Can remote smartphone cognitive tests detect early Alzheimer's disease?

Testing whether a phone battery measures cognition

The paper asks whether cognitive testing done at home on a smartphone can produce usable measurements in the earliest stages of Alzheimer's disease. The authors built the AD Digital Assessment Suite (AD-DAS), nine active tasks covering episodic memory, working memory, language, processing speed, response inhibition and motor speed, and deployed it remotely for 28 days.

The design matters for interpreting every number that follows. 123 participants were enrolled across four known groups: 32 healthy controls, 31 people with subjective cognitive decline and a negative amyloid-PET scan (SCDn), 30 with subjective cognitive decline and a positive amyloid-PET scan (SCDp), and 30 with early Alzheimer's disease (eAD). Amyloid-PET status was centrally read, so the SCD split reflects measured amyloid pathology rather than clinical impression alone.

The study is cross-sectional and the sample is small and demographically limited, which the authors state requires validation in an independent, larger study before any diagnostic accuracy claim can be made. The results below are therefore preliminary validity evidence, not a screening performance estimate.

Most participants finished, but eAD users needed more time

Completion was high: 97.6% of participants finished the study. The group breakdown is more informative than the headline. Completers were 97% of healthy controls, 100% of SCDn, 100% of SCDp and 93% of eAD; the two voluntary withdrawals were both in the eAD group, alongside one healthy volunteer who withdrew because of COVID.

Daily burden was modest but not identical across groups. The overall median time spent on remote assessments was 14.5 minutes per day, ranging from 13.2 minutes in healthy controls to 16.6 minutes in eAD. Adherence among completers was about 96%, roughly 27 to 28 of 28 days, although a lower proportion of healthy controls completed the tests (89.3%) than the other subgroups.

Acceptability was measured with a user-experience questionnaire. Over 85% of participants rated the overall experience as good (46.7%) or very good (38.5%); six participants, three SCDp and three eAD, rated it poor, and none rated it very poor. Study partners reported that 9 of 30 eAD participants needed assistance with the tasks, compared with 4 of 30 SCDp, 1 of 31 SCDn and 0 of 32 healthy controls, so unsupervised use is not equally realistic across the spectrum.

Reliability was moderate to excellent, task by task

Test-retest reliability asks whether a task gives a similar result when repeated, which is a precondition for using it to track change. The AD-DAS metrics ranged from an intraclass correlation coefficient (ICC) of 0.53 for speeded tapping with the non-dominant hand to 0.91 for speeded tapping with the dominant hand, a span the authors describe as moderate to excellent.

The range is the point: reliability is not a single property of the battery. A reader comparing tasks should treat the dominant-hand tapping metric and the non-dominant-hand metric as different measurements with different stability, rather than assuming all AD-DAS outputs carry the same precision.

Phone metrics track clinic tests only modestly

Convergent validity was tested by comparing AD-DAS metrics with established in-clinic neuropsychological tests. At least one metric from every AD-DAS task was associated with its clinical comparator in the expected direction, with the exception of the Fairytale task, whose results were not shown because of missing data.

The effect sizes were modest: adjusted R-squared values ranged from 0.01 to 0.24. For example, total time on Gallery Game Learning was negatively associated with FCSRT free recall during encoding (R-squared 0.15), and the percentage of correctly learned objects was positively associated with the same measure (R-squared 0.18). Gallery Game Recognition accuracy was only marginally associated with FCSRT sensitivity to cueing (R-squared 0.04 for under 24 hours, 0.01 for over 24 hours).

The authors attribute the relatively low adjusted R-squared values to covariate adjustment and limited sample size. Practically, this means the smartphone tasks overlap with clinic tests but are not interchangeable with them; a low R-squared is not evidence that a task is invalid, but it does mean the phone metric carries information the clinic test does not capture.

eAD separated from known groups on most metrics

Group comparisons used the log odds of group membership, and the paper reports absolute differences of 0.6 to 3.28 between eAD and the other known groups. Most AD-DAS metrics differentiated eAD from the other groups. Two exceptions are informative: 30-second dual-task walking gait speed and non-dominant-hand mean tapping interval separated eAD from healthy controls and from SCDp, but not from SCDn.

The in-clinic scales behaved differently. A-IADL-Q-SV, FCSRT, TMT, BNT, VFT and DSCT differentiated groups, while the Timed 25 Foot Walk and Timed Up and Go separated eAD from healthy controls only, not from the other groups. So the digital battery is not uniformly more sensitive than clinical scales; the comparison depends on which task and which group pair is examined.

The authors also note that case-control differences on Gallery Game performance were not as pronounced as expected, and speculate that its incremented-difficulty design may have let participants, including eAD patients, benefit from a spacing effect. That is an author hypothesis, not a tested mechanism in this study.

Memory tasks split SCD by amyloid status

The most distinctive result concerns the two subjective cognitive decline groups, which differ by amyloid-PET status rather than by clinical diagnosis. SCDp participants performed worse than SCDn participants on Gallery Game Recognition (under 24 hours), Gallery Game Free Recall (under 24 hours) and the highest level reached on Find the Egg, a visuospatial working memory task.

Among the classical neuropsychological scales, which were not used to include or exclude SCD participants, FCSRT free recall after a 30-minute delay and BNT showed nominal SCDn-SCDp differences. The authors read the digital findings as preliminary evidence that AD-DAS metrics can detect subtle cognitive change early in the AD spectrum, and they connect the episodic memory difference to a heightened risk of progression to MCI, while stating that the predictive significance of the working memory impairment needs further investigation.

This is a group-level difference in a small sample, not a demonstrated ability to classify an individual. It also does not establish that these tasks predict who will progress.

Task scores line up with atrophy in expected regions

As an orthogonal check, the authors related AD-DAS performance to gray matter volume from structural MRI using whole-brain voxel-based morphometry. Whole-brain analysis matters because it tests specificity: a valid task should correlate with atrophy in regions that support the function it claims to measure, and not in unrelated regions.

The patterns followed that expectation. Longer time to reach criterion on Gallery Game Learning was associated with reduced gray matter in the right middle occipital lobe and bilateral perirhinal cortices, regions implicated in visual processing and episodic learning. Longer completion time on Gallery Game Free Recall (under 24 hours) was associated with reduced volume in the left perirhinal and entorhinal cortices. Response inhibition accuracy on the Tilt Task was associated with clusters including the calcarine sulcus, bilateral hippocampi and left inferior orbital cortex and insulae.

Some results were null in a way that constrains interpretation. No clusters were identified for Gallery Game Recognition accuracy or completion time, although the over-24-hour hit rate was associated with left thalamus volume. Noun production rate in Story Time Narration, BNT and VFT did not yield supra-threshold clusters, and no clinical measure of visuospatial working memory was available to compare with Find the Egg.

A candidate tool awaiting confirmatory validation

The paper's own priority is not diagnostic accuracy but the sequence from feasibility to neuroanatomic validity: first show that people will do the tests at home, then that the metrics are stable, then that they relate to clinic measures and to brain structure in the expected places. On that sequence, the strongest evidence is operational, with 97.6% completion and about 96% adherence, and the weakest is the modest overlap with clinic comparators (adjusted R-squared 0.01 to 0.24).

The result most worth watching is the SCDn versus SCDp split on episodic and working memory tasks, because it is the comparison closest to the preclinical end of the spectrum and the one where amyloid status, not clinical diagnosis, defines the groups. The authors frame AD-DAS as a candidate tool pending confirmatory validation, and they call for larger normative and longitudinal datasets to control demographic effects and to establish norms for change.

For a reader weighing whether remote smartphone testing is ready for early Alzheimer's detection, the paper supports a narrower judgment: the battery is deployable and reliable enough to justify that larger study, and its memory metrics are the ones that should be tested first for sensitivity to amyloid-defined early change.

Sources used in this answer