol Approvals, Registrations, and Patient Consents

The Partners Human Research Ethics Committee at MGH approved the study and written informed consent was obtained from all participants prior to study procedures.

Neuroimaging

A total of 143 participants underwent PET with ¹¹C Pittsburgh compound B (PiB) and ¹⁸F flortaucipir (FTP).² All PET imaging was obtained on a Siemens ECAT HR + PET scanner. PiB images were acquired using a 60-minute dynamic acquisition and FTP images were acquired over 75–105 minutes. PET images were coregistered to corresponding T1 images using Freesurfer-based (v6) structural regions of interest (ROIs) mapped into native PET space using SPM12. FTP was expressed as a standard uptake volume ratio and PiB as the distribution volume ratio. The reference region was cerebellar gray using an MRI-based method. FTP-PET data were corrected for partial volume effects and calculated in entorhinal and inferior temporal lobe regions for each participant. For PiB, a global cortical aggregate was calculated in frontal, lateral temporal, and retrosplenial regions and was examined both continuously and dichotomously for each participant. Low (Aβ−) vs high (Aβ+) groups were determined based on a cutoff of 1.185.¹⁹ The DCTclock occurred within 36 months of PiB-PET (mean 11.04 ± 8.17) and within 18 months of FTP-PET (mean 8.53 ± 4.57) collection.

DCTclock

DCTclock was originally designed and developed at Lahey Clinic and the Massachusetts Institute of Technology and further developed into an Food and Drug Administration–cleared product by Digital Cognition Technologies Inc. and licensed for the purpose of this research. DCTclock was developed based on the familiar paper and pencil task²⁰ and contains a command and copy condition. Participants were initially presented with a blank piece of paper that has a faint grid pattern and handed a digital pen that looks and functions like a normal pen. They were instructed to “draw the face of a clock, put in all of the numbers, and set the hands to 10 after 11.” When they finished the drawing, the paper was taken away. The participants were then asked to complete the copy condition with the digital pen. They were given a paper with a drawn clock and instructed to replicate the picture with the hands set to 10 after 11.

DCTclock contains multiple objective measurements that were derived from approximately 5,000 digital clock drawings using machine learning algorithms to precisely evaluate nuances in performance beyond successful task completion.²¹ To acquire these performance nuances, the digital pen contains a camera sensor that captures pen position every 12 ms. This allows for the capture of hundreds of clock drawing features to be analyzed^21,22 (e.g., pen velocity, ink time, pen off and on the page) as a series of time-stamped (x,y) coordinates. Machine learning algorithms were previously developed to calculate meaningful clock scores based on their ability to discriminate performance between thousands of healthy controls and patients from different diagnostic groups including amnestic MCI, AD dementia, Parkinson disease, and other neurodegenerative disorders.²¹ The DCTclock algorithm analyzes the drawing signal in 2 steps. First, artificial intelligence algorithms were used to classify the individual pen strokes into drawing components (i.e., clock face, hands, numbers, noise), similarly to how a human eye would categorize a clock drawing. Second, the classified pen strokes were used in combination with the raw drawing signal to calculate variables that can be predictive of cognitive state, such as the distribution of latencies in the drawing process, pen velocities at different points in the drawing, and accuracy in completion of the task.²² These variables from the command and copy versions were combined from the machine learning calculations into an overall score ranging from 0 to 100. A similar technique was used for the domain-specific subscores of drawing efficiency, information processing, simple motor, and spatial reasoning. The subcomponents for each of these domain-specific scores were also available; e.g., the spatial reasoning domain examines clock face circularity (i.e., how accurate the circle was drawn), component placement (i.e., the accuracy of where the hands and numbers are placed around the clock face), vertical/horizontal spatial arrangement (i.e., where the clock is placed on the page), and drawing size (figure 1). These fine-grained parameters provide additional detailed information about performance that are not typically available using traditional hand-scoring techniques.

Figure 1. Example of Variables Comprising the Digital Clock Drawing Test (DCTclock).

DCTclock includes a total summary score, which ranges from 0 to 100. Domain-specific scores are produced for both command and copy versions of the task (e.g., spatial reasoning) in addition to subcomponent scores for each domain-specific score (e.g., clock face circularity). Subcomponents are only shown for spatial reasoning.

To compare digitized scoring to the traditional hand-scored clock, all clock drawings were independently scored by 2 trained neuropsychologists. Whereas there are several scoring systems available for clock drawing,^23,24 we chose the Montreal Cognitive Assessment 3-point scoring system,²⁵ a standard, well-established method that requires the rater to subjectively assess 0 to 3 points, with 3 being the highest score, based on (1) circle contour—no points were given if the circle was incomplete, elongated, or not properly closed; (2) number placement—no points were given if the numbers were outside their respective quadrants, in the wrong sequence, or missing; and (3) hand placement—no points were given if the time was inaccurate, the hour hand was larger than the minute hand, or the hands did not join in the center of the clock face. Raters scored each command and copy drawing without knowledge of each other's scores or participant cognitive/biomarker status. Consensus was used for discrepant scores.

Statistical Analyses

Differences in demographic characteristics were described using standardized difference scores. The first objective was to determine whether the DCTclock summary score could distinguish between diagnostic groups (CN vs MCI/early AD).²⁶ The MCI and early AD participants were combined into one diagnostic group because clinical severity on CDR was equivalent. A receiver operating characteristic curve (ROC) analysis was completed to further evaluate the extent to which the DCTclock summary score, in comparison with the hand-scored clock and PACC, could discriminate between CN and MCI/AD diagnostic groups. Also, partial correlations were used to examine the association between the DCTclock summary score and cognitive composites for memory, executive functioning, and speed while controlling for age, sex, and education, to determine which cognitive domains were most associated with DCTclock performance.

The next primary objective was to determine whether the DCTclock summary score could detect amyloid and tau burden, even earlier along the AD trajectory in CN older adults. This analysis will provide Class II evidence that the DCTclock results are associated with amyloid and tau burden among the clinically normal older adults.

To explore this objective, those CN participants who had PET imaging were classified as Aβ+ vs Aβ− using 1.185 as the cutoff.¹⁹ A second ROC analysis was performed to evaluate the extent to which the DCTclock summary score, compared with the hand-scored clock and the PACC, could discriminate Aβ+ from Aβ− groups in CN participants only. A logistic regression was used to assess the ability of the DCTclock summary score, PACC, and age to predict Aβ status among CN participants. Partial correlations controlling for age, sex, and education were computed to assess the relationship between the DCTclock summary score, continuous Aβ, and entorhinal and inferior temporal tau among CN participants. Correlations between biomarkers and DCTclock scores (e.g., domain-specific scores and subcomponent scores) were only examined if the superordinate score was significantly associated with the biomarker to minimize type 1 error. In addition, Bonferroni corrections were applied for comparisons between biomarkers and the DCTclock summary score, biomarkers and domain-specific scores, and biomarkers and subcomponent scores. We also evaluated PiB-PET amyloid signal, vertex-wise in CN participants (vertex threshold p = 0.05 and cluster extent >200 mm²). Finally, linear regression models, adjusted for age, sex, and education, were used to evaluate the relationship of the DCTclock summary score and PACC with fibrillar amyloid deposition on PiB-PET.

Data Availability

DCTclock data will be made available by the corresponding author upon request. All baseline imaging and cognitive data from the HABS are available to the research community upon request at nmr.mgh.harvard.edu/lab/harvardagingbrain.

Results

Demographic Characteristics

Sample characteristics are provided in table 1 for the entire sample (n = 300) as well as the CN subset (n = 143) who had amyloid or tau PET imaging. As expected, MCI/early AD participants were slightly older, had lower years of education, and performed worse on the MMSE and LMDR compared with CN participants. Among the CN participants only, Aβ+ participants were slightly older, with lower MMSE scores compared with Aβ− participants (table 1).

Table 1.

Sample Characteristics

Relationship of Digitally Scored Clock Performance to Processing Speed and Executive Functions in CN Participants

Among CN participants, older age (r = −0.389, p < 0.001) and lower education (r = 0.169, p = 0.045) were associated with worse DCTclock summary scores. There were no sex differences in performance (t = 0.31, p = 0.759).

The DCTclock summary score was also correlated with a composite of processing speed (r = 0.427, p < 0.01) and executive functions (r = 0.289, p = 0.001) but not significantly related to memory (r = 0.148, p = 0.083).

Digitally Scored Clock Across the Diagnostic Spectrum (CN to MCI/Early AD)

The PACC, a composite of standardized cognitive tests, had greater diagnostic discriminative ability between groups (264 CN, mean 71.23 ± 21.87 vs 36 MCI/early AD, mean 35.94 ± 22.09) with an area under the ROC curve (AUC) of 0.95 (95% confidence interval [CI], 0.93–0.97). However, the DCTclock summary score also had an excellent level of discrimination between groups, with an AUC of 0.86 (95% CI, 0.77–0.90) (figure 2A). The effect size difference between diagnostic groups was also greater for the PACC (Cohen d = 2.42), than the DCTclock summary score (Cohen d = 1.55). In contrast, DCTclock provided a 24% increase in diagnostic discriminability compared with the hand-scored clock (AUC 0.67 [95% CI, 0.58–0.75]). A DCTclock overall score below 60/100 exhibited 83% sensitivity and 70% specificity. Put differently, the odds ratio for being cognitively impaired if scoring <60 is equal to 11 (95% CI, 4.94–22.53).

Figure 2. Receiver Operating Characteristic Curves for the Digital Clock Drawing Test (DCTclock).

(A) Diagnostic discriminability (clinically normal [CN] vs mild cognitive impairment [MCI]/early dementia). There is a 24% increase in diagnostic discriminability (area under the receiver operating characteristic curve [AUC]) between 264 CN individuals and 36 patients with MCI/early Alzheimer disease using digitally scored vs hand-scored methods for clock drawing. A multidomain composite (Preclinical Alzheimer's Cognitive Composite [PACC]) supersedes clock drawing in diagnostic discriminability. (B) β-amyloid (Aβ) ± status discriminability in CN individuals only. In CN individuals only, digitally scored clock drawing discriminated between Aβ+ (n = 40) and Aβ− (n = 103) groups whereas hand-scored clock did not. The DCTclock summary score outperformed the PACC in discriminating between groups (14% increase in AUC). CI = confidence interval.

Digitally Scored Clock and AD Biomarkers in CN Participants

Examining CN participants only, worse DCTclock performance was associated with Aβ burden, while controlling for age, sex, and education (r = −0.241, p < 0.01), in contrast to a lower PACC performance, which was nonsignificant (r = −0.100, p = 0.295) (table 2). In CN participants only, the DCTclock summary score exhibited a stronger level of discrimination between Aβ± groups with an AUC of 0.72 in contrast to the PACC (AUC 0.63) (figure 2B). The effect size difference in the DCTclock summary score between Aβ± groups was medium to large (Cohen d = 0.76), and greater than the effect size group difference using the PACC (Cohen d = 0.30). Furthermore, in a logistic regression using age and DCTclock to predict Aβ group, inclusion of the PACC did not improve the model (table 3). Lower DCTclock summary scores were associated with higher levels of entorhinal tau (r = −0.191, p < 0.05) but not inferior temporal tau (r = −0.072, p = 0.321) (table 2). Interestingly, the spatial reasoning domain-specific score (figure 1) from the DCTclock copy condition, not the command condition, was related to amyloid (r = −0.358, p < 0.01) and entorhinal tau (r = −0.314, p < 0.01). In addition, the subcomponent scores for spatial reasoning (figure 1) involving “component placement” (i.e., the accuracy of where the hands and numbers are placed around the clock face) remained significant after multiple comparison correction, but not the subcomponents of face circularity, vertical/horizontal spatial arrangement, or drawing size. However, worse vertical spatial arrangement (i.e., where the drawing was placed on the page) was associated with higher levels of entorhinal tau (table 2).

Table 2.

Correlations Between Digital Clock Drawing Test (DCTclock) Drawing Measures and Alzheimer Disease Biomarkers in Clinically Normal (CN) Individuals

Table 3.

Results of Logistic Regressions Predicting Clinically Normal Individuals With Aβ− and Aβ+ Groups

In vertex-wise analyses of amyloid in CN participants, the DCTclock summary score was associated with amyloid signal in regions commonly involved in amyloidosis, including anterior medial frontal and posterior medial parietal/precuneus (figure 3). Linear regression models adjusted for age, sex, and education confirmed these findings and showed that the DCTclock summary score was associated with amyloid deposition in regions found within the frontoparietal and default mode networks and not with regions outside of these networks, such as the pericalcarine and precentral (figure 4). In comparison, PACC performance in the CN only group showed less of an association with vertex-wise amyloid signal and was also not significantly associated with amyloid burden in any of the regions involved in the frontoparietal and default mode networks. Vertex-based tau analyses in CN participants confirmed the findings in ROIs (data not shown).

Figure 3. Digital Clock Drawing Test (DCTclock) Summary Score is Associated With Fibrillar Amyloid Burden in Clinically Normal Individuals (CN).

DCTclock summary score (A) is associated with Pittsburgh compound B retention distribution volume ratio in aggregate region of interest (A.a, p < 0.01), and at surface vertices (A.b). The Preclinical Alzheimer's Cognitive Composite (PACC) (B) was not significantly associated with neocortical amyloid (B.a, p = 0.295) and showed less of an association across the surface (B.b). Vertex-wise analyses are smoothed on the surface with a minimum cluster extent of 200 mm² and a threshold of p = 0.05 for each contributing vertex. These analyses were adjusted for age, sex, and education. FLR = frontal, lateral temporal, and retrosplenial regions.

Figure 4. Digital Clock Drawing Test (DCTclock) Summary Score is Associated With Amyloid Deposition in the Frontoparietal and d Default Mode Networks in Clinically Normal Individuals (CN).

DCTclock summary score (A) was associated with amyloid burden in regions found in the frontoparietal and default mode networks in CN individuals only, but not with amyloid outside of these regions, such as the pericalcarine and precentral. Preclinical Alzheimer's Cognitive Composite (PACC) performance (B) showed no significant association with amyloid deposition in any of these regions in CN individuals. The figure displays β estimates with a standard error bar showing 95% confidence intervals; data are adjusted for age, sex, and education. FLR = frontal, lateral temporal, and retrosplenial regions.

Discussion

As AD prevention trials and potential AD treatments gain momentum, refining our ability to proficiently identify at-risk individuals in the clinic setting will be critical. We found that a familiar and quick (<5 minutes) digital clock-drawing test, the DCTclock, was correlated with standard neuropsychological tests, particularly tests of processing speed and frontal executive functions, and was able to differentiate CN from MCI and early AD dementia as efficiently as traditional methods. Furthermore, DCTclock in CN tracked with AD biomarker severity of both global amyloid and regional tau and operated comparably to, if not better than, a 30-minute cognitive battery designed to be sensitive at differentiating Aβ+ from Aβ− groups. Similar to other traditional tests, older age was associated with worse DCTclock performance. However, unlike most traditional cognitive tests, there were nonsignificant differences in sex and education among CN participants. These findings suggest the digitally derived clock may have greater generalizability than many other traditional neuropsychological measures.

The earliest documented use of clock drawing, i.e., where an individual is asked to draw the face of a clock and set the hands to a specific time, dates to nearly a century ago when neurologists were asked to evaluate the mental state of soldiers who had parietal and occipital head wounds.²⁷ Observing how a person completes the task allows us to glean information about his or her cognitive state; e.g., did he or she put the numbers on the clock in an organized fashion, did he or she hesitate when placing the first or second hand, did he or she draw quickly with confidence or slowly with effort?

Whereas this traditional method of administration has been clinically useful, digital technology has the potential of improving upon clinician assessment by systematically quantifying countless performance details, thus providing a substantially more nuanced window into seeing how an individual completes the task, even if the finished drawing appears accurate. The DCTclock, used in this study, has been used by other researchers who explored its performance in various diagnostic groups, including healthy young and older adults, as well as those with depression, multiple sclerosis, Parkinson disease, MCI, AD dementia, and other neurodegenerative disorders.^21,28,29 The premise of these and other research studies^30–32 was to explore whether digital clock performance could capture added features in cognitive task processing not easily captured with a stopwatch and paper and pencil administration. For example, algorithmic digital software used in the DCTclock has the potential of measuring thinking processes, i.e., mental speed, efficiency, and time to decision-making by capturing “pen-time” on and off the page. This fine-grained approach exposes the efficient or inefficient strategies that individuals use to complete a task, even if the finished drawing appears accurate. For example, one study found that “time to decision-making” was able to discriminate patients with MCI from healthy controls with a sensitivity of 81.3% and specificity of 72.2%, compared with traditional clock drawing scores,³² and with slightly better diagnostic accuracy (81.5%) than traditional cognitive tests (77.5%).³¹ Another study found that digital clock drawing measurements for psychomotor speed could differentiate between the motor and cognitive aspects of psychomotor slowing in aging and depression.²⁸ In our sample, diagnostic discriminability using the DCTclock was increased by 24% when compared with traditional clock scoring, and scores below 60/100 exhibited 83% sensitivity and 70% specificity between diagnostic groups, similar to what others have found.^31,32 Interestingly, DCTclock discriminability between diagnostic groups, although excellent, was lower in comparison with the PACC, a composite of standardized neuropsychological tests. However, these gains in discriminability between diagnostic groups were reached at a cost (i.e., <5 minutes vs 30 minutes). Nonetheless, these findings are promising and suggest that digitized tests could be useful without compromising the benefits obtained from traditional measures that are acquired over longer administration times. In addition, digitized assessments have the potential of reducing rater error, which may be particularly helpful in decreasing variability when assessing performance over time.

More importantly, we found that digital clock deficits were noted even prior to diagnosis, i.e., at the preclinical stage of AD. Among CN individuals, the magnitude of the group differences between Aβ± groups were larger for the DCTclock (Cohen d = 0.76) compared with the PACC (Cohen d = 0.30) and the probability of the DCTclock score to predict Aβ status was also slightly better than the PACC. These findings suggest that digitized assessments have the potential of detecting Aβ status among CN individuals better than traditional cognitive measures. Future work will determine whether digital clock parameters can be improved upon for identifying Aβ+ among CN individuals, e.g., in conjunction with genetic and other demographic variables, and whether this mode of assessment can be useful in tracking decline over the course of preclinical AD.

Clock drawing performance, while simple and easy to administer, encompasses a broad variety of cognitive functions including visuospatial processing, executive skills, and semantic memory.³³ A recent fMRI study examined the underlying neural correlates of clock drawing in cognitively healthy older adults.³⁴ The authors used a novel touch-sensitive digitized tablet in the fMRI scanner to provide real-time visual feedback of hand and stylus position during clock drawing. The underlying neural correlates of positive task-related activity were observed in bilateral frontal, occipital, parietal, and inferior temporal cortices and, like our study, were negatively correlated with age. Because decreased neural activity was observed as a function of aging, the authors raised concerns about the validity of clock drawing as a screening tool for pathologic impairment in elderly populations. However, their participants were not screened with in vivo biomarkers that could reduce the confound of other neurodegenerative processes beyond aging. In our study, vertex-wise analyses of amyloid showed that the digital clock summary score was associated with signal in regions commonly involved in amyloidosis, including anterior medial frontal and posterior medial parietal/precuneus areas, suggesting that the underlying neural correlates of clock drawing, noted in the above study, may be associated not only with aging but other copathologies as well. Nonetheless, their concerns are relevant and further work will be needed to reduce the potential of false-positive results, if the test is to be used more universally in clinic and research settings.

Although it is widely known that total summary scores of the traditional clock-drawing test can differentiate diagnostic groups,^21,22,32 more fine-grained performance features, only available in a digitized format, may provide additional information about cognitive and neural network function in normal community-dwelling elders. Lamar and colleagues³⁰ found individual differences in graphomotor organization and executive functioning on the basis of whether individuals anchored the numbers (i.e., 12, 6, 3, 9) around a clock face. Those who anchored the numbers had better performances on standardized executive tasks and a higher degree of modular integration of neurocircuitry in the ventral visual processing stream in contrast to those who did not. Similar to this and other studies,^24,30,33 we also found that the DCTclock summary score was strongly associated with executive functions as well as amyloid deposition in regions found within the frontoparietal and default mode networks. However, it was the fine-grained features in spatial reasoning of the copy clock that were more strongly associated with increased amyloid and tau PET signal. When exploring further the performance details of spatial reasoning, we found that the spatial arrangement of the numbers around the clock face and placement of the hands were associated with amyloid burden and increased regional tau in the entorhinal and inferior temporal cortices. There were no associations found between biomarkers and the other subdomains of clock-drawing efficiency, simple motor, or information processing, suggesting that these cognitive processes may not be predominant at the preclinical stage but rather at later stages of dementia severity or in other neurodegenerative diseases. Interestingly, it was the copy and not the command clock that was associated with biomarker burden. It is well known that patients with AD dementia improve from the command to the copy drawing^23,35 on the premise that the command clock requires greater executive resources and semantic memory than the copy stage.^23,24,35 The fact that we found the opposite, that the copy version of spatial reasoning was associated with preclinical AD, may suggest that the DCTclock may be exposing some drawing hesitation through “timed pen lifts,” thus capturing greater thinking effort in the copy clock not seen in the command version. This greater thinking effort may potentially be revealing early pathologic burden seen in the frontoparietal control network in Aβ+ CN. While numerous studies^7,36,37 report memory changes at the preclinical stage, the digitally derived clock may be capturing early changes in frontal executive function that have gone unobserved cross-sectionally³⁷ but emerge longitudinally in preclinical AD.³⁸ This more refined examination of digital clock performance may be providing a window into the executive deficits that are occurring earlier along the preclinical continuum. In essence, attention to these fine-grained performance details of cognitive change, rather than a total score or time to completion, has the potential of improving our understanding of the earliest cognitive alterations in preclinical AD.

There are several limitations that need to be addressed. Use of the digital pen, in this study, required specialized paper and a docking station, in order to capture all the performance details in the digitized drawing. In order for this test to become suitable in clinical settings, a technology that is more widely available, such as an iPad or tablet, will be necessary and is under development. Also, to deploy this technology in clinical settings requires more than the purchase of a digital pen. It was the extensive work that went into creating the machine learning algorithms from various neurologic disorders that makes the digitized data valuable.²² Although clock drawing is easy to administer and quick, more work needs to be done to understand its test–retest reliability, longitudinal performance, and false-positive rate in the clinical setting. We would also like to increase our sample of MCI and AD patient groups to explore further its diagnostic discriminability and for detecting biomarkers of AD pathology, not only in CN individuals, but across the AD diagnostic spectrum. This will require further honing of the algorithms that were initially developed to discriminate diagnostic groups, to be sensitive to AD biomarkers. Further work will also be needed to determine how this digitized assessment operates over time. Finally, digitizing other standard cognitive tests that enjoy widespread use may be fruitful in producing more fine-grained knowledge about cognitive performance beyond what we currently derive from our traditional measures.

Developing tools that can efficiently measure cognitive deficits in otherwise CN older individuals is a major challenge particularly as potential treatments for AD prevention become available. Measuring cognitive performance with digital technology and interpreting performance based on big-data techniques is a groundbreaking approach that allows for an unprecedented number of cognitive performance features to be easily collected and mined for relevant signal. This fine-grained approach has the potential for delivering a cost-effective, simple method for detecting at-risk individuals who require further diagnostic workup. However, further development and validation will be needed before this technology can enjoy widespread use in the clinical setting.

Glossary

Aβ

β-amyloid

AD

Alzheimer disease

AUC

area under the receiver operating characteristic curve

CDR

Clinical Dementia Rating

CI

confidence interval

CN

clinically normal

DCTclock

digital clock-drawing test

DSST

Digit Symbol Substitution Test

FCSRT

Free and Cued Selective Reminding Test

FTP

18F flortaucipir

HABS

Harvard Aging Brain Study

LMDR

Logical Memory Delayed Recall

MCI

mild cognitive impairment

MGH

Massachusetts General Hospital

MMSE

Mini-Mental Status Examination

PACC

Primary Alzheimer Cognitive Composite

PiB

Pittsburgh compound B

ROC

receiver operating characteristic curve

ROI

region of interest

Appendix. Authors

Study Funding

National Institute of Aging for the Harvard Aging Brain Study (P01AG036694), Fidelity Biosciences Corporation, The National Science Foundation (SF-12-512).

Disclosure

D.M. Rentz has served as a consultant for Eli Lilly, Janssen, Biogen Idec, and Digital Cognition Technologies; serves on the scientific advisory board for Neurotrack; and was funded for this work by grants from Fidelity Biosciences Corporation, the National Science Foundation (SF-12-512), and the National Institute of Aging as an Administrative Supplement to the Harvard Aging Brain Study (P01AG036694). K.V. Papp has served as a consultant for Biogen Idec and Digital Cognition Technologies and is funded by a NIH K23 grant (5K23AG053422-04) and a fellowship grant from the Alzheimer's Association. R.A. Sperling has served as a consultant for AC Immune, Biogen, Janssen, Neurocentria, and Roche and has received research support from Eli Lilly and Eisai and funding from National Institute of Aging for the Harvard Aging Brain Study (P01AG036694). K.A. Johnson has served as a consultant for Novartis, AC Immune, Biogen, Janssen, Takeda, Merck, Life, Cerveau, and Roche; has received research support from Eli Lilly, Cerveau, and Eisai; and receives funding from the National Institute of Aging (R01 AG046396). W. Souillard-Mandar is a stakeholder and employee of Digital Cognition Technologies, Inc. and Linus Health, Inc. He has received research funding from the National Institute of Aging (1R43 AG066291-01). J.S. Sanchez, H. Klein, and D.V. Mayblyum report no disclosures. Go to Neurology.org/N for full disclosures.