A multimodal digital NeuroMarker platform with composite and domain-level validity for identification of mild cognitive impairment: a cross-sectional study

Summary

Background

Mild cognitive impairment (MCI) is a clinically heterogeneous syndrome that can represent an early manifestation of neurodegenerative disease. Diagnosis relies on assessments that are difficult to scale, leaving many individuals unidentified.

Methods

In this cross-sectional study conducted in the United States, Australia, and Spain between May 2022 and October 2024, 638 adults aged ≥ 50 years were classified as cognitively unimpaired (CU; n = 441) or MCI (n = 197) using Petersen/NIA-AA criteria. Participants completed a self-administered, tablet-based assessment (Altoida NeuroMarker Platform) integrating motor, speech, and augmented reality tasks. An XGBoost classifier was trained to distinguish CU from MCI. Diagnostic performance, test-retest reliability, and convergent validity were evaluated.

Findings

Median assessment time was 13.8 min. Participants reported comfort with digital technology (67%), willingness to engage in cognitive screening (90%), and positive usability. Altoida’s NeuroMarker differentiated MCI with ROC-AUC 0.90, 95% CI 0.88–0.92 and comparable performance in an independent cohort (ROC-AUC 0.87, 95% CI 0.80–0.93; n = 75). Cognitive composite and domain scores were lower in MCI than in CU. Classifier agreement across repeated assessments was κ = 0.60, and reliability of composite and domain scores was moderate-to-good (ICC 0.55–0.80). The composite score correlated with the Mini-Mental State Examination (MMSE) (r_s = 0.40, p < 0.001) and Repeatable Battery for the Assessment of Neuropsychological Status (RBANS) Total Index (r_s = 0.58, p < 0.001). The platform also showed convergent validity of cognitive domains (|r_s| = 0.27–0.50, p < 0.001).

Interpretation

A brief, self-administered multimodal assessment identified clinically adjudicated MCI and demonstrated convergent validity, test-retest reliability, and consistent performance in an independent cohort. Prospective studies should evaluate its implementation in real-world workflows.

Funding

Alzheimer’s Drug Discovery Foundation; Altoida Inc.

Research in context.

Evidence before this study

We searched PubMed for articles indexed up to March 1, 2026, using the terms (“mild cognitive impairment” OR “MCI”) AND “digital biomarkers” AND “machine learning” AND “early detection”. No language restrictions were applied. The search yielded 29 publications investigating digital approaches for early identification of cognitive impairment, published between 2020 and 2026. Most studies examined specific digital biomarkers, including speech, gait, sensor-derived behavioral signals, and virtual or augmented reality tasks. Some studies applied machine-learning models to identify MCI, with reported ROC-AUC values ranging from approximately 0.70 to 0.90. However, much of the available evidence derives from single-site cohorts or single modalities (e.g., speech- or gait-only). Few studies have evaluated self-administered multimodal tools in large, geographically diverse cohorts, or reported multiple dimensions of validation, including independent performance, calibration, reliability, and construct validity.

Added value of this study

This study provides evidence that a multimodal, self-administered digital assessment can accurately identify MCI and characterize cognition through interpretable composite and domain-level cognitive measures validated against established neuropsychological assessments, within a format compatible with scalable deployment.

Implications of all the available evidence

Our findings suggest that multimodal, self-administered digital cognitive assessments may help address barriers to standardized cognitive evaluation in frontline care. If validated in prospective real-world studies, such a tool could support more efficient clinical decision support, facilitate earlier identification of cognitive impairment, and improve pathways to biomarker testing and timely treatment. Future research should evaluate implementation across diverse care settings, including integration with biomarker testing and longitudinal monitoring of cognitive change.

Introduction

Mild cognitive impairment (MCI) is characterized by cognitive decline beyond what is expected for normal aging, with preservation of functional independence. It affects approximately 15–22% of individuals over 65 years.¹ Medicare-based estimates suggest that up to 90% of MCI cases go undiagnosed, reflecting broader system-level challenges.^2,3 Alzheimer’s disease (AD) is among the most common underlying pathologies of MCI.⁴ The heterogeneity of MCI etiologies underscores the need for scalable tools that support early identification and etiologic characterization.

A major barrier to timely MCI identification is the absence of a single definitive test or universal diagnostic criteria, leading to variability in clinical practice.⁵ Diagnosis relies on lengthy cognitive batteries requiring trained personnel for administration, limiting scalability.⁶ As a result, cognitive evaluation is rarely feasible in frontline settings, where brief consultation times, competing clinical priorities, limited access to neuropsychological testing, insufficient training, and the normalization of memory decline as part of aging further contribute to delayed diagnosis.⁷ Consequently, there remains an unmet need for scalable approaches that enable cognitive assessment with minimal personnel requirements while preserving clinical interpretability.

Digital cognitive assessments (DCAs) have emerged as scalable alternatives to traditional testing, with approaches based on speech, eye tracking, digital pen input, smartphone tasks, and wearable sensors.⁸ These technologies have demonstrated promise for improving access to cognitive assessment. However, DCAs represent a heterogeneous field, with substantial differences in intended use, cognitive constructs, underlying technologies, validation strategies, and clinical maturity. The Altoida NeuroMarker Platform is a multimodal digital platform designed to capture complementary aspects of cognitive and functional performance within a brief, self-administered assessment (Fig. 1). It captures multimodal features from motor, augmented reality (AR), and speech tasks, which facilitate training of machine-learning (ML) models. Designed for self-administration with minimal staff involvement, the Platform is intended to augment, rather than replace, a clinical evaluation by providing a brief, objective assessment of whether a participant’s performance is consistent with MCI.

Fig. 1.

Representative screenshots from the Altoida cognitive assessment illustrating motoric, augmented reality (AR), and speech-based tasks. Motoric tasks involve tracing and tapping activities designed to capture fine motor performance, processing speed, coordination, and executive function. AR tasks include “Back In Time” and “Day Out,” which require participants to place, locate, and recall virtual objects within their physical environment while completing goal-directed activities. These tasks engage learning and memory, spatial navigation, planning, and task sequencing. Speech tasks involve verbal description and recall of visual stimuli. The assessment is self-administered on an iOS tablet and produces multimodal digital features that reflect cognitive and functional performance across multiple cognitive domains. Assessment outputs include an MCI classification, a global composite score, and scores across five cognitive domains: learning and memory, complex attention, executive function, perceptual-motor function, and language.

Early studies using the Altoida NeuroMarker Platform demonstrated feasibility for predicting progression from MCI to AD⁹ and detecting individual-level change.¹⁰ Building on this work, subsequent studies showed that the Platform could distinguish prodromal AD, while also demonstrating greater sensitivity to preclinical AD compared with established assessments.¹¹ Since that earlier work, the Platform has evolved, incorporating a larger and more diverse training dataset and additional modalities, including speech-derived features. More broadly, the clinical landscape has also evolved. Neurology has entered the biomarker era, but cognitive assessment still relies largely on clinician-administered pen-and-paper instruments developed in the previous century. DCAs, therefore, represent an opportunity to modernize cognitive evaluation and support more efficient allocation of specialist expertise. The approval of amyloid-lowering therapies has further increased the importance of timely MCI identification, placing greater emphasis on accurate and scalable detection in frontline healthcare, where current pathways remain poorly equipped to respond.¹²

To support this evolving clinical role, rigorous evaluation of the diagnostic performance and clinical validity of DCAs is essential. Here, we evaluated the performance, reliability, and clinical validity of the Altoida NeuroMarker Platform for MCI identification. Beyond MCI classification, the Platform also generates a global composite and domain-level cognitive scores to support interpretable characterization of cognitive function. We also evaluated the reliability and convergent validity of these measures against established neuropsychological assessments.

Methods

This study followed the STROBE reporting guidelines for cross-sectional studies.

Study participants

The ML training dataset included 638 participants classified as cognitively unimpaired (CU) or with MCI from the United States, Australia, and Spain, aged ≥ 50 years and of both sexes, recruited and assessed between May 2022 and October 2024. An additional 75 participants (115 assessments) from two independent cohorts were used for external validation; 50 participants recruited in Chicago, Illinois, with available plasma pTau217 measurements (LucentAD), and 25 participants recruited in Barcelona, Spain (Supplementary Table S1).

Participants were prospectively recruited from the community, research cohorts, and memory clinics to capture the heterogeneity of individuals undergoing cognitive evaluation across different clinical pathways.13, 14, 15 Participants completed the Altoida NeuroMarker during a study visit. Cognitive status was established either through clinical evaluation performed during the Altoida study visit or through prior clinical characterization within established cohorts, provided that the cognitive classification had been established no more than six months before Altoida administration. Clinical characterization included demographics, medical history, current medications, general cognitive screening with the Mini-Mental State Examination (MMSE), and neuropsychological assessment. Clinical classification was therefore based on an integrated clinical evaluation rather than cognitive test scores alone. While specific tests varied across cohorts, all evaluations adhered to recognized clinical standards, ensuring consistent diagnostic rigor. Additional cohort-specific recruitment procedures are described in the Supplementary Methods.

Individuals presenting with MCI were required to meet clinical criteria for MCI based on Petersen/2011 NIA-AA guidelines.⁶ MCI inclusion was not restricted by cognitive subtype or presumed etiology. Participants classified as CU did not meet study criteria for MCI based on the overall clinical and neuropsychological assessment. This category included individuals with no evidence of objective cognitive impairment, as well as some individuals presenting with cognitive concerns or isolated findings on individual cognitive measures whose overall clinical picture was not consistent with MCI.

Biological sex, race, and ethnicity were self-reported. Ethnicity was collected as Hispanic or Latino or Not Hispanic or Latino. Race was collected as White, Black or African American, American Indian or Alaska Native, Asian, Native Hawaiian or Other Pacific Islander.

Altoida NeuroMarker platform

Participants were prospectively invited to complete either one or two Altoida assessments at the study sites using iOS tablets. Additional details are in the Supplementary Methods.

Participant survey

A subset of participants (n = 382) completed an online survey assessing their attitudes toward digital cognitive testing, including impressions of the Altoida NeuroMarker (Supplementary Table S2). Usability observations were also recorded for a subset of participants (n = 438).

Ethics approval

Ethical approval was obtained from a central Institutional Review Board in the US (Advarra; Pro00076590) and from the local Research Ethics Committees of each institution in Spain (β-AARC_BBRC2021; HCB/2022/1173; EC/23/225/7331) and Australia (HREC 028/06). All participants provided written consent before participation. The study was designed and carried out in accordance with the ethical principles outlined in the Declaration of Helsinki.

ML model training

Using data from the Altoida NeuroMarker, we trained an XGBoost binary classification model to identify MCI. The model was trained using five-fold cross-validation with participant-level grouping, with each fold serving as an internal holdout set for performance estimation. An optimal classification threshold was first determined independently within each fold to maximize sensitivity and specificity. The mean of these fold-specific thresholds (0.315) was then applied in a second round of cross-validation from which the final performance metrics were derived. A total of 638 participants (970 assessments) were included in the ML training dataset (Table 1). Data exclusion decisions are presented in Supplementary Fig. S1. The MCI classifier used 22 features reflecting performance on motoric, AR, and speech tasks, and included four demographic features: age, sex, education level, and dominant hand. Model hyperparameters were tuned using a Bayesian optimization method. Additional details are in the Supplementary Methods. The final model was then evaluated in the external validation cohort described above.

Table 1.

Demographic characteristics of the study population.

	CU	MCI	Total	p-value
Sample size, N (%)
	441 (69.1)	197 (31.9)	638 (100)
Country, N (%)				p < 0.001b
Australia	55 (12.5)	19 (9.7)	74 (11.6)
Spain	242 (54.9)	70 (35.5)	312 (48.9)
USA	144 (32.6)	108 (54.8)	252 (39.5)
Biological Sex, N (%)				p = 0.002a
Female	272 (61.7)	95 (48.2)	367 (57.5)
Age				p < 0.001b
Median (IQR)	67.0 (61.0, 72.0)	71.0 (65.0, 76.0)	68.0 (62.0–73.0)
Education, N (%)				p < 0.001a
Less than High School	38 (8.6)	33 (16.8)	71 (11.0)
High School Diploma	123 (27.9)	78 (39.6)	203 (31.7)
College/University	280 (63.5)	86 (43.6)	366 (57.3)
Ethnicity, N (%)				p < 0.001a
Hispanic or Latino	27 (6.1)	43 (21.8)	70 (11.0)
Not Hispanic or Latino	413 (93.7)	154 (78.2)	567 (88.8)
Not Available	1 (0.2)	0 (0)	1 (0.2)
Race, N (%)				p = 0.02a
White	388 (88.0)	161 (81.7)	549 (86.1)
Black or African American	29 (6.6)	28 (14.2)	57 (8.9)
Other	21 (4.8)	7 (3.6)	28 (4.4)
Not Available	3 (0.6)	1 (0.5)	4 (0.6)
MMSE Median (IQR)	29 (28–30)	27 (25–28)	29 (27–30)	p < 0.001b
Number of Altoida Assessments, N (%)				p = 0.30a
1 Assessment	218 (49.4)	88 (44.7)	306 (48.0)
2 Assessments	223 (50.6)	109 (55.3)	332 (52.0)

CU = cognitively unimpaired. IQR = interquartile range. MCI = mild cognitive impairment. MMSE = Mini Mental State Examination.

^aChi-Squared.

^bMann Whitney U test.

Additionally, a logistic regression model using only the four demographic variables was developed as a comparator. When restricted to these demographic variables, logistic regression achieved a higher ROC-AUC than an XGBoost model trained using the same demographic variables and was therefore selected as the comparator to the Altoida MCI NeuroMarker.

Altoida cognitive domains and composite score

In addition to MCI classification, the Altoida NeuroMarker generates a composite score and domain-level scores organized according to DSM-5 cognitive domains (perceptual-motor function, learning and memory, complex attention, executive function, and language). Domain scores summarize patterns of performance derived from multiple digital task features that predominantly reflect the corresponding cognitive construct. By design, individual features may contribute to more than one domain score because many digital tasks engage multiple cognitive processes simultaneously, and the underlying cognitive processes are not fully separable. Consequently, domain scores are not intended to be orthogonal. Assignment of features to cognitive subdomains was informed by the cognitive neuroscience literature, factor analysis, and expert review. Scores are expressed as age- and sex-adjusted percentiles relative to CU participants (Supplementary Table S3).

Statistical analysis

Statistical analyses were conducted using Python (v3.12) and R (v4.4). Continuous variables were compared using the Mann–Whitney U test and categorical variables using the χ² test. Unless otherwise indicated, continuous variables are summarized as medians with interquartile ranges (IQR), and categorical variables are presented as counts and percentages. Survey responses were summarized descriptively using frequencies or percentages. Associations between participant characteristics, self-reported digital comfort, and perceived tutorial helpfulness were evaluated using ordinal logistic regression. Calibration of the MCI classifier was assessed in the independent validation cohort using the Brier score and calibration plots. The observed Brier score was compared with that of a prevalence-based null model using the Wilcoxon signed-rank test. Group differences in composite and cognitive domain scores were assessed using t-tests or Mann–Whitney U tests as appropriate, with effect sizes (Cohen’s d) reported. Subdomain analyses were exploratory (Supplementary Fig. S2). All tests are reported as two-tailed, and significance was set at p < 0.05.

For the MCI classification outcomes, test-retest reliability was measured using Cohen’s Kappa (κ). Classification agreement was interpreted based on established κ ranges.¹⁶ This analysis was conducted using data from the validation cohort, which included 40 participants who completed two Altoida assessments (26 CU and 14 MCI). For the composite and domain scores, test-retest reliability was measured using the intraclass correlation coefficient (ICC). This analysis included participants who completed repeat assessments between two and six weeks apart (n = 220). Interpretation of ICC values followed established criteria.¹⁷

Convergent validity was evaluated using Spearman correlations between the composite score and measures of global cognition (MMSE and RBANS Total Index), and between domain-level scores and corresponding neuropsychological tests assessing similar constructs. Exploratory correlations are reported in the Supplementary Results. Adjustment for multiple comparisons was performed using the Bonferroni correction. To further characterize construct validity, an additional analysis compared the association of Altoida cognitive domains with a corresponding neuropsychological measure and a selected non-corresponding comparator (Supplementary Materials).

Role of the funding source

The funding organizations had no role in data analysis, interpretation of the results, or the decision to submit the manuscript for publication, except that Altoida-affiliated authors participated in the study design, data analysis and interpretation, and preparation of the manuscript in their capacity as authors. No author received payment from a pharmaceutical company or other agency for writing this manuscript.

No author was precluded from accessing data relevant to their contributions. The corresponding author had full access to the study data and had final responsibility for the decision to submit for publication.

Results

Participant characteristics

Participants were grouped according to their clinical status, as CU (n = 441) or MCI (n = 197) (Table 1). The distribution of participants did not significantly differ between recruitment countries (p > 0.05). There was a greater proportion of females in the CU group (61.7% vs. 48.2%; p < 0.01). CU participants were younger than those with MCI (median age 67.0 vs. 71.0 years; p < 0.001), differed in education level (p < 0.001), and showed higher median MMSE scores (29.0 vs. 27.0 points; p < 0.001). Age, biological sex, education level, and handedness were included as demographic features in the ML model.

Compared with CU participants, the MCI group included a higher proportion of Hispanic or Latino participants (21.8% vs. 6.1%; p < 0.001). Racial composition was comparable across groups (p > 0.05), predominantly White (86%), with Black or African American participants comprising the second largest racial group (8.9%). Among the participants who completed two assessments, only three out of 321 transitioned from CU to MCI at the follow-up visit.

Acceptability and user experience

The median time to complete the Altoida assessment was 13.8 min (IQR 11.9–16.1). Participants with MCI required more time than CU participants, with a median completion time of 15.9 min (IQR 13.5–18.7) vs. 12.9 min (IQR 11.6–14.6), respectively (p < 0.001).

Survey respondents indicated high digital comfort, with 67% reporting feeling comfortable using digital devices, 22% reporting neutrality, 11% indicating discomfort. Increasing age was associated with lower digital comfort (p < 0.01), whereas cognitive status, biological sex, and education level were not significant predictors. Many participants (66%) also reported experiencing frustration or worry about their memory in the past month, and most (90%) indicated they would be open to completing a short cognitive assessment as part of routine medical care.

Regarding the Altoida assessment, 70% found the tutorial helpful. Education level was not associated with perceived tutorial helpfulness, whereas older participants and participants with MCI were more likely to report that the tutorial was helpful (p < 0.001 and p < 0.01, respectively). Among assessment components, AR-based tasks were more frequently identified as the most challenging, particularly among participants with MCI compared with CU (AR “Back In Time”: 28.6% vs. 17.4%; AR “Day Out”: 43.6% vs. 31.0%). These responses reflect perceived relative difficulty among tasks rather than inability to complete them. In contrast, shape-drawing tasks were most frequently identified as the least challenging.

Most participants selected home as the preferred setting for completing a cognitive assessment, whereas fewer preferred primary care or specialist settings, and pharmacy was the least preferred. Among participants who provided usability feedback during the assessment, only 16% reported challenges, primarily related to AR navigation or interpretation of task instructions.

Performance of the Altoida NeuroMarker for MCI identification

The Altoida NeuroMarker demonstrated strong discrimination for MCI (ROC-AUC 0.90, 95% CI 0.88–0.92), with a sensitivity of 0.83 (95% CI 0.74–0.93) and a specificity of 0.83 (95% CI 0.75–0.90) (Fig. 2a). The model achieved an overall accuracy of 0.82 (95% CI 0.79–0.86), a positive predictive value (PPV) of 0.68 (95% CI 0.54–0.83), and a negative predictive value (NPV) of 0.91 (95% CI 0.85–0.97) (Supplementary Table S4).

Fig. 2.

A) Receiver operating characteristic (ROC) curves with area under the curve (AUC) from each fold of the MCI model’s five-fold cross-validation, overlaid with the curve reflecting the mean of the five individual folds based on 970 assessments from 638 participants (training set). B) Calibration plot comparing predicted probabilities of MCI generated by the Altoida NeuroMarker with the observed proportion of participants classified as MCI across probability bins. The dashed diagonal line represents perfect calibration, where predicted probabilities equal observed event frequencies. The Altoida NeuroMarker achieved a Brier score of 0.18, significantly lower than the prevalence-based null model (Brier = 0.25; Wilcoxon signed-rank p < 0.001), indicating that the model’s predicted probabilities are informative beyond base rate prediction. The null model, which assigns all assessments the observed MCI prevalence of the validation cohort, is shown as a single reference point on the diagonal.

In an independent validation cohort (n = 75), the MCI NeuroMarker showed comparable performance (ROC-AUC 0.87, 95% CI 0.80–0.93), with a sensitivity of 0.75 (95% CI 0.63–0.85), a specificity of 0.86 (95% CI 0.75–0.93), and an accuracy of 0.81 (95% CI 0.73–0.87). A sensitivity analysis restricted to the first assessment from each participant in the validation cohort yielded performance metrics comparable to those obtained using all available assessments (Supplementary Table S5). Exploratory analyses of model performance by biological sex and education level are reported in the Supplementary Results.

Calibration of the Altoida NeuroMarker was assessed using the Brier score, which compares predicted probabilities with observed outcomes. The Altoida NeuroMarker achieved a Brier score of 0.18, compared with 0.25 for a prevalence-based null model (Fig. 2b), indicating improved probabilistic accuracy relative to the prevalence-based null model (p < 0.001).

To further assess the contribution of assessment-derived features beyond demographic characteristics, the Altoida NeuroMarker was compared with a model based solely on demographic variables. In five-fold cross-validation, this demographics-only model showed limited discrimination for MCI (ROC-AUC 0.68, 95% CI 0.63–0.73). Consistent with these findings, the Altoida NeuroMarker significantly outperformed the demographics-only model in the independent validation cohort (AUC = 0.87 vs. 0.68, ΔAUC = 0.19, DeLong z = 3.63, p < 0.001), indicating that the performance of the Altoida NeuroMarker was primarily driven by assessment-derived features rather than demographic characteristics.

Composite score and cognitive domains

Participants with MCI showed significantly lower overall performance on the Altoida composite score relative to the CU group (mean 41.27 vs. 49.22; p < 0.001; Fig. 3a). Participants with MCI also showed lower scores across all cognitive domains (Fig. 3b). The largest difference was observed in Language (p < 0.001), followed by Complex Attention (p < 0.001), Perceptual-Motor Function (p < 0.001), Learning and Memory (p < 0.001), and Executive Function (p < 0.001). Because composite and cognitive domain scores represent age- and sex-matched percentiles relative to CU reference subsets, mean values near (but not exactly at) 50 are expected for the CU group. Group means below 50 in the MCI group indicate performance below that of age-and sex-matched CU peers. Across the 14 subdomains, mean scores were lower in the MCI group (Supplementary Fig. S2). The largest differences were observed in object naming, flexibility, visual perception, sustained attention, and processing speed.

Fig. 3.

Box plots depicting Altoida composite scores (A) and cognitive domain scores (B) in participants identified as cognitively unimpaired (CU) or with mild cognitive impairment (MCI). Data is presented as box and whiskers plots, with the median and 25 and 75 percentiles. ∗∗∗p < 0.001, Mann Whitney U test.

Test re-test reliability

The MCI classifier demonstrated consistent agreement across repeated assessments (κ = 0.60). The composite score and cognitive domains also showed good test-retest reliability (ICC = 0.80, 95% CI [0.75, 0.84]), with the highest ICC values observed for Complex Attention (ICC = 0.79, 95% CI [0.74, 0.84]), then Perceptual-Motor Function (ICC = 0.77, 95% CI [0.71, 0.82]), Language (ICC = 0.74, 95% CI [0.67, 0.80]), Executive Function (ICC = 0.64, 95% CI [0.56, 0.71]), and Learning and Memory (ICC = 0.55, 95% CI [0.45, 0.63]). Across subdomains, this pattern remained consistent (Supplementary Table S6).

Convergent validity

The Altoida Composite Score showed moderate positive correlations with MMSE scores (r_s = 0.40, r_s² = 0.16, n = 633, p < 0.001) and RBANS Total Index (r_s = 0.58, r_s² = 0.34, n = 228, p < 0.001), supporting convergent validity (Fig. 4).

Fig. 4.

Spearman correlation analysis between the Altoida composite score and MMSE (left) and RBANS Total Index scores (right). Significant positive associations with MMSE (rs = 0.40, n = 633, p < 0.001) and RBANS index (rs = 0.58, n = 228, p < 0.001) support convergent validity. The regression line is included for visualization only.

The Altoida Learning and Memory domain demonstrated small-to-moderate positive correlations with the FCSRT Immediate Free Total Recall (r_s = 0.37, r_s² = 0.14, n = 333, p < 0.001), Delayed Total Recall scores (r_s = 0.31, r_s² = 0.09, n = 333, p < 0.001), the RBANS Immediate Memory Index (r_s = 0.33, r_s² = 0.11, n = 228, p < 0.001) and the RBANS Delayed Memory Index (r_s = 0.27, r_s² = 0.08, n = 228, p < 0.001) (Fig. 5).

Fig. 5.

Analysis of convergence validity between the Altoida cognitive domains and established neuropsychological tests assessing the same constructs. Associations were evaluated using the Spearman rank correlations. Positive correlations indicate higher scores on both measures, whereas inverse correlations with Trail Making Test (TMT) scores reflect better performance with shorter completion times. Regression lines are included for visualization only.

The Altoida Complex Attention domain showed moderate positive correlations with the RBANS Attention Index (r_s = 0.46, r_s² = 0.22, n = 228, p < 0.001) and RBANS Coding (r_s = 0.50, r_s² = 0.25, n = 228, p < 0.001), together with a moderate inverse correlation with TMT-A completion time (r_s = −0.40, r_s² = 0.16, n = 568, p < 0.001; Fig. 5). The Executive Function domain exhibited small-to-moderate inverse correlations with TMT-A (r_s = −0.27, r_s² = 0.07, n = 568, p < 0.001), TMT-B (r_s = −0.35, r_s² = 0.12, n = 563, p < 0.001), and the difference between the TMT-A and TMT-B completion times (r_s = −0.33, r_s² = 0.11, n = 561, p < 0.001; Fig. 5).

The Altoida Language domain showed a small positive correlation with the RBANS Language Index (r_s = 0.30, r_s² = 0.09, n = 227, p < 0.001; Fig. 5), while the Perceptual-Motor Function domain demonstrated a moderate positive correlation with the RBANS Line Orientation subtest (r_s = 0.38, r_s² = 0.15, n = 228, p < 0.001; Fig. 5). Additional correlations for all domains are presented in Supplementary Table S7. Cross-domain analyses comparing corresponding and non-corresponding neuropsychological measures are presented in Supplementary Table S8.

Discussion

This study demonstrates that the Altoida NeuroMarker Platform can accurately identify clinically adjudicated MCI using a brief, self-administered assessment with a median completion time of under 15 min. Unlike digital versions of conventional cognitive tests, the Altoida Platform was designed to capture complementary aspects of cognitive and functional performance through multimodal motoric, speech, and augmented reality tasks. Being self-administered, the platform requires limited staff involvement, which may facilitate integration into frontline care. Depending on the clinical setting, this may include completion before the consultation following an initial clinical concern, or remote completion before the visit.

The Altoida Platform showed positive usability. Older participants and those with MCI were more likely to report that the tutorial was helpful, suggesting that the self-administered workflow remained accessible in groups expected to require greater support. Although augmented reality tasks were most frequently identified as challenging, study staff did not report participants requesting to stop the assessment because of distress, frustration, or unwillingness to continue. More broadly, survey responses suggested good acceptance of DCAs and potential acceptability of remote pre-visit workflows, although their feasibility and impact on diagnostic performance require further study.

Beyond accurate MCI identification, this study evaluated calibration, independent validation, test-retest reliability, and construct validity. Performance was maintained in an independent cohort, providing initial evidence of generalizability. Composite and cognitive domain scores also differed significantly between CU and MCI, supporting their ability to reflect differences in cognitive performance associated with clinical status.

Correlation analyses showed modest-to-moderate associations between the Altoida composite score, cognitive domains, and established neuropsychological measures assessing similar cognitive constructs, providing evidence of convergent validity. The magnitude of these associations is consistent with the Platform capturing related, rather than identical, aspects of cognition compared with conventional neuropsychological measures. Additional analyses comparing corresponding and non-corresponding neuropsychological measures further characterized construct validity. Differences between the corresponding and non-corresponding correlation coefficients were modest across domains and reached statistical significance only for the Language domain. These findings support the interpretation that the cognitive outputs predominantly reflect their intended cognitive constructs while retaining biologically plausible overlap with related cognitive processes.

Cognitive subdomain analysis provided additional insight into the breadth of measures captured by the Platform. While memory deficits are considered a classical hallmark of amnestic MCI, the Platform revealed differences across subdomains related to attention, executive function, visual perception, and perceptual motor coordination, signaling impairments that extend beyond memory and which are often underrepresented in traditional testing batteries.¹⁸ This broader coverage may be particularly relevant for capturing cognitive changes across heterogeneous and atypical MCI presentations. Our MCI cohort was not restricted to amnestic MCI and likely reflects etiologic and phenotypic heterogeneity.

The composite score and cognitive domains exhibited moderate to good test-retest reliability (ICC range 0.55–0.80), consistent with unsupervised digital cognitive tools.¹⁹ Higher reliability in Language and Attention scores likely reflects the relative stability of these cognitive abilities across retest intervals, whereas lower ICCs in Learning and Memory and Executive function are consistent with greater day-to-day variability in these measures and susceptibility to practice effects. Together with the reproducibility of the MCI classifier, these findings support further evaluation of the Platform for longitudinal monitoring of cognitive change across clinical and research settings.

Timely MCI identification has implications beyond treatment initiation. Recognition of cognitive impairment in primary care provides an opportunity to evaluate potentially reversible contributors to cognitive decline, optimize vascular and other modifiable risk factors, review medications that may affect cognition, engage patients and families in care planning, and establish appropriate longitudinal follow-up.²⁰ These clinical actions remain valuable regardless of whether an individual ultimately undergoes biomarker testing or becomes eligible for disease-targeting therapy. At the same time, identification of MCI should be accompanied by appropriate patient counselling, recognizing that a diagnosis may also generate anxiety, concerns about stigma, and uncertainty regarding prognosis. These considerations reinforce the importance of integrating objective cognitive assessment within comprehensive clinical care rather than using it in isolation.

Self-administered assessments may reduce barriers to evaluation by enabling testing in accessible settings and minimizing stigma associated with cognitive assessment.²¹ At the same time, blood-based biomarkers, such as pTau217 and pTau181, are transforming AD diagnostics by enabling less invasive detection of underlying pathology. However, while blood-based provide biological information, they do not directly capture cognitive or functional impairment and therefore require a clinical context for interpretation.²² Consequently, there is growing interest in cognitive assessment approaches that are accurate, scalable, and acceptable to patients, while also being capable of integrating into diagnostic pathways alongside emerging biomarkers.²³

Within this evolving landscape, DCAs represent a rapidly expanding class of tools with varying intended uses.²⁴ Although brief screening instruments such as the Mini-Cog and MoCA are established methods for identifying cognitive impairment and can be used in primary care, many newer DCAs seek to provide additional advantages, including automated scoring, richer quantitative outputs, self-administered workflows, and reduced reliance on trained examiners. Many DCAs focus on single modalities or a limited subset of cognitive domains.^19,25 The Altoida NeuroMarker Platform is distinguished by integrating motoric, speech, and augmented reality tasks within a single assessment. Because MCI is heterogeneous and extends beyond isolated memory impairment, this multimodal approach is designed to support broader characterization of cognitive and functional performance.

The outputs provided by the Platform are intended to complement, rather than replace, clinical evaluation. The MCI classifier is designed to identify the clinical syndrome of MCI and is not intended to determine its underlying etiology. Accordingly, the Platform should be interpreted alongside clinical history, medication use, and other components of a clinical evaluation. In frontline care, these complementary measures may help guide decisions regarding specialist referral, additional diagnostic evaluation (including biomarker testing where appropriate), and longitudinal monitoring of individuals who can be managed within primary care.

As with any clinical decision-support tool, the clinical interpretation of positive and negative results depends on the prevalence of MCI in the population being evaluated. The Altoida Platform is not intended for population-wide screening of asymptomatic individuals, but rather for use in people presenting with cognitive concerns in whom further cognitive evaluation is being considered. Within this context, the observed positive and negative predictive values reported in this study support the role of the Altoida Platform as an adjunctive tool, where a negative result may provide reassurance that MCI is unlikely, whereas a positive result should prompt further clinical evaluation rather than be interpreted as diagnostic in isolation.

Despite the strengths and scale of the training dataset, there are study limitations to consider. The training cohort intentionally combined participants recruited from community, research, and specialist memory clinic settings. Although this strategy exposed the ML model to a broad spectrum of cognitive presentations, performance within individual recruitment settings was not established. Additional validation will therefore be important to evaluate performance across specific clinical contexts. In addition, biomarker measures (CSF, PET, or blood) were not systematically available for etiologic characterization of the clinical syndrome. Accordingly, the cohort included participants with diverse neurodegenerative and non-neurodegenerative causes of cognitive impairment, and future studies incorporating fluid or imaging biomarkers will be important to evaluate performance across biologically defined subgroups. Despite the geographic diversity of the cohort, participants were predominantly White and non-Hispanic, with high self-reported digital comfort. Expanding diversity across race, ethnicity, socioeconomic status, and education remains a priority to ensure broad applicability. Equitable implementation will also require strategies that promote accessibility across diverse healthcare settings, accommodate varying levels of digital literacy, and ensure evaluation in underserved populations to avoid widening existing healthcare disparities. Finally, larger prospective validation across diverse healthcare settings will be important before broad clinical implementation.

In conclusion, the Altoida NeuroMarker provides a brief, self-administered approach for identifying MCI and generates interpretable cognitive metrics to support clinical evaluation. As neurological care enters the biomarker era, scalable DCAs may play an increasingly important role alongside biological testing.²⁶

Contributors

MFI, VBK, VBR, NG, SF, MC, MJ, and ES conceptualized and designed the study. JH contributed to the design and interpretation of the convergent validity analyses. GSB, CM, OGR, MSC, and CPM coordinated recruitment, clinical characterization, and data generation within the BETA-AARC cohort. EVC and DA coordinated recruitment, clinical characterization, and data generation within the SPIN cohort. LM, LW, JR, and JD coordinated recruitment, clinical characterization, and data generation within the ADNeT and AIBL cohorts. ATM and MB coordinated recruitment, clinical characterization, and data generation within the Hospital Clinic cohort. VBR performed data curation. VBR, GSB, EVC, LM, and ATM verified the study data. VBR, NG, and SI conducted data analyses and data visualization. IT and MJ contributed to funding acquisition. MFI interpreted the data and wrote the first draft of the manuscript. All authors contributed to data interpretation, critically revised the manuscript, and approved the final submitted version.

Data sharing statement

Deidentified participant data may be made available on a case-by-case basis for scientifically sound research proposals, subject to review and approval by the data owners, and execution of appropriate data use and confidentiality agreements. Data may be requested following publication by contacting the corresponding author. Availability of supporting documents, including the study protocol, will be considered as applicable and subject to the same review requirements.

Declaration of interests

MFI, VBR, NG, SF, MJ, and ES are employees of Altoida Inc. and hold stock options. VBK is a co-founder of deepPath.AI and Cognimark.AI. VBK and JH are paid members of Altoida’s Advisory Board. JH also reports royalties from Blackwell Publishers and Oxford University Press; consulting fees through Metis Cognition Ltd. from ADvantage Therapeutics, Actinogen, AlzeCure, Boehringer Ingelheim, Cambridge Cognition, Cognition Therapeutics, Curasen, EQT Life Sciences, Gallymedix, GfHEU, Julius Clinical, Kynexis, Manus Neuro, Novartis, Pearson, Scottish Brain Sciences, Shionogi, Signant, Teitur Therapeutics, Therini Bio, Transposon, Virogenics, Vivoryon Therapeutics, Winterlight Labs, and Xylo Bio; and stock or stock options in Scottish Brain Sciences. DA participated in advisory boards from Fujirebio-Europe, Roche Diagnostics, Grifols S.A., Schwabe, and Eli Lilly, received speaker honoraria from Fujirebio-Europe, Roche Diagnostics, Nutricia, Krka Farmacéutica S.L., Zambon S.A.U., Neuraxpharm, Alter Medica, Eli Lilly, and Esteve Pharmaceuticals S.A.; and received travel support from Fujirebio-Europe, Eli Lilly, Nutricia, and Novo Nordisk. DA declares holding a patent for markers of synaptopathy in neurodegenerative disease (licensed to ADx NeuroSciences N.V., WO2019175379). ATM declares receiving lecture honoraria from Eli Lilly. MB declares receiving lecture honoraria and conference travel support from Eli Lilly. SI served as a statistical consultant to Altoida from December 2025 to April 2026. OGR has given lectures sponsored by Roche Diagnostics and Idorsia Pharmaceuticals Ltd. MSC reports consultancy fees and/or participation on advisory boards for Eli Lilly, Grifols, Novo Nordisk, and Roche Diagnostics; and lecture honoraria from Roche Diagnostics, Almirall, Eli Lilly, Novo Nordisk, Quanterix, Biogen, and Beckman Coulter. MSC’s institution received in-kind biomarker measurements or materials from Roche Diagnostics International, Avid Radiopharmaceuticals (Eli Lilly), Janssen Research & Development, ADx Neurosciences, Alamar Biosciences, Fujirebio, Meso Scale Discovery, and ALZpath. MSC also served as a site investigator for a Roche Diagnostics clinical study. Other authors have nothing to disclose.