Diagnostic accuracy of digital clock drawing test for Alzheimer disease and mild cognitive impairment

Abstract

Alzheimer’s disease (AD) and mild cognitive impairment (MCI) are major public health concerns, requiring accurate and scalable diagnostic tools. The digital clock drawing test (dCDT) captures drawing data and enables extraction of process-related features that may improve diagnostic performance. However, existing evidence remains inconsistent, highlighting the need for a systematic synthesis to support its clinical translation. We searched Web of Science, Embase, PubMed, PsycINFO, IEEE Xplore, CNKI, and Wanfang from inception to January 8, 2026. A bivariate mixed-effects model was used to pool sensitivity and specificity. A total of 13 studies comprising 17 diagnostic tests were included, and risk of bias was notable across studies. For MCI, the standalone dCDT showed pooled sensitivity of 0.765 (95% CI: 0.683–0.832), specificity of 0.752 (95% CI: 0.673–0.817), and pooled area under the summary receiver operating characteristic curve (AUC) of 0.825 (95% CI: 0.790–0.856). When both standalone and augmented dCDT tests were considered for MCI, the pooled sensitivity and specificity were 0.760 and 0.800, respectively, and the pooled AUC increased to 0.845. For AD, the pooled sensitivity, specificity, and AUC of dCDT were 0.820 (95% CI: 0.721–0.889), 0.897 (95% CI: 0.860–0.923), and 0.928 (95% CI: 0.902–0.948), respectively. Exploratory subgroup analyses of standalone dCDT for MCI suggested diagnostic performance appeared higher in studies employing algorithm-based approaches than in those using traditional-scoring approaches. Overall, the available evidence supports dCDT as a promising digital screening tool for cognitive impairment. Further multicenter studies and standardized protocols are needed to enhance its role in early diagnostic and clinical practice.

Introduction

With the accelerating pace of global population aging, the burden of Alzheimer’s disease (AD) and related dementias is increasing dramatically^1–3. These conditions not only diminish patients’ quality of life but also impose substantial caregiving and economic challenges on families and society^4,5. Mild cognitive impairment (MCI) is widely recognized as a prodromal stage of AD⁶, and early detection and intervention may delay disease progression and alleviate the societal burden^7–9. However, mental health services for older adults remain insufficient, particularly in low- and middle-income countries, where limited specialist resources and unequal distribution of care exacerbate disparities^10–13. Therefore, there is a pressing need for a brief, accurate, and scalable screening tool suitable for use in community and primary care settings to support early identification and large-scale population screening^14–16.

Among various cognitive assessment tools, the clock drawing test (CDT) has been widely applied due to its simplicity, time efficiency, and ability to evaluate multiple cognitive domains simultaneously¹⁷. The traditional paper-and-pencil CDT has been used in clinical practice for decades and has played an important role in screening for dementia and cognitive impairment^18,19. However, its reliance on manual scoring introduces subjectivity, with results varying according to examiner experience and scoring criteria, leading to limited reliability and consistency^18,20,21. Moreover, conventional CDT captures only the final static drawing, failing to reflect dynamic processes during task performance, which restricts its sensitivity for detecting subtle early cognitive deficits^19,22.

With the rapid advancement of digital technologies, the digital CDT (dCDT) has emerged as a promising digital alternative to the traditional CDT. The dCDT is a computerized version of the conventional paper-and-pencil CDT that uses a digital pen or tablet to capture both the final drawing and the entire drawing process²³. Unlike the traditional CDT, which records only the static end product, the dCDT continuously collects real-time information on pen strokes, pauses, pen lifts, timing, and movement dynamics^24,25. These real-time, process-based data provide deeper insights into patients’ cognitive processing patterns, offering more objective and quantitative evaluations with the potential for automated analysis and large-scale implementation^22,26. Emerging evidence also suggests that dCDT may outperform the traditional CDT in distinguishing MCI and AD, particularly at early disease stages^22,27,28.

Nevertheless, current evidence regarding the diagnostic performance of dCDT remains inconsistent. Limitations associated with small sample sizes, together with differences in devices and analytic algorithms, have led to wide variability in reported sensitivity and specificity, preventing consensus^22,29–31. Furthermore, previous reviews have primarily focused on summarizing the diagnostic utility of the traditional CDT or broadly discussing the advantages of digital assessment tools, while lacking a systematic synthesis of the diagnostic accuracy of dCDT specifically^21,27,32. This gap in evidence not only limits clinical adoption and standardization of dCDT but also hinders validation of its role in the early diagnostic evaluation of cognitive impairment.

Therefore, we conducted a systematic review and meta-analysis of available studies to critically appraise the diagnostic performance of the dCDT in detecting MCI and AD. The findings are intended to provide rigorous evidence for clinical practice and to inform the agenda for future research.

Results

Study selection

The search retrieved 4919 records from databases including Web of Science, Embase, PubMed, PsycINFO, IEEE Xplore, CNKI, and Wanfang, up to January 8, 2026. After removing 1351 duplicate records and 647 review studies or conference abstracts, 2921 records were screened by title and abstract. Following this screening, 326 full-text studies were assessed for eligibility. Additionally, one study was identified by screening the reference lists of relevant review articles. Ultimately, 13 studies met the inclusion criteria and were included in this systematic review and meta-analysis (Fig. 1).

Fig. 1. Literature screening process.

Illustration of the identification, screening, eligibility assessment, and inclusion of studies in the systematic review.

Characteristics of included studies

A total of 17 diagnostic tests from the 13 included studies provided data for the current analysis^{22,29–31,33–41}. Of the included diagnostic tests, 12 evaluated MCI, including nine assessing standalone dCDT and three assessing augmented dCDT; the remaining five diagnostic tests evaluated AD. Among the 17 included diagnostic tests, four presented results from multiple diagnostic models, of which two evaluated MCI and two evaluated AD. The detailed characteristics of the included studies are shown in Table 1. In addition, the included studies differed in feature extraction and input variables; detailed methods are provided in Supplementary Table 1.

Table 1.

Characteristics of included studies

Study	Region	Device type of dCDT (Details)	Type of dCDT (Details)	Analytical algorithm	Type of cognitive disorder	Diagnostic criteria
Davis et al.33	America	Pen (A digitizing ballpoint pen)	Standalone dCDT (The test instructs the subject to draw on a blank page a clock showing 10 min after 11, then asks the subject to copy a predrawn clock showing that time.)	Linear SVM	AD	Clinical diagnosis
Garre-Olmo et al.29	Spain	Tablet and pen (A commercial digitizing tablet and a pressure-sensitive ink pen)	Standalone dCDT (Participants were required to draw a clock face with all numbers correctly positioned and to set the hands to “10 past 11 o’clock.”)	Discriminant Analysis	MCI	Petersen criteria
Müller et al.22	Germany	Digitizer and pen (A Windows Surface Pro 4 digitizer and a handheld stylus pen)	Standalone dCDT (Participants were required to draw a clock face with all numbers correctly positioned and to set the hands to “10 past 11 o’clock.”)	ROC curve analysis	MCI	Petersen criteria
					AD	NINCDS-ADRDA
Müller et al.36	Germany	Tablet and pen (Microsoft Surface Pro 4 tablet and digital pen)	Standalone dCDT (Participants were required to draw a clock face with all numbers correctly positioned and to set the hands to “10 past 11 o’clock.”)	LR	MCI	Petersen criteria
					AD	NINCDS-ADRDA
Zhang et al.37	China	Tablet and pen (Huawei 10.8‑inch tablet equipped with a dedicated stylus)	Standalone dCDT (Participants were required to draw a clock face with all numbers correctly positioned and to set the hands to “10 past 11 o’clock.”)	LR	MCI	The 2011 NIA-AA criteria
Zheng et al.31	China	Tablet (The electromagnetic tablets)	Standalone dCDT (Participants were required to draw a clock face with all numbers correctly positioned and to set the hands to “10 past 11 o’clock.”)	XGBoost	AD	NINCDS-ADRDA
Lü40	China	Tablet and pen (A 10.8-inch tablet equipped with a dedicated stylus)	Standalone dCDT (Participants were required to draw a clock face with all numbers correctly positioned and to set the hands to “10 past 11 o’clock.”)	LR	MCI	Cognitive impairment (self/informant/clinician), preserved basic ADL, mild IADL impairment, and MMSE/MoCA/CDR
Liu38	China	Tablet and pen (HUAWEI Mate Pad Pro 10.8-inch tablet and HUAWEI M-Pencil stylus)	Standalone dCDT (Participants were required to draw a clock face with all numbers correctly positioned and to set the hands to “10 past 11 o’clock.”)	AlexNet-Trace-BP	AD	Alzheimer's Disease Diagnostic and Treatment Guidelines
Li et al.35	China	Tablet and pen (iPad 2019 tablet and Apple Pencil)	Standalone dCDT (Participants were required to draw a clock face with all numbers correctly positioned and to set the hands to “10 past 11 o’clock.”)	LR	MCI	MoCA; CDR; comprehensive medical evaluations conducted by experienced neurologists
			Augmented dCDT (Participants were required to perform a single dCDT task while completing a task related to language function.)
Um Din et al.34	France	Tablet (Two tablets of the same size)	Standalone dCDT (The researchers presented the participants with a predefined circle. The participants were required to use the stylus to draw all the numbers on a clock face on this circle. Then, they needed to draw the hands to indicate 11:10.)	The Greco and the Codex scoring systems	MCI	DSM-5 criteria
Chen et al.39	China	Tablet and pen (Professional stylus pen and 10.8-inch tablet)	Standalone dCDT (The participants were asked to draw a clock, input all the numbers, and set their hands to read 10 after 11.)	RF	MCI	Cognitive impairment (self/informant/clinician), MMSE, CDR, preserved ADL
			Augmented dCDT (The participants were asked to draw a clock, input all the numbers, and set their hands to read 10 after 11.)	RF	MCI
Zhang et al.41	China	Tablet and pen (The tablet devices and digital pen)	Augmented dCDT (The test consists of three steps: 1. The evaluator clicks the audio button to randomly play one of the six word lists and instructs the subject to memorize the words. 2. The subject uses a stylus to draw a clock indicating “11:10” on the tablet; 3. The subject recalls the three words memorized in Step 1)	LightGBM	MCI	Petersen criteria and DSM-4
Wu et al.30	China	Tablet	Standalone dCDT (Draw a clock face, including numbers and hands, setting the hands to 11:10.)	CLOX	MCI	ICD-11 and Peterson’s criteria

dCDT digital clock drawing test, AD Alzheimer’s disease, MCI mild cognitive impairment, SVM support vector machine, LR logistic regression, XGBoost eXtreme Gradient Boosting, BP back propagation, RF random forest, LightGBM light gradient boosting machine, NINCDS-ADRDA National Institute of Neurological and Communicative Disorders and Stroke - Alzheimer’s Disease and Related Disorders Association, NIA-AA National Institute on Aging-Alzheimer’s Association, MMSE Mini-Mental State Examination, MoCA Montreal Cognitive Assessment, IADL instrumental activities of daily living, ADL activities of daily living, CDR clinical dementia rating, DSM-5 Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, DSM-4 Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, CLOX an executive clock drawing task, ICD-11 The International Classification of Diseases, 11th Revision.

RoB of included studies

The majority of the included studies were judged to have a high risk of bias (RoB) (Fig. 2). The most common source of bias was related to patient selection (n = 13), as most studies employed a case-control design. Additional concerns were identified in the conduct of the index test (n = 11) as well as in flow and timing (n = 9). Nevertheless, the overall applicability of the included studies was considered good.

Fig. 2. Risk of Bias Assessment of Included Studies According to QUADAS-2.

The figure displays the overall risk of bias assessments across all included studies using the QUADAS-2 tool. a Risk of bias assessment-methodological quality graph. b Risk of bias assessment-methodological quality summary. QUADAS-2 Quality Assessment of Diagnostic Accuracy Studies 2.

Performance of dCDT in the diagnosis of MCI

A total of nine studies evaluating standalone dCDT were included^{22,29,30,34–37,39,40}, comprising 457 patients with MCI and 457 controls. Across the included studies, sensitivity ranged from 0.455 to 0.861 and specificity from 0.500 to 0.889. The pooled sensitivity and specificity of the standalone dCDT for the diagnosis of MCI were 0.765 (95% CI: 0.683–0.832) and 0.752 (95% CI: 0.673–0.817), respectively (Fig. 3). The pooled area under the summary receiver operating characteristic curve (AUC) was 0.825 (95% CI: 0.790–0.856), and the diagnostic odds ratio (DOR) was 9.883 (95% CI: 5.071–19.263) (Supplementary Fig. 1; Table 2). Substantial between-study heterogeneity was observed, with I² values of 68.61% for sensitivity and 68.75% for specificity. The Spearman correlation between logit(sensitivity) and logit(1−specificity) was weak and not statistically significant (ρ = −0.383, p = 0.309), providing no clear evidence of a threshold effect. Deeks’ funnel plot asymmetry test yielded a p value of 0.088 (Supplementary Fig. 2); therefore, the possibility of small-study effects cannot be excluded.

Fig. 3. Forest plot of diagnostic accuracy (sensitivity and specificity) for MCI in the standalone dCDT.

MCI mild cognitive impairment, dCDT digital clock drawing test.

Table 2.

Diagnostic performance of the dCDT for MCI and AD

Type of disease	Subgroup	No. of study	Sensitivity (95% CI)	Specificity (95% CI)	LR+ (95% CI)	LR− (95% CI)	DOR (95% CI)	AUC (95% CI)	p valuea
MCI	Overall	12	0.760 (0.691–0.818)	0.800 (0.722–0.861)	3.809 (2.673–5.428)	0.300 (0.227–0.395)	12.717 (7.386–21.895)	0.845 (0.810–0.874)
	Augmented dCDT	3	0.706 (0.656–0.752)	0.897 (0.861–0.927)	6.879 (5.012–9.441)	0.328 (0.279–0.386)	20.967 (13.878–31.667)	0.910R (0.643–0.922)	0.047
	Standalone dCDT	9	0.765 (0.683–0.832)	0.752 (0.673–0.817)	3.084 (2.193–4.337)	0.312 (0.217–0.449)	9.883 (5.071–19.263)	0.825 (0.790–0.856)
	Input variables and analytical algorithms
	Traditional-scoring approaches	2	0.667 (0.533–0.783)	0.557 (0.441–0.669)	1.505 (1.109–2.042)	0.598 (0.398–0.900)	2.514 (1.253–5.046)	0.579R (0.452–0.617)	0.001
	Algorithm-based approaches	7	0.797 (0.722–0.856)	0.809 (0.759–0.851)	4.177 (3.306–5.276)	0.251 (0.183–0.345)	16.648 (10.747–25.790)	0.856 (0.823–0.884)
	LR	4	0.837 (0.792–0.874)	0.797 (0.748–0.838)	4.117 (3.280–5.169)	0.205 (0.159–0.264)	20.124 (13.389–30.247)	0.887 (0.856–0.912)	0.253
	Non-LR	3	0.705 (0.591–0.803)	0.822 (0.715–0.902)	3.960 (2.370–6.616)	0.359 (0.250–0.514)	11.037 (5.099–23.891)	0.837R (0.675–0.900)
	Region
	China	5	0.718 (0.567–0.833)	0.775 (0.656–0.861)	3.192 (1.796–5.671)	0.363 (0.208 -0.635)	8.781 (2.953–26.116)	0.815 (0.779–0.847)	0.135
	Western countries	4	0.809 (0.725–0.872)	0.720 (0.621–0.801)	2.888 (0.977–4.220)	0.265 (0.167–0.421)	10.893 (4.871–24.359)	0.835 (0.800–0.865)
	RoB
	Relatively high RoB	5	0.734 (0.589–0.842)	0.742 (0.620–0.835)	2.846 (1.735–4.669)	0.358 (0.209–0.613)	7.953 (3.008–21.027)	0.802 (0.765–0.835)	0.723
	Relatively low RoB	4	0.798 (0.711–0.864)	0.766 (0.655–0.850)	3.414 (2.115–5.510)	0.263 (0.166–0.418)	12.973 (5.265–31.962)	0.852 (0.818–0.881)
AD	Overall	5	0.820 (0.721–0.889)	0.897 (0.864–0.923)	7.980 (5.646–11.280)	0.201 (0.124–0.325)	39.765 (18.434–85.777)	0.928 (0.902–0.948)

For subgroups with fewer than four studies, MIDAS could not be applied. Pooled estimates were therefore calculated by pooling the 2 × 2 tables (fixed-effect micro-average) and should be interpreted as exploratory; ^RAUC was estimated in R using a bivariate model (Reitsma) with bootstrap 95% CI and should also be interpreted cautiously; ^aThe p value for heterogeneity between subgroups was assessed using meta-regression analysis.

dCDT digital clock drawing test, AD Alzheimer’s disease, MCI mild cognitive impairment, LR+ positive likelihood ratio, LR− negative likelihood ratio, DOR diagnostic odds ratio, AUC area under the summary receiver operating characteristic curve, LR logistic regression, RoB risk of bias.

Supplementary analysis

After incorporation of three augmented dCDT studies^35,39,41, the pooled diagnostic performance across all included diagnostic tests (n = 12)^{22,29,30,34–37,39–41} yielded a sensitivity of 0.760 (95% CI: 0.691–0.818), specificity of 0.800 (95% CI: 0.722–0.861), and the pooled AUC increased from 0.825 for the standalone dCDT subgroup to 0.845 (95% CI: 0.810–0.874). (Table 2; Supplementary Fig. 3; Supplementary Fig. 4). An exploratory meta-regression demonstrated a significant difference in diagnostic performance between augmented dCDT (n = 3) and standalone dCDT (n = 9) (p = 0.047); however, this finding should be interpreted cautiously given the small number of augmented dCDT studies.

Sources of heterogeneity in the meta-analysis of the diagnostic performance of the standalone dCDT for MCI

First, in exploratory subgroup analyses by input variables and analytical algorithms, studies employing algorithm-based approaches (n = 7) yielded higher pooled diagnostic performance (AUC = 0.856; DOR = 16.648)^{22,29,35–37,39,40}, whereas the traditional-scoring subgroup (n = 2) yielded lower pooled diagnostic performance (DOR = 2.514) (Table 2; Supplementary Fig. 5)^30,34. In a sensitivity analysis excluding the two studies that used traditional-scoring, the pooled AUC increased from 0.825 to 0.856. An exploratory meta-regression suggested higher diagnostic performance for algorithm-based approaches compared with traditional-scoring approaches (p = 0.001); however, this comparison was based on only two traditional-scoring studies and should be interpreted cautiously. Among the seven studies using algorithm-based approaches, four employed logistic regression (LR)^35–37,40 and yielded a pooled AUC of 0.887, with pooled sensitivity and specificity of 0.837 and 0.797, respectively (Table 2; Supplementary Fig. 6). The remaining three studies assessed non-LR approaches: a random forest (RF) model (sensitivity = 0.595, specificity = 0.889; AUC = 0.811)³⁹, discriminant analysis (sensitivity = 0.818, specificity = 0.833; AUC not reported)²⁹, and an ROC curve analysis based on time in air (sensitivity = 0.813, specificity = 0.722; AUC = 0.880)²². Exploratory meta-regression indicated no significant difference between studies using LR and those using non-LR (p = 0.253).

Second, exploratory subgroup analyses showed no significant differences in pooled diagnostic performance by geographic region, comparing five studies from China^{30,35,37,39,40} with four from Western countries^22,29,34,36, or by RoB, comparing four studies with relatively low RoB^29,34,35,37 with five with relatively high RoB^{22,30,36,39,40}; meta-regression likewise identified neither geographic region nor RoB as a significant source of heterogeneity (Table 2; Supplementary Figs. 7 and 8).

Sensitivity analyses of standalone dCDT for MCI

A sensitivity analysis was conducted using a leave-one-out approach. The pooled sensitivity, specificity, DOR, and AUC varied only slightly after exclusion of each study sequentially (sensitivity: 0.760–0.789; specificity: 0.731–0.780; DOR: 8.437–13.038; AUC: 0.809–0.851; Supplementary Table 2).

Performance of dCDT in the diagnosis of AD

Among the five included studies^{22,31,33,36,38}, a total of 378 patients with AD and 940 controls were analyzed. The pooled sensitivity and specificity of the dCDT for AD diagnosis were 0.820 (95% CI: 0.721–0.889) and 0.897 (95% CI: 0.864–0.923), respectively (Fig. 4). The pooled AUC was 0.928 (95% CI: 0.902–0.948), and the DOR was 39.765 (95% CI: 18.434–85.777) (Table 2; Supplementary Fig. 9). Substantial heterogeneity was observed (I² = 77.19% for sensitivity and 61.46% for specificity). The Spearman correlation between logit(sensitivity) and logit(1−specificity) was negative and not statistically significant (ρ = −0.700, p = 0.188), providing no clear evidence of a threshold effect, although subtle threshold-related variation cannot be excluded. The Deeks’ funnel plot asymmetry test was not statistically significant (p = 0.767) (Supplementary Fig. 10).

Fig. 4. Forest plot of diagnostic accuracy (sensitivity and specificity) for AD in dCDT.

AD Alzheimer’s disease, dCDT digital clock drawing test.

Discussion

This systematic review and meta-analysis included 13 studies with 17 diagnostic tests evaluating the diagnostic accuracy of dCDT for MCI and AD. For MCI, pooled estimates from studies using standalone dCDT showed good diagnostic performance, with a sensitivity of 0.765, specificity of 0.752, and an AUC of 0.825. After incorporation of augmented dCDT studies, the overall pooled estimates were 0.760 for sensitivity, 0.800 for specificity, and 0.845 for AUC, with exploratory meta-regression indicating a significant difference between augmented and standalone dCDT. These findings suggest that augmenting dCDT with one additional task or one set of variables may offer incremental diagnostic value for MCI, although the evidence remains preliminary and should be interpreted cautiously. For AD, the corresponding values for pooled sensitivity, specificity, and AUC were 0.820, 0.897, and 0.928, yielding higher pooled point estimates than those for MCI. Clinically, the higher point estimates for AD may suggest that dCDT is more sensitive to moderate-to-severe impairment, whereas its sensitivity for early-stage MCI remains relatively limited⁴². Optimizing analytic algorithms to enhance early detection, therefore, represents a key area for future research²⁸.

The diagnostic advantages of dCDT may derive from its ability to capture dynamic drawing features such as pauses and speed, which are more sensitive to subtle cognitive deficits than the static final images of traditional CDT²². In addition, its quantitative and automated scoring reduces subjective bias and improves objectivity and reproducibility^20,27. Thus, dCDT holds promise not only as a clinical adjunct but also for large-scale screening and telemedicine applications^28,43,44. Compared with other digital cognitive tests (e.g., digital Rey–Osterrieth Complex Figure Test, AUC = 0.82³⁵; digital Mini-Mental State Examination, AUC = 0.82³⁰), the dCDT is brief to administer yet captures rich spatiotemporal drawing-process information. Its diagnostic performance appears broadly comparable across digitized screening tools²⁷.

Substantial heterogeneity was observed across studies. Exploratory subgroup and sensitivity analyses suggested that input variables and analytical algorithms may partly contribute to this heterogeneity: algorithm-based studies showed good diagnostic performance, and excluding the two traditional-scoring studies increased the pooled AUC from 0.825 to 0.856, which suggests that the overall accuracy estimate was somewhat higher in studies that used features beyond traditional-scoring approaches. This pattern may reflect that algorithm-based approaches can leverage richer, time-resolved drawing-process features (e.g., time in air, number of strokes, and pressure-related metrics) that may capture subtle motor–cognitive changes associated with MCI, potentially improving diagnostic sensitivity²⁷. Exploratory meta-regression also favored algorithm-based approaches, but this finding should be interpreted cautiously, given the very small number of traditional-scoring studies and the risk of model instability. Within algorithm-based approaches, LR showed good pooled accuracy (AUC = 0.887), whereas evidence for other algorithms (e.g., RF, discriminant analysis) was limited to single studies, precluding algorithm-specific subgroup meta-analyses.

Exploratory analyses suggested no meaningful differences in diagnostic performance between studies conducted in China and those from Western countries. Therefore, regional variation should not be over-interpreted based on the current evidence base. Nonetheless, cross-setting differences (e.g., participant characteristics, education, cultural context) may still influence performance^28,45–47 and should be addressed in future multicenter studies using standardized protocols and more diverse populations. Similarly, exploratory analyses suggested no meaningful differences between the relatively low and high RoB groups, likely because the overall RoB was high across studies, such that both subgroups remained subject to bias, thereby limiting discrimination between them. In addition, leave-one-out sensitivity analysis indicated that the diagnostic accuracy of the standalone dCDT for MCI diagnosis was not substantially affected by individual studies. However, these findings should be interpreted cautiously and not over-emphasized, given the limited evidence base.

Differences in device types (e.g., Huawei tablets or iPads) and their sampling characteristics may influence how temporal and kinematic features are captured⁴⁷. Scoring algorithms and feature-processing pipelines also varied across studies, potentially shifting sensitivity and specificity through different feature sets or cut-point optimization⁴⁵. Educational level may further affect drawing efficiency and dynamics because clock drawing depends on visuospatial construction, executive planning, and familiarity with analog clocks; lower education or limited clock-reading experience may lead to slower or less organized drawing patterns independent of neurodegeneration⁴⁸. Collectively, these factors may also contribute to between-study heterogeneity.

Several limitations should also be acknowledged. First, the number of eligible studies was limited, particularly for AD, which may reduce statistical power. Second, most included studies employed case-control designs and were rated as high RoB in the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2), particularly in patient selection and index test domains; such designs may overestimate diagnostic accuracy, and the pooled estimates should therefore be interpreted cautiously. Third, Deeks’ funnel plot asymmetry test suggested possible small-study effects for the MCI analyses, whereas no significant asymmetry was observed for the AD analyses; nevertheless, selective reporting of positive results cannot be ruled out. In addition, Deeks’ test is underpowered when only a few studies are available—especially for the AD analyses—so non-significant results should not be interpreted as definitive evidence of absence of bias. Fourth, small sample sizes and single-center settings in many studies may restrict the generalizability of findings. Subgroup and meta-regression analyses should be interpreted as exploratory, as the small number of studies and correlation between study characteristics (e.g., analytical algorithm and region) may lead to residual confounding. Finally, although ranking the diagnostic accuracy of dCDT and other screening tools is clinically relevant, the scarcity of head-to-head comparative studies limits the feasibility of conducting a network meta-analysis.

Future research should prioritize prospective, multicenter studies with large samples to improve representativeness and robustness. In addition, there is relatively limited foundational work on standardizing task procedures, data acquisition, and feature extraction. More research in this area is needed to facilitate the broader dissemination and scaling of dCDT diagnostic models. From a methodological perspective, leveraging artificial intelligence with deep learning (e.g., sequence- and multimodal models) could further boost the sensitivity of dCDT to early MCI⁴⁹. Moreover, combining dCDT with other biomarkers—such as neuroimaging, blood-based, or cerebrospinal fluid markers—may enable the development of a more comprehensive framework for early diagnosis of cognitive disorders^50–52. Importantly, future studies should increase standardized head-to-head comparisons between dCDT and established tools to support evidence synthesis (e.g., network meta-analysis) aimed at identifying the most effective screening tool.

In conclusion, this study provides systematic evidence on the diagnostic utility of dCDT for MCI and AD, demonstrating overall good diagnostic performance. Current evidence suggests that dCDT demonstrates relatively good diagnostic accuracy for AD, whereas its performance for MCI remains comparatively modest, underscoring the need for further high-quality, standardized, prospective multicenter studies to improve and validate MCI detection. Overall, these findings support further development of dCDT toward practical implementation, highlighting its strong potential for large-scale screening and clinical application.

Methods

The systematic review and meta-analysis were performed according to the standard guidelines for the systematic review of diagnostic studies, including the Preferred Reporting Items for Systematic Reviews and Meta-Analyses⁵³ (Supplementary Table 3) and the guidelines proposed by the Cochrane Diagnostic Test Accuracy Working Group⁵⁴. This study is registered as CRD420251053336 in PROSPERO.

Search strategy

We searched international databases (Web of Science, Embase, PubMed, PsycINFO, and IEEE Xplore) as well as Chinese databases (CNKI and Wanfang) from their respective inception dates until January 8, 2026. Keywords included “Alzheimer’s Disease,” “Alzheimer,” “AD,” “Cognitive Dysfunction,” “Cognitive Impairment,” “MCI,” “dCDT,” “digital clock,” “clock test,” “sensitivity and specificity,” “accuracy,” and “diagnostic”. Additionally, the reference lists of eligible studies and relevant reviews were manually searched to identify any additional studies. No language restriction was adopted. To ensure transparency and reproducibility, the complete search strings used for each database are provided in Supplementary Table 4.

Inclusion and exclusion criteria

Studies were included if they met the following inclusion criteria:

1. The study aimed to evaluate the diagnostic accuracy of the dCDT for MCI or AD, with the index test being either standalone dCDT or augmented dCDT, defined as dCDT combined with one additional cognitive task or one prespecified set of non-drawing variables.

2. The study included clinically diagnosed patients with MCI or AD.

3. Studies must include at least one group of healthy control participants (or individuals with normal cognitive function) and explicitly compare these to the group with AD or MCI.

4. Participants with MCI or AD were confirmed based on standardized diagnostic criteria, including the Petersen Criteria⁵⁵ and the recommendations of the National Institute on Aging and Alzheimer’s Association⁵⁶, among others.

5. The study must provide data that allows for the extraction of true positives, false positives, true negatives, and false negatives to construct a 2 × 2 contingency table, or provide diagnostic indicators such as sensitivity and specificity.

6. When studies reported both training and validation results, only the validation set data were extracted. In cases where multiple models or thresholds were presented, we extracted the results from the author-designated primary or clinically recommended model to minimize the RoB from post-hoc model selection.

Studies were excluded if dCDT was combined with two or more additional tasks or variable sets, because the primary objective was to evaluate the diagnostic performance of dCDT; such multi-component tests would substantially increase heterogeneity and obscure whether dCDT remains the primary contributor to the observed diagnostic performance.

Data extraction

Two investigators (CSH and LHL) independently assessed the relevance of the retrieved records and extracted the data into an Excel spreadsheet. The spreadsheet was used to document key characteristics of the included studies, including year of publication, study location, task details, and devices used for the dCDT, and the diagnostic criteria applied to define patients with MCI or AD. In addition, we extracted the numbers of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN) for each dCDT to enable subsequent analysis. Because several included studies were published in Chinese, the full texts were reviewed directly by two bilingual investigators whose native language is Chinese, and no machine translation tools were used. Any discrepancies in study eligibility or data extraction were resolved through discussion, and when consensus could not be reached, a third investigator (ZBL) made the final decision.

RoB and reporting quality

We used the QUADAS-2 tool⁵⁷ to assess the RoB of the included studies. QUADAS-2 evaluates methodological quality across four domains: Patient Selection, Index Test, Reference Standard, and Flow and Timing, and also considers concerns regarding applicability. In the primary assessment, RoB judgments were reported descriptively at the domain level, rather than combined into a composite score, in accordance with QUADAS-2 guidance. Because all included studies were overall judged to be at high RoB, we conducted an exploratory subgroup analysis to explore whether pooled diagnostic accuracy differed according to the relative extent of bias within this high-RoB set. For this purpose, studies with low RoB in at least two QUADAS-2 domains were operationally categorized as relatively low RoB, whereas studies with low RoB in one or fewer domains were operationally categorized as relatively high RoB. This post-hoc categorization was used solely for exploratory comparisons and should not be interpreted as a formal QUADAS-2 summary score.

Outcomes

The outcomes of interest were diagnostic accuracy indicators from the included studies, including AUC, sensitivity, and specificity.

Data synthesis and statistical analysis

A bivariate mixed-effects model was used to combine the sensitivity and specificity of each dCDT⁵⁸. Forest plots were used as a graphical representation to pool the sensitivity and specificity. The DOR was used as an indicator of the test performance of the dCDT for different cognitive impairments⁵⁹. A summary receiver operating characteristic curve was generated, and the AUC was calculated. Statistical heterogeneity among the trials was assessed using I². Threshold effects were examined by calculating Spearman’s rank correlation (ρ) between logit(sensitivity) and logit(1−specificity); a significant positive correlation (p < 0.05) would indicate heterogeneity driven by varying diagnostic thresholds across studies. The subgroup analysis and meta-regression were conducted to explore the sources of heterogeneity. The meta-regression models were used to compare the diagnostic performance differences between subgroups, with p < 0.05 indicating statistical significance. Publication bias was investigated through Deeks’ funnel plot asymmetry test, with p < 0.10 indicating significant asymmetry⁶⁰. Sensitivity analysis for the diagnostic performance of the dCDT in MCI was conducted using a leave-one-out approach. In addition to the primary meta-analysis of standalone dCDT, a supplementary analysis incorporated both standalone and augmented dCDT diagnostic tests. For studies in which standalone dCDT and augmented dCDT were evaluated in the same participant cohort, the overlapping case and control samples were split between the two comparisons to avoid double-counting and unit-of-analysis errors. The split was deterministic: when the total number was even, it was divided equally; when the total number was odd, the extra participant was assigned to the augmented dCDT comparison and the remaining participant to the standalone dCDT comparison, with the same rule applied separately to cases and controls. TP and TN were derived from the allocated denominators using the reported sensitivity and specificity, respectively, and rounded to the nearest integer; FN and FP were calculated as the residual counts. This approach was used in line with Cochrane methodological guidance to prevent reuse of the same participants across multiple comparisons^61,62. The main statistical analyses were performed with the Midas procedure in STATA version 18.

Author contributions

All authors contributed to the study conception and design. S.-H.C., H.-L.L., and H.Q. developed the protocol with guidance from and under the supervision of B.-L.Z. S.-H.C., H.-L.L., and H.Q. conducted the literature screening, searched the electronic databases, and performed backward and forward reference list checking. The study selection process, data extraction, and risk of bias assessment were carried out by S.-H.C., H.-L.L., and B.-L.Z. Data synthesis was conducted by H.Q. and Z.-T.L. The introduction section was written by S.-H.C.; the “Methods,” “Results,” and “Discussion” sections were written by H.-L.L., H.Q., Z.-T.L, and B.-L.Z. J.L. guided and actively participated in the manuscript revision, contributing to additional subgroup analyses and strengthening the “Discussion.” All authors have read and approved the manuscript for publication and agree to be accountable for all aspects of the work.

Data availability

The raw data extracted from the included studies to support the findings of this systematic review and meta-analysis are provided in Supplementary Table 5.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41746-026-02687-2.