Evaluating digital cognitive tests for clinical use in Alzheimer's disease: A novel framework and scoping review

Abstract

Mild cognitive impairment (MCI), a prodromal stage of Alzheimer's disease (AD), remains undiagnosed in > 90% of individuals, delaying access to timely evaluation and interventions. Self‐administered digital cognitive assessments (SA‐DCAs) offer scalable approaches for early detection, yet their real‐world validation and clinical readiness remain uncertain. We developed a use‐case–specific framework to evaluate SA‐DCAs intended for community and primary‐care MCI screening and applied it to a comprehensive scoping review of published evidence (2012‐2025). Among 79 identified SA‐DCAs, only four tools met predefined framework criteria across nine eligible studies. Common limitations included restricted population representativeness, inconsistent diagnostic performance reporting, limited biomarker anchoring, and reliance on prefiltered cohorts. Overall, the current evidence base is methodologically heterogeneous and incomplete for clinical deployment. The proposed framework characterizes requirements including anchoring strength, prevalence‐adjusted performance reporting, and representative sampling establishing a foundation for advancing robust real‐world evidence needed to translate SA‐DCAs from research to clinical practice.

Highlights

• Self‐administered Digital Cogntive Assessment (SA‐DCA) tools may be able to detect MCI.

• We evaluated SA‐DCA evidence for use in real‐world (RW) care using a novel framework.

• Scoping review identified 79 SA‐ DCAs, four met the framework criteria.

• Studies lacked appropriate comparators, biomarker anchoring, and/or performance metrics.

• The framework provides a pathway to generate evidence for using SA‐DCA in RW settings.

1. BACKGROUND

Mild cognitive impairment (MCI) is defined as a decline in cognitive function that exceeds normal age‐related changes but does not significantly interfere with instrumental activities of daily living. ¹ Diagnosis typically involves clinical history, cognitive testing, neurological examination, laboratory evaluation, and, when indicated, neuroimaging. In many cases, MCI is preceded by subjective cognitive decline (SCD), a perception of decline that may or may not be accompanied by objective cognitive deficits. Once MCI is confirmed, reversible causes such as sleep deprivation, metabolic abnormalities, medication side effects, and sensory impairments are assessed and treated appropriately. ² While some individuals revert to normal cognition, the majority (55–65%) progress to dementia. ¹ , ³ Among adults aged 50‐80 years with MCI, approximately half show evidence of an amyloid‐β pathology consistent with prodromal AD or MCI due to AD (MCI‐AD). ⁴ , ⁵ , ⁶ This subgroup may be eligible for treatment with symptomatic and, more recently, disease targeted therapies. ⁷ , ⁸

Early detection of MCI offers critical opportunities for intervention ³ , ⁷ , ⁸ , ⁹ where addressing some modifiable risk factors may reduce dementia risk. ¹⁰ , ¹¹ Yet over 90% of MCI cases remain undiagnosed. ¹² Failing to identify MCI‐AD results in missed treatment opportunities. Furthermore, a recent report estimates that in the U.S. it might take people on average six years to know if they are eligible for anti‐amyloid therapies if the provider uses only brief cognitive assessments to guide referrals. ¹³

Compounding these challenges is the absence of sensitive, standardized, quick, and scalable cognitive assessment tools suitable for primary care. ¹⁴ , ¹⁵ , ¹⁶ Widely used traditional instruments (Table 1), such as the Mini‐Mental State Examination or Mini‐Cog lack sensitivity for early‐stage cognitive changes as they were mostly designed to detect and stage moderate to severe cognitive impairment (CI). ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ Additionally, these paper‐based tools require trained administration, clinicians time, ¹⁰ and are vulnerable to inter‐rater variability and ceiling effects. ²¹ , ²² , ²³ , ²⁴

TABLE 1.

Common cognitive assessments used in clinical practice (non‐exhaustive) – all require administration by trained clinicians, not applicable for MCI.

Tool *	Time to administer	Who administers	Domains assessed	Strengths	Limitations
MoCA	∼10‐15 min	Clinician or trained staff	Memory, executive function, attention, language, visuospatial	More sensitive to MCI than MMSE	Requires training and permission to use, less sensitive to MCI
MMSE	∼7‐10 min	Clinician	Orientation, recall, attention, language, basic calculation	Widely used; easy to administer	Poor sensitivity to MCI, ceiling effects
Mini‐Cog	∼3 min	Clinician or medical assistant	Recall, clock drawing	Fast and easy; minimal training required	Less sensitive than MoCA, limited domains
GPCOG	∼5 min	Clinician or nurse	Cognitive function + informant report	Designed for primary care; includes informant input	May not detect subtle impairment
AD8	∼3‐5 min (informant)	Informant or caregiver	Functional changes across multiple domains	Quick; informant‐based; useful for MCI clues	Requires a knowledgeable informant

^*In many health systems, particularly outside high‑income countries, the availability of trained examiners and protected clinical time further limits use of traditional instruments. Furthermore, many countries have different versions of these instruments.

**Sensitivity to MCI is relative to scope, scores, and data available along performance reported (whether it can detect MCI correctly).

Abbreviations: AD8, Ascertain Dementia 8‐Item; GPCOG, General Practitioner Assessment of Cognition; MCI, mild cognitive impairment; Mini‐Cog, Mini short Cognitive screening test; MMSE, Mini‐Mental State Examnation; MoCA, Montreal Cognitive Assessment.

Likewise, globally used cognitive assessments, even if translated often carry forward education‐, language‐, and culture‑related biases and depend on trained examiners who are unequally distributed across countries and care settings. Other factors that contribute to diagnostic gaps include limited provider training, disparities in healthcare access, time‐constrained visits and prolonged specialist waitlists. ¹⁰ These constraints highlight the need for scalable, self‑administered tools that reduce examiner burden and support culturally and linguistically adapted cognitive testing to identify those who would benefit from further evaluation. ¹⁶

Differentiating MCI‐AD from other causes of cognitive decline or from typical aging remains a significant clinical challenge due to the subtle and heterogeneous nature of its etiology and clinical presentation. ²⁵ Diagnostic modalities such as cerebrospinal fluid (CSF) analysis and positron emission tomography (PET) imaging can be used to confirm the presence of an amyloid‐β pathology. However, access to these gold‑standard tests remains limited or logistically infeasible (even in many high income countries, such as rural settings), creating substantial inequities. However, the advent of highly accurate blood‐based biomarkers (BBMs) used to confirm an amyloid‐b pathology offers a new opportunity to identify an amyloid‐β pathology and improve diagnostic precision. ⁹ , ¹⁰ , ¹¹ , ¹² , ¹³ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ , ²⁰ , ²¹ BBM confirmatory tests can be used to rule in or out AD as the likely cause of an individual's MCI, ²⁶ reducing unnecessary costs and invasive diagnostic workup. Despite their promise, BBMs are not yet widely implemented in clinical practice. ¹² , ²¹

Self‐administered digital cognitive assessments (SA‐DCAs) have emerged as promising tools to help bridge these gaps. These tools, ranging from digital adaptations of paper‐based tests to artificial intelligence (AI) ‐driven platforms, enable scalable, often low‐burden screening across single or multiple cognitive domains in approximately 5‐15 minutes. ²⁷ A key advantage of SA‐DCAs is their ability to extend cognitive testing beyond the clinic via mobile device and web‐based application deployment. ²⁸ , ²⁹ SA‐DCAs can reduce the need for in‐person visits and examiners. ²⁷ , ³⁰ When designed for ease‐of‐use, they can increase access to screening, reduce healthcare burden, and allow for frequent testing in familiar settings. ³¹ Some SA‐DCAs incorporate “zero‐effort” technologies that use passive data collection (e.g., activity monitoring), which require minimal user engagement while providing complementary data that can enhance the assessment's ecological validity and clinical utility.

Large‐scale studies have shown that SA‐DCAs are feasible for population‐level MCI screening, with good performance (e.g., area under the curve [AUC] = 0.85, accuracy in MCI classification compared to those cognitively unimpaired) and participant engagement across diverse populations. ³² , ³³ Nevertheless, their adoption remains limited outside of research and academic medical centers. Current validation approaches, developed primarily for traditional cognitive instruments, do not adequately address the unique characteristics and deployment settings of self‐administered digital tools. These include incorporating equity safeguards such as rigorous translation procedures, validation using local normative data, differential item functioning (DIF) analyses, and device‑agnostic delivery as essential components to ensure that SA‐DCAs can be deployed fairly and effectively across global populations. Furthermore, the U.S. Preventive Services Task Force has highlighted a lack of high‐quality evidence supporting cognitive MCI screening tools (including DCAs) and the need for consistent, replicable validation studies in primary care settings. ³⁴

Most validation studies of SA‐DCAs have adopted paradigms developed for clinician‐administered neuropsychological tests without assessing whether these approaches are appropriate for self‐administered screening tools deployed at scale. Importantly, SA‐DCAs involve different tradeoffs versus other types of evaluation; neuropsychological tasks function as pragmatic comparators grounded in observable behavior, whereas biological anchoring reflects pathophysiological evidence. Each form of anchoring contributes differently to validation and are not interchangeable. Clarifying this distinction is critical in the context of screening for MCI, where pathophysiological confirmation increasingly informs therapeutic eligibility and care pathways.

Therefore, we sought not only to catalog SA‐DCAs, but to critically examine the methodological assumptions underlying their validation. Specifically, we developed a use‐case–specific framework that distinguishes biological, clinical, and proxy anchoring strategies; emphasizes equity in deployment; and aligns performance reporting with screening‐context prevalence and decision‐making needs. By applying this framework to the current evidence base, we aimed to clarify what level of evidence exists and validation is necessary for SA‐DCAs to transition from research tools to clinically useful screening instruments.

2. METHODS

2.1. Study design

This review aimed to map and evaluate published evidence on SA‐DCAs designed to detect MCI in community and primary care settings. This manuscript was prepared according to Preferred Reporting Items for Systematic Reviews and Meta‐analysis for Scoping Reviews (PRISMA‐ScR) Checklist, ³⁵ which can be found in Supplementary Table 1. This study followed PRISMA‐ScR guidelines to ensure methodological transparency and reproducibility. The scoping review format was selected to capture the breadth and heterogeneity of evidence across diverse technologies and study designs rather than to quantitatively synthesize effect sizes. The process consisted of five key steps: (1) identifying the objectives; (2) identifying relevant studies eligibility criteria, information sources, search criteria; (3) selecting studies; (4) charting the data; and (5) synthesis, summarizing, collating, and reporting the results.

2.2. Objectives

The objectives were to: (1) Identify and categorize SA‐DCAs applicable to community or primary‐care screening for MCI; (2) Describe and map the extent, characteristics, and methodological quality of the evidence supporting these tools; and (3) Apply a novel framework—informed by regulatory, clinical, and digital medicine standards—to evaluate readiness for real‐world implementation.

2.3. Information sources and search strategy

We conducted a structured multi‐source search from January 2012 to March 2025, aligning with the publication from the 2011 National Institutes of Aging Alzheimer's Association (NIA‐AA) MCI criteria. The search covered scientific databases (PubMed and ClinicalTrials.gov), technology and innovation sources: (Google Scholar, Google Patents, Crunchbase, Digital.Health, CBInsights, Growjo, and LinkedIn). The following key terms and Boolean operators were applied:

(“Screening” OR “Diagnosis” OR “Identification” OR “Validation” OR “Validity”) AND (“Cognitive Dysfunction” OR “Mild Cognitive Impairment” OR “Cognitive Impairment” OR “Mild Neurocognitive Disorder” OR “Early MCI” OR “MCI” OR “MCI due to AD” OR “Early AD” OR “Cognitive Decline”) AND (“Digital” OR “Wearable” OR “Voice” OR “Vocal” OR “Actigraph” OR “Digital Cognitive Test” OR “Remote” OR “Self‐administered” OR “Smartphone” OR “App” OR “Digital Biomarker” OR “Digital Test” OR “Smart Device” OR “Remote Assessment” OR “Remote Administration” OR “At‐Home”).

Two sequential searches were conducted, tool identification to map all SA‐DCAs (using the search criteria above) and peer reviewed publication evidence retrieval (using the same search criteria along with the SA‐DCAs identified in the first search, listed in Table 2a,b) to identify validation publications for each tool.

TABLE 2A.

Study performance and characteristics summarized a. Study characteristics summarized for each SA‐DCA study identified.

SA‐DCA	Yr.	Testing Location	Longitudinal data and duration	Sample size	Conv (Y/N)	BBM (Y/N)	Imag‐ing (Y/N)	Details on blood/imaging biomarkers	Diagnostic criteria	Se/Sp Reported	Conditions compared	Gold standard test	Effect size on gold standard comparison	Single vs. multiple test battery	Name of key tests
CANTAB 47	2021	Specialist	Y, 36 mo.; time points at 0/18/36 mo.	149 SCC‐stable; 32 MCI‐stable; 27 MCI‐worsened	N	N	N	NA	Standard criteria (Albert et al. 2011; Dubois et al. 2007; Petersen, 2004)	Y	SCC‐stable vs. MCI‐stable vs. MCI‐worsened	NA	NA	Multiple	PAL; PRM; DMS; SSP
CANTAB 54	2024	At‐home	Y; every 6 mo.	MCI: 496; 54: AD; SMC: 6162; AD family history: 5038	N	N	N	NA	Self‐report	N	Self‐report MCI vs. no self‐report MCI	NA	NA	Single	PAL
CBB 49	2018	At‐home	N	n = 6463; 5748 CN, 638 MCI, 77 AD	N	N	N	NA	BHR medical questionnaire	NA	MCI vs. CN	NA	NA	Multiple	L/WM Composite
CBB 50	2020	Specialist	Y, 3.5 yrs.	Future diagnoses: 450 CU stable; 150 CU converter. Biomarker profile: 146 CU A–T–; 61 CU A+T–; 33 CU A+T+	Y, CU to MCI/ dementia	N	Y	PiB‐PET and tau‐PET are acquired with a PET/CT, 3‐D modality analyzed	study coordinator, examining physician, and neuropsychologist diagnosed CU, MCI, dementia	N	CU stable vs. converter	NA	NA	Multiple	ONB and OCL tests
CBB 48	2021	Specialist	N	CU, n = 2866; MCI, n = 226; CU A–T–, n = 146; MCI A+T+, n = 15; MCI A–T–, n = 14	N	N	Y	Amyloid‐PET, PiB and tau PET (AV1451) scans were used	Clinician diagnosis: CU, MCI, or dementia. HCPs blinded to prior clinical, diagnoses, biomarker info	Y	MCI vs. CU	NA	NA	Multiple	L/WM Composite
CBB 53	2024	PCP	N	CN: 338; MCI: 185; AD: 7	N	Y	Y	All—Aβ‐ Tau‐PET, by 90 d after CBB test. Stratified by FBP (n = 295) FBB (n = 235), SUVR	N/A	N	MCI vs. CN	NA	NA	Multiple	CBB (average of all four tasks)
CBB 52	2024	PCP	N	Total: 3982 (size of each group is not available)	N	N	Y	A4/ADNI cohorts, scans, Aβ‐PET positivity based on SUVr across all cortices, cerebellar gray matter	CL > 15 for PET imaging – FBP, FBB tracer from A4 ADNI, equivalent to an SUVr threshold of 1.08 of FBP PET	N	Aβ+ vs. Aβ‐ among CU	Amyloid PET imaging	d  =  0.177 (CBB L/WM)	Multiple	L/WM, BPI
ki:e SB‐C 46	2023	At‐home	N	Memory clinic: n = 121; MCI = 52; SCI = 69; Birth cohort: n = 404; MCI = 48; CI = 356	N	N	N	NA	DeepSpA: MCI—DSM‐5 criteria. H70 sample: MCI was classified based on Petersen criteria.	Y	MCI vs. SCI; MCI vs. CI	0.71‐0.78	NA	Single	ki:e SB‐C
MyCog Mobile 51	2024	At‐home	N	Suspected CI: 20; unsuspected CI: 180	N	N	N	NA	Mini‐Cog < 3	Y	Suspected CI vs. no suspected CI	Mini‐Cog	NA	Multiple	MyCog Mobile

Abbreviations: AD, Alzheimer's Disease; ADNI, Alzheimer's Disease Neuroimaging Initiative; aMCI, amnestic mild cognitive impairment; A/T, amyloid or tau; AUC, area under the curve; Aβ , Amyloid β; BPI, [^{1 1}C]‐PiB; CBB, Cogstate Brief Battery; CL, Centiloid; combo, combination; CI, cognitive impairment; Conv, conversion to AD; CU, cognitive unimpaired; d, day; DeepSpA, Deep Speech Analysis; FBB, [^{1 8}F]‐florbetaben; FBP, [^{1 8}F]‐florbetapir; HC, healthy control; ki:e SB‐C , ki elements speech biomarker‐cognition; L/WM, learning / working memory; MCI, mild cogntive impairment; N, no; NA, not applicable; ONB, one back task; OCL, one card learning; PAL, paired associative learning; PET, positron emissions tomography; PiB, Pittsburgh compound b; PCP, primary care physician; SCI, subjective cognitive impairment; se/sp, sensitivity / specificity; SUVR, standardized uptake value ratio; Y, yes; Yr, year.

TABLE 2B.

Reported performance in each study.

Company	Publication Year	Accuracy	AUC	Sensitivity	Specificity	Positive predictive value	Negative predictive value
CANTAB 47	2021	NA	0.86 for combo of all 4; PAL AUC = 0.78 (6 items); PRM AUC = 0.79; DMS AUC = 0.71; SCD AUC = 0.71	NA	NA	NA	NA
CANTAB 54	2024	NA	0.68	NA	NA	NA	NA
CBB 49	2018	68.3% (95% CI [66.0%–70.6%] (compared to 57% demographics data only)	NA	NA	NA	NA	NA
CBB 50	2020	NA	0.70 (0.65, 0.74)	0.25 (0.18, 0.32)	0.88 (0.85, 0.91)	NA	NA
CBB 48	2021	NA	0.75 (0.71, 0.78)	0.38 (0.32, 0.45)	0.91 (0.90, 0.92)	NA	NA
CBB 53	2024	NA	0.7	NA	NA	NA	NA
CBB 52	2024	NA	Cohen's d  =  0.172	NA	NA	NA	NA
ki:e SB‐C 46	2023	0.71‐0.78	0.73‐0.81	0.62‐0.79	0.72‐0.86	0.28‐0.76	0.75‐0.96 (11.9% in H70 and 42% in DeepSpA)
MyCog Mobile 51	2024	NA	0.83	0.76	0.88	NA	NA

Note: performance is not meant to be comparative given different study designs. Only those with unfiltered populations and specific comparisons that either partially or fully met criteria are included (see Figure 1B and Figure 3 for categorization). NA means not reported.

Abbreviations: AUC, area under the curve; CANTAB, Cambridge Neuropsychological Test Automated Battery; CBB, Cogstate Brief Battery; CI, confidence interval; combo, combination; NA, not applicable.

2.4. Eligibility criteria

Studies were included if they met all the following criteria: evaluated a self‐administered and remotely deployable SA‐DCA with algorithmic or automated scoring, published in more than one manuscript in peer‐reviewed journals, reported diagnostic performance metrics and assessed performance in populations reflecting primary‐care or community screening contexts. More specifically these will be referred to as “unfiltered populations” (e.g., those that are not pre‐selected based on clinical diagnosis or biomarker status).

Studies were excluded if: SA‐DCA tools required examiner supervision or specialized hardware (e.g., electroencephalogram [EEG], eye‐tracking), studies with only feasibility or usability outcomes without diagnostic performance data (e.g., accuracy), or population comparisons outside the intended use case (e.g., dementia vs. healthy controls). Two independent reviewers (B.M., V.M.) screened titles and abstracts using predefined inclusion criteria. Full‐texts of potentially eligible studies were retrieved and reviewed. Discrepancies were resolved through discussion and, if necessary, consensus with a third reviewer (M.D.).

2.5. Data charting process

A standardized data charting form was developed and piloted prior to full data extraction. Extracted variables included: SA‐DCA name, platform, and administration mode, comparator groups and reference standards, performance metrics, use of biomarkers or neuroimaging; and population characteristics (filtered vs. unfiltered) and study characteristics (year, setting, sample size, demographics). Two reviewers (B.M., V.M.) independently extracted data, cross‐checked for consistency, and consolidated in summary tables.

2.6. Critical appraisal and framework application

Formal meta‐analysis and GRADE certainty assessments were not conducted; we used risk‐of‐bias tools (QUADAS‐2) to evaluate methodological rigor and risk of bias across studies. To contextualize the evidence, we developed and applied a SA‐DCA Framework, integrating guidance from regulatory authorities (United States Food and Drug Administration [FDA], European Medicines Agency [EMA], National Institute for Health and Care Excellence [NICE]) and digital medicine (Digital Medicine Society, or DiME) as well as established reporting standards (Supplementary Table 2). The framework evaluates four core domains relevant to real‑world implementation—population selection, comparator appropriateness, performance reporting, and clinical anchoring—with additional considerations for equity, representativeness, and usability in self‑administered settings (Figure 1).

FIGURE 1.

Proposed framework and model classification. (A) The framework distinguishes biological (Level 1), clinical (Level 2), and proxy anchoring (Level 3) and were performed during abstractions; validating SA‑DCAs solely against traditional cognitive tests risks embedding existing measurement biases, underscoring the need for biological biomarker anchoring or bridging analyses. Methodological representation of population categories, “filtered” versus “unfiltered” and performance classification. This figure illustrates the distinction between common population selection strategies used in clinical research versus real‐world applications. On the left, different types of population comparisons are shown (green represent the populations we are focused on as part of the framework's inclusion criteria, red text refers to comparison excluded). In the middle are “filtered”, pre‐selected populations based on known clinical diagnosis or biomarker confirmation. Whereas “unfiltered” populations represent the real‐world primary care setting, where individuals are not pre‐stratified, and early MCI cases are unidentified or blinded. Aβ, amyloid‐beta; ACC, accuracy; AD, Alzheimer's disease; ADHD, attention deficit hyperactivity disorder; BD, bipolar disorder; AUC, area under the curve; CDR, Clinical Dementia Rating; CN, cognitively normal; CU, cognitively unimpaired; HC, healthy controls; MCI, mild cognitive impairment; mda‐MCI, multi‐domain MCI; PD, Parkinson's disease; preAD, preclinical AD; proAD, prodromal AD; PSCI, post‐stroke cognitive impairment; ROC, receiver operator curve; SCD, subjective cognitive decline; SCI, subjective cognitive impairment; sda‐MCI, single‐domain MCI. (B) Categorization of SA‐DCAs identified according to framework criteria. Applying the proposed framework criteria to categorize the 37 SA‐DCAs. Performance metrics were based on cutoffs specified in the table for each color coated category (left side, chart) and the performance requirements across the three‐color coated categories were used to determine, in part, whether the SA‐DCAs met the framework criteria. Three categories were assigned (“does not meet”, “minimally meets”, “partially meets”, “fully meets”) in color coded text (right side) based on performance, whether the cohort was pre‐filtered (e.g., a pre‐selected clinical trial population), what population comparison used in the study—the “focus population” (see Figure 1a). Those that were from a population outside our focus were removed. SA‐DCA, self‐administered digital cognitive assessment.

2.6.1. Anchoring hierarchy

Given that traditional neuropsychological tests are influenced by education, language, culture, and familiarity, we implemented a three‑level anchoring hierarchy to contextualize the strength of validation evidence. Level‑1 anchoring reflects biological confirmation using BBMs, CSF, or PET. Level‑2 anchoring reflects comprehensive clinical adjudication using multidomain neuropsychological assessment and clinician consensus. Level‑3 anchoring reflects proxy anchors such as brief screens. Because therapeutic eligibility and disease staging increasingly rely on biological confirmation, we defined Level‐1 biomarker anchoring as the highest evidentiary standard for screening tools intended to guide downstream AD diagnostic pathways. Validating SA‑DCAs solely against neuropsychological tests—many of which exhibit education, language, and cultural performance biases—risks propagating those biases into digital tools. Accordingly, our framework requires either Level‑1 biological anchoring or prespecified, statistically supported bridging that demonstrates concordance between Level‑2 and Level‑1 anchors in comparable populations; Level‑3 proxies alone are insufficient for clinical readiness in heterogeneous screening contexts. This anchoring hierarchy reflects a broader conceptual distinction between behavioral construct validation and biological validation.

2.6.2. Population representativeness

Given that SA‑DCAs are intended for widespread use in community and primary‑care settings, the framework emphasizes validation in unfiltered populations, reflecting the demographic, linguistic, educational, socioeconomic, and device‑access diversity of real‑world screening environments. Studies were evaluated for (1) sampling strategies supporting the right populations; (2) linguistic and cultural adaptation procedures, including standardized translation/back‑translation; (3) literacy and accessibility considerations; and (4) DIF analyses where multi‑language deployment was claimed. Lack of representativeness was treated as a major limitation for clinical generalizability.

2.6.3. Performance reporting

For performance evaluation, we extracted the following metrics: area under the receiver operator characteristic (ROC) AUC, sensitivity, specificity, accuracy, positive predictive value (PPV), negative predictive value (NPV), and Cohen's d. We considered AUC > 0.80, sensitivity > 75%, specificity > 60%, and Cohen's d ≥1.2 as indicators of strong discriminative performance, consistent with established benchmarks. ³⁶ , ³⁷ , ³⁸ We also assessed whether studies reported confusion matrices, detection rates, false‑positive rates, or stratified performance (e.g., by age, education, language, or device). For machine‑learning–derived tools, cross‑validation using independent training and test sets was considered essential to mitigate overfitting. Because real‑world screening requires estimating accuracy at plausible prevalence levels, we evaluated whether PPV/NPV calculations were adjusted accordingly. ³⁹

2.6.4. Clinical benchmarks and categorization

To ensure alignment with clinical decision‑making pathways, we also evaluated whether SA‐DCAs benchmarked their performance against Level‑1 or Level‑2 reference standards. Using the combined criteria described above, SA‑DCAs were categorized into four evidence levels: (1) Fully meets: tested in unfiltered populations, used appropriate comparisons (e.g., MCI vs cognitively unimpaired [CU], see Figure 1a) incorporated Level‑1 anchoring or valid bridging to Level‑1, and demonstrated robust performance (AUC > 0.80 or Cohen's d > 1.2), (2) Partially meets: met most criteria, but used pre‐filtered populations lacked biological biomarker anchoring, or reported lower accuracy, (3) Minimally meets: evaluated primarily in pre‐filtered cohorts substantially limited metrics reported, 4) Does not meet: did not evaluate the relevant population or lacked diagnostic performance data (e.g. a feasibility or usability studies). Tool categorization followed the PRISMA‑ScR flow shown in Figure 2.

FIGURE 2.

SA‐DCA landscape and literature review: PRISMA‐ScR flow diagram. Three step process (left, blue) showing the number SA‐DCAs identified to be able to detect MCI, then those that were able to self‐administer testing with sufficient validation studies. Peer reviewed studies were evaluated for all identified SA‐DCAs and reported study findings were reviewed. Studies were filtered using the framework's inclusion and exclusion criteria (accuracy was reported, comparison matched the context of use, and populations were “unfiltered”). Note: “unfiltered” meant outside a pre‐selected clinical trial cohort, see Figure 1 for the focus population comparisons used). PRISMA‐ScR, Preferred Reporting Items for Systematic Reviews and Meta‐analysis for Scoping Reviews; SA‐DCA, self‐administered digital cognitive assessment.

Final determinations were jointly made by three reviewers (B.M., V.M., M.D.), with discrepancies resolved by consensus. Because included studies varied in design and no head‑to‑head comparisons were performed, categorizations do not imply superiority among SA‑DCAs but instead reflect how completely each tool aligns with the framework's real‑world evidence ,   requirements.

2.7. Synthesis of results

The resulting framework was applied to all eligible studies to assess real‐world validation readiness. Data were summarized descriptively and tabulated. Quantitative metrics were synthesized narratively due to heterogeneity in study design and endpoints. Thematic synthesis was applied to highlight recurring methodological patterns and evidence gaps, emphasizing readiness for validation in primary care and community settings.

3. RESULTS

3.1. Framework development

We first reviewed existing frameworks to assess their applicability to SA‐DCAs intended for MCI screening in community and primary care settings (see Supplementary Table 2). Framework design was informed by three practical use cases: self‐reported SCD, informant concern, and clinician suspicion during a routine visit. The resulting framework provided a structured approach for evaluating SA‐DCAs intended for population‐level MCI screening which can be adapted for similar use cases. Importantly, the framework was constructed with the explicit assumption that tools intended for large‐scale screening must be evaluated against the demographic, linguistic, educational, and technological diversity of their intended deployment contexts. Frameworks such as the DiME V3 and V3+ models provided foundational principles for verification (ensuring the tool is built correctly), analytical validation (accuracy and reliability), and clinical validation (relevance to clinical outcomes), usability, and human factors ⁴⁰ , ⁴¹ but lacked use‐case–specific criteria for self‐administered cognitive screeners. To address this gap, we developed a tailored framework (Figure 1) emphasizing four core domains: (1) appropriate comparator groups (MCI vs. CU or cognitively normal [CN] participants); (2) population selection; (3) comprehensive reporting of diagnostic performance metrics; and (4) clinical anchoring using neuropsychological and biomarker‐based reference standards.

In assessing appropriate comparator group, the framework excluded groupings where MCI and dementia cases were combined given that this can overestimate test accuracy and introduce a pre‐selection bias—patients with mild dementia are easier to detect. ⁴² To be considered for implementation, validation cohorts should reflect diversity across age, sex, race, ethnicity, language, and socioeconomic status. Thus, a key component of a validation study is that the results can be generalizable to a screening population for which the screener will be used. Additionally, for the evaluation of performance, key metrics should include accuracy, sensitivity, specificity, PPV and NPV (Supplementary Figure 1). PPV should be adjusted based on estimated population prevalence. Assessing accuracy with multiplex markers should report cross‐validation using model training and test sets to avoid overfitting. Finally, to ensure translation into clinical workflows, a clinically anchored validation should include diagnostic performance measures that are benchmarked to clinical‐biological gold standards (e.g., neuropsychological evaluation and/or biomarker confirmation with BBMs, CSF biomarkers, and/or amyloid PET).

3.2. Identification and selection of eligible studies

Selection outcomes are summarized in the PRISMA‐ScR flow diagram (Figure 2). The multi‐source search (January 2012–March 2025) identified 79 self‐administered, remotely deployable SA‐DCAs. Following screening, 37 SA‐DCAs met eligibility for evidence review having more than one peer‐reviewed validation study (for detailed characteristics of SA‐DCAs, see Supplementary Table 3). A total of 161 peer‐reviewed published records were screened in full. After exclusion of studies lacking diagnostic performance metrics (n = 97), appropriate comparisons (MCI versus CN, n = 72), and using pre‐filtered clinical trial cohorts, nine studies remained. These nine studies evaluated four SA‐DCAs—the ki:elements (ki:e) Speech Biomarker–Cognition (SB‐C) battery, Cambridge Neuropsychological Test Automated Battery (CANTAB), My Cognition (MyCog Mobile), and Cogstate Brief Battery (CBB)—which together represented the current evidence base for these studies (Figure 3). ⁴³ , ⁴⁴ , ⁴⁵ , ⁴⁶ , ⁴⁷ , ⁴⁸ , ⁴⁹ , ⁵⁰ , ⁵¹

FIGURE 3.

SA‐DCA performance characteristics reported from selected studies: applying framework criteria. SA‐DCA were categorized by applying the criteria framework using the study designs and performance data reported. This illustration categories each SA‐DCA based on what is reported and whether it meets our framework criteria for a screening test to detect MCI in a real‐world population. Top black, yellow, red, and blue boxes correspond to color‐coded categories on the bottom row. Note: Given that these are not head‐to‐head studies, and the study designs differ, the categories are not meant to ascribe superiority of one SA‐DCA over the other. See Supplemental Materials for more information. SA‐DCA, self‐administered digital cognitive assessment.

3.3. Characteristics of included studies

The 161 studies collectively enrolled more than 12,000 participants across diverse settings including primary care, community‐based, and at‐home testing environments. Sample sizes ranged from less than 50 to over 6,000 participants. Most studies employed cross‐sectional designs; eight included longitudinal follow‐ups ranging from two weeks to five years. In those that had the appropriate comparisons, twenty‐eight studies were conducted in primary care and/or at‐home contexts, the rest involved specialists. Eighteen studies incorporated biomarker confirmation with amyloid‐β PET, while only three included correlations with BBM confirmatory tests. Most studies relied on established neuropsychological criteria (e.g., NIA–AA or Petersen definitions) for MCI classification.

Risk‐of‐bias assessment using QUADAS‐2 S indicated moderate‐to‐high risk primarily due to convenience sampling, limited blinding, and heterogeneous reference standards (see Supplementary Figure 2, Supplementary Table 4).

3.4. Application of the framework

Applying the predefined framework (Section 2.7), all nine studies were categorized according to the degree to which they met criteria for real‐world readiness (Figure 3). None of the SA‐DCAs fully met all framework criteria. MyCog Mobile, CANTAB, and ki:e SB‐C met most validation components, including use of unfiltered populations and accuracy and sensitivity/specificity reporting, and were thus classified as meeting framework criteria. Cogstate Brief Battery met core technical requirements but had lower performance and was therefore considered to partially meet framework criteria. Tools with incomplete diagnostic performance reporting or reliance on filtered populations were excluded prior to framework application.

Only one study provided a complete confusion matrix: most reported partial metrics. Moreover, no study reported PPV or NPV adjusted to the expected prevalence of MCI in community screening populations or provided stratified accuracy metrics across demographic or linguistic subgroups, limiting the interpretability and transportability of reported diagnostic performance. Several promising SA‐DCAs were excluded but may become eligible as additional evidence emerges.

3.5. Summary of diagnostic performance and validation levels

Across all nine eligible studies, none met full real‐world representativeness criteria as defined by the framework. Reported diagnostic performance for MCI classification relative to CU/CN comparators ranged from AUC = 0.68–0.86 (Table 2a,b). CANTAB demonstrated the highest AUC (0.86) when combining multiple subtests, although performance decreased to 0.68 in unsupervised, at‐home settings, and most cohorts were drawn from registries or higher resource environments. MyCog Mobile recruited from primary‑care settings but relied on a Level‑3 proxy cognitive anchor and was validated only in English, limiting generalizability to diverse linguistic populations reporting an AUC of 0.83 (sensitivity = 0.76, specificity = 0.88). ki:e SB‐C, a speech‐based tool, demonstrated moderate performance (AUCs of 0.73–0.81) but lacked biomarker anchoring and validation across global or multilingual cohorts with sensitivity reported (0.62–0.79). Cogstate Brief Battery studies reported AUCs of 0.70–0.75 with high specificity (> 0.85) but variable sensitivity (0.25–0.45) across specialist and research cohorts, with inconsistent reporting of language versions, education levels, and device modalities. To clarify what representativeness was possible within each study's design, we provide a structured summary of recruitment settings, language availability, education distributions, and device modalities (see Supplementary Table 5).

According to the predefined interpretive thresholds (AUC ≥ 0.80 or Cohen's d ≥ 1.2 for strong discriminative validity), only MyCog Mobile and CANTAB achieved high performance in unfiltered populations. Overall, three of four tools met most framework criteria with substantial methodological variability across studies and incomplete reporting.

3.6. Synthesis of evidence gaps

Thematic synthesis revealed several recurring gaps across studies. Samples were frequently trial‐enriched or demographically narrow, limiting generalizability. Reporting completeness was inconsistent. Reference standards varied widely, ranging from comprehensive clinical evaluations to brief cognitive screeners, limiting comparability across tools. Longitudinal validation was rare, with only a few studies assessing test–retest reliability or cognitive trajectories. Finally, biomarker anchoring was seldom included, and no studies incorporated BBMs, hindering integration with disease‐specific diagnostic workflows.

Together, these findings underscore that while SA‐DCAs demonstrate technical feasibility and moderate diagnostic accuracy, the evidence supporting their implementation remains limited. Expanded validation in unfiltered, diverse primary/community care populations, adherence to standardized reporting, and alignment with biological confirmation pathways are critical next steps for advancing clinical utility.

4. DISCUSSION

Early detection of MCI remains a pressing global healthcare challenge. Without appropriate tools, clinicians may overlook early‐stage impairment in patients presenting with SCD, delaying diagnosis, and narrowing the window for intervention. ⁵¹ , ⁵² By 2030, one in six people globally will be aged ≥ 60 years, underscoring the urgency for scalable screening approaches. ⁵² SA‐DCAs have the potential to expand access to cognitive screening through their scalability, remote deployment, ease‐of‐use, and ability to support longitudinal testing. They offer an accessible pathway for individuals who may not otherwise seek evaluation, including those in rural, underserved, or resource‐limited settings.

The findings of this review highlight both the promise and the current evidentiary limitations of SA‐DCAs for real‐world MCI screening. Notably, none of the evaluated SA‐DCAs fully satisfied all framework domains simultaneously, highlighting the current gap between available digital tools and the necessary evidence required for real‐world screening readiness. It also reveals a fragmented and inconsistent validation landscape. Namely, many studies relied on pre‐selected cohorts drawn from clinical trials or memory clinics, which can overestimate diagnostic performance and limit generalizability across real‐world screening populations. Some tools performed better in pre‐filtered cohorts than in real‐world populations, suggesting potential overfitting.

A central implication of this review is that validation of SA‑DCAs must clearly distinguish between behavioral construct validation and biological disease confirmation. Traditional neuropsychological test results are influenced by education, language, culture, and test familiarity. When SA‑DCAs are validated exclusively against these instruments, their biases may be propagated. This concern is amplified in global and culturally diverse deployment contexts. Our framework prioritized biological anchoring using BBMs, CSF, and/or PET imaging, or prespecified bridging analyses demonstrating concordance between clinical adjudication and biological confirmation. These anchoring strategies are not interchangeable and are essential for screening tools intended to inform downstream diagnostic pathways.

Applying a global lens further underscores the limitations of the current evidence base. None of the evaluated SA‑DCAs demonstrated representativeness across language, education, socioeconomic status, and care settings. In several cases, these gaps reflected structural constraints of available datasets such as specialist‑enriched cohorts or English‑only language rather than intrinsic tool limitations. The absence of culturally adapted validation, education‑stratified normative data, and DIF analyses limits confidence in generalizability, particularly in low‑ and middle‑income settings. Future validation should therefore incorporate multilingual deployment and culturally adapted norms in populations where screening would be performed. Reporting practices also varied substantially. Many studies failed to report full diagnostic metrics or adjust for prevalence. Frameworks such as the Standards for Reporting of Diagnostic Accuracy Studies (STARD) provide transparent guidelines for study design, blinding, and reference standards, and statistical reporting of accuracy metrics. ⁵³ Applying these principles to SA‐DCA validation studies would improve scientific reporting rigor and enable meaningful cross‐tool comparisons.

Recent recommendations from the Global CEO Initiative (CEOi) provide important guidance on acceptable performance standards for clinic‑based DCAs, emphasizing validation standards, representative normative data, and clinically interpretable outputs. ⁵⁴ However, these recommendations do not yet address requirements unique to unsupervised self‑administered, remotely deployed screening tools, which the CEOi has identified as a priority for its forthcoming remote‑assessment workstream. Our framework extends CEOi guidance by incorporating appropriate population selection, performance reporting, and structured clinical and biological anchoring. Applying this framework allows DCA developers to design validation studies aligned with the tool's intended use and to generate the level of evidence required for real‑world deployment. These distinctions are particularly relevant for SA‐DCAs deployed in remote or unsupervised contexts, where issues of population representativeness, device variability, and biological anchoring become especially important.

BBM‑anchored diagnostic approaches provide a model for staged evidence development in AD, beginning with defined acceptable performance criteria, ⁵⁵ progressing to implementation recommendations ⁵⁶ and culminating in clinical practice guidelines. ⁵⁷ In contrast, DCAs measure performance on cognitive tasks, which is influenced by behavioral, contextual, and human factors (e.g., fatigue, stress, mood, digital literacy, and cultural context), introducing variability that fundamentally differs from biological biomarkers. ⁴¹ These differences further underscore the importance of a validation framework tailored to the unique characteristics of DCAs.

Given that cognitive changes over time are often smaller than natural intra‐individual natural variability, trend‐based longitudinal monitoring may offer greater diagnostic value. ⁵⁸ DCAs can also generate a digital cognitive phenotype, a personalized profile of cognitive function over time that may serve as a proxy for brain health analogous to blood pressure monitoring for cardiovascular risk. Establishing a cognitive baseline and tracking longitudinal changes allows for detection of deviations and may differentiate disease‐related decline from other transient factors such as medications or sleep disruption. This approach may also increase sensitivity to subtle cognitive changes, by smoothing day‐to‐day variability, yielding a more stable representation of cognitive function. Yet few studies incorporated such analyses. To address this, we propose two complementary strategies (Figure 4):

1. Longitudinal Cognitive Monitoring (LCM): Repeated assessments over time can reduce random error, mitigate practice effects, and enhance sensitivity to meaningful change. Modeling trajectories may enable earlier detection of progressive decline and distinction between neurodegenerative and stable conditions. AI‐enabled trend analysis could further improve predictive accuracy. ⁵⁹ , ⁶⁰

2. Cognitive Staging Using Composite Assessments (CSUCA): Domain‐specific composite scores, modeled after validated batteries may improve sensitivity and specificity. A two‐tiered approach (high‐sensitivity screening followed by confirmatory composite staging) could streamline triage and optimize resource utilization.

FIGURE 4.

Digital cognitive phenotyping methods for measuring meaningful changes using SA‐DCAs within the diagnostic and care pathways. (A) Digital cognitive phenotyping using Longitudinal Cognitive Monitoring (LCM) demonstrating value of multiple measures over time in detecting cognitive decline. Shown is a hypothetical example compared to changes over time from normal aging versus those that develop CI. (B) Cognitive Staging Using Composite Assessments (CSUCA) such that combining tests with different composite types and scores can improve sensitivity and assess overall cognitive staging to assess disease. MoCA, Montreal Cognitive Assessment; Mod, Moderate; No/Lo, No or Low Risk; PACC5, Preclinical Alzheimer Clinical Composite‐5. (C) Proposed care pathway: DCA and digital cognitive phenotyping in MCI detection linked into clinical workflows. Role of SA‐DCAs in screening for MCI in those at‐risk who have SCD and need to be tested: shifting the care pathways in the initial testing towards a patient‐enabled screening (or physician‐directed) modality. SA‐DCA results prompt those being screened to seek care with their primary care provider to further assess them for MCI. When other common causes (e.g., reversible etiologies) are ruled out by a clinician and if MCI is confirmed AD is suspected, a DCA result could help with clinical decisioning (e.g., order an AD laboratory test to “rule‐in”, or “rule‐out”, AD). Validations in real‐world settings are imperative to demonstrate that the SA‐DCA solutions are easy to use, results are interpretable by those self‐testing and it is clear when they should seek further care with their provider. Providers then should be able to interpret the results, which should inform them of their next best action. AD, Alzheimer's disease; CI, cognitive impairment; SA‐DCA, self‐administered digital cognitive assessment; MCI, mild cognitive impairment; MRI, magnetic resonance imaging; MoCA, Montreal Cognitive Assessment; PACCs, Preclinical Alzheimer's Cognitive Composite; RW, real‐world.

Importantly, SA‑DCAs can function as upstream screening tools to identify individuals who may benefit from further clinical and/or biomarker evaluation, not to establish a diagnosis or replace neuropsychological assessment. Results should be interpreted alongside patient‑reported outcomes, informant observations, and clinical context. Validated SA‑DCAs may improve referral efficiency by increasing pre‑test probability and supporting more targeted use of confirmatory biomarker testing and specialist assessments. Generating real‑world evidence including integration with plasma biomarkers will be essential to strengthening clinician confidence and clarifying how SA‑DCAs can improve referrals pathways. Incorporating longitudinal follow‑up into validation studies will help establish the prognostic relevance of SA‑DCAs alongside diagnostic utility. For clinical adoption, SA‐DCAs need to yield results that are compatible with routine clinical workflows. Prevalence‐adjusted confusion matrices and visual summaries could assist clinicians in translating results into actionable next steps. Ultimately, successful tools must demonstrate technical and clinical utility with results that are interpretable, actionable, and trusted by patients and primary and/or community care providers.

The future of digital diagnostics lies in multimodal integration. In addition to measuring performance on structured cognitive tasks, digital approaches can be designed to capture behavioral and decisional changes related to underlying cognitive and neurobiological dysfunction, such as mild behavioral impairment (MBI) or mild decisional impairment (MDI). These constitute distinct measurement modalities that index related aspects of cognitive dysfunction and may serve as more sensitive manifestations of early cognitive change, particularly in the earliest stages of AD. ⁶¹ When combined with passive digital phenotyping (e.g., navigation patterns, hesitation, and error correction), such approaches may further enhance screening sensitivity, although many remain in proof‑of‑concept stages. Detecting subtle cognitive, behavioral, and decisional changes may help identify preclinical AD (asymptomatic individuals with biological evidence of an amyloid‐β pathology) who may be at increased risk for progression. Emerging technologies incorporating novel digital biomarker measures may further improve early detection capabilities. Digital measures including eye tracking, ²⁶ , ⁶² saccadic movements, ⁶³ , ⁶⁴ speech and voice analytics, ⁶⁵ gait and movement sensors, ⁶⁶ , ⁶⁷ and sleep monitoring ⁶⁸ combined with advanced psychometric methods such as computerized adaptive testing and multidimensional item response theory, may enable high‐resolution modeling of disease progression. ³⁰ , ⁶⁹

This review has limitations. Conference abstracts and posters were excluded due to insufficient methodological detail. Excluding gray literature may have omitted emerging tools while including it risks unverifiable data. Heterogeneity in reference standards from clinical judgment to biomarker confirmation made comparability across studies impractical. Although we did not define performance cutoffs in this study, we reference previously established thresholds to guide interpretation. No head‐to‐head comparisons were available, limiting comparisons across tools. The review mapped available evidence rather than synthesizing effect sizes; thus, GRADE criteria were not applied. Risk‐of‐bias assessments (QUADAS‐2) are summarized (Supplementary Figure 2, Supplementary Table 5). Future research should address usability, digital literacy, language adaptation, ecological validity, socioeconomic and data governance and privacy. Several observed limitations in representativeness reflect constraints independent of the tools themselves, further reinforcing the need for purpose‑built, unfiltered, and culturally diverse validation studies.

This framework addresses a central methodological gap in the SA‐DCA literature: the absence of explicit criteria linking validation design to the intended screening population. As screening for early AD increasingly intersects with clinical–biological staging and therapeutic eligibility, validation strategies must account for both behavioral performance and underlying disease biology. Transparent anchoring hierarchies, prevalence‐adjusted performance reporting, and representative validation will therefore be essential to ensure SA‐DCAs are evaluated according to their intended clinical role. By establishing evidentiary requirements aligned with clinical use, this framework provides a structured approach for evaluating SA‐DCA screening tools intended for real‐world deployment. In doing so, it supports earlier identification of cognitive impairment and a timely clinical evaluation and biological confirmatory testing necessary for an accurate AD diagnosis. Future validation studies conducted in representative real‐world populations and integrated with biomarker‐based diagnostic pathways will be critical to translating SA‐DCAs from promising research tools into clinically actionable screening solutions.

AUTHOR CONTRIBUTIONS

All authors (Harald Hampel, Yosuke Nakamura, Joanne Bell, Suzanne Hendrix, Miron Derchansky, Bruno Montezano, Luke E. Stoeckel, Stefan Teipel, Deanna R. Willis, Jeffrey L. Cummings, and Bryan R. Cobb) were involved in the design, results, interpretations, and discussion components of this study and all authors have read and approved the manuscript. Bryan R. Cobb, Vera Maljkovic, Miron Derchansky and Bruno Montezano, contributed to the data analysis and abstractions. Bryan R. Cobb completed most of the writing. All authors participated in manuscript development, design and reviewed the manuscript. This manuscript was developed in accordance with Good Publication Practice 2022 (GPP2022) guidelines and followed ICMJE standards.

CONFLICT OF INTEREST STATEMENT

Luke E. Stoeckel is an employee of the NIH. The content is solely the responsibility of the authors and does not represent the official views of the NIH or the U.S. Government. Jeffrey L. Cummings has provided consultation to Acadia, Acumen, ALZpath, Annovis, Artery, Axsome, Biogen), Biohaven, Bristol‐Myers Squib, Cervomed, Eisai, Fosun, GAP Foundation, Green Valley, Hummingbird Diagnostics. IGC, Janssen, Kinoxis, Lighthouse, Lilly, Lundbeck, LSP/eqt, Merck, MoCA Cognition, Novo Nordisk, NSC Therapeutics, Optoceutics, Otsuka, Praxis, ReMYND, Roche, Scottish Brain Sciences, Signant Health, Simcere, sinaptica, T‐Neuro, TrueBinding, and Vaxxinity pharmaceutical, assessment, and investment companies. Dr. Cummings is co‐founder CNS Innovations and Mangrove Therapeutics. Suzanne Hendrix is owner and employee of Pentara Corporation, a company that consults with dozens of companies in the neurodegenerative space. Miron Derchansky is owner and employee of Blackletter Consulting Incorporated, an advisory firm that consults with multiple pharmaceutical companies in the digital innovation space. Deanna R. Willis has no competing interests. Other disclosures: Harald Hampel is an employee of Bristol Myers Squibb (and was an employee of Eisai, Inc. at the time the manuscript was being prepared) and served as Reviewing Editor and previously as Senior Associate Editor for the journal Alzheimer's & Dementia, the journal of the Alzheimer's Association. He is member of the scientific advisory boards of Sinaptica Therapeutics, Qynapse, and PAM Theragnostics. He is inventor of 121 patents and has received no royalties: In Vitro Multiparameter Determination Method for The Diagnosis and Early Diagnosis of Neurodegenerative Disorders Patent Number: 8916388. In Vitro Procedure for Diagnosis and Early Diagnosis of Neurodegenerative Diseases Patent Number: 8298784 Neurodegenerative Markers for Psychiatric Conditions Publication Number: 20120196300, In Vitro Multiparameter Determination Method for The Diagnosis and Early Diagnosis of Neurodegenerative Disorders Publication Number: 20100062463, In Vitro Method for The Diagnosis and Early Diagnosis of Neurodegenerative Disorders Publication Number: 20100035286, In Vitro Procedure for Diagnosis and Early Diagnosis of Neurodegenerative Diseases Publication Number: 20090263822, In Vitro Method for The Diagnosis of Neurodegenerative Diseases Patent Number: 7547553, CSF Diagnostic in Vitro Method for Diagnosis of Dementias and Neuroinflammatory Diseases Publication Number: 20080206797, In Vitro Method for The Diagnosis of Neurodegenerative Diseases Publication Number: 20080199966, Neurodegenerative Markers for Psychiatric Conditions Publication Number: 20080131921​, Method for diagnosis of dementias and neuroinflammatory diseases based on an increased level of procalcitonin in cerebrospinal fluid: Publication number: United States Patent 10921330. CSF diagnostic in vitro method for diagnosis of dementias and neuroinflammatory diseases. Publication number: 20160097782. Jeffrey L. Cummings has provided consultation to Acadia, Acumen, ALZpath, AnnovisBio, Artery, Axsome, Biogen, Biohaven, Bristol‐Myers Squib, Eisai, Fosun, GAP Foundation, Hummingbird Diagnostics. IGC, Janssen, Julius Clinical, Kinoxis, Lighthouse, Lilly, Lundbeck, LSP/eqt, Merck, MoCA Cognition, Novo Nordisk, NSC Therapeutics, Optoceutics, Otsuka, Praxis, ReMYND, Roche, Scottish Brain Sciences, Signant Health, Simcere, sinaptica, T‐Neuro, TrueBinding, and Vaxxinity pharmaceutical, assessment, and investment companies. Dr. Cummings is co‐founder of CNS Innovations and Mangrove Therapeutics. JLC is supported by NIGMS grant P20GM109025; NIA R35AG71476; NINDS RO1NS139383; Alzheimer's Disease Drug Discovery Foundation (ADDF); Ted and Maria Quirk Endowment; Joy Chambers‐Grundy Endowment. Author disclosures are available in the Supporting Information.

Supporting information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information