JMIR Publications

ABSTRACT

Background:

Scalable tools to identify individuals likely to have cognitive impairment (CI) are important for Alzheimer disease and related dementias research. The Everyday Cognition scale (ECog) and its short form (ECog12) capture subjective cognitive and functional decline and are useful in predicting CI. Automated, speech-based story recall tasks have separately shown sensitivity to early Alzheimer disease. Remote digital cohorts increasingly collect questionnaire- and speech-based measures from the same unsupervised participants at scale, but whether combining these measures improves discrimination of CI beyond a self-reported ECog12 score alone has not been examined.

Objective:

This study aimed to evaluate the ability of the self-reported ECog12 (Self-ECog12), alone and combined with demographics and self-reported memory items, to discriminate CI from cognitively unimpaired (CU) individuals in a remote digital cohort, and to test whether adding speech-recall features further improved discrimination.

Methods:

Participants were 7268 individuals from the Alzheimer's Disease Neuroimaging Initiative 4 (ADNI4) Remote Digital cohort with complete baseline Self-ECog12, memory item, demographic, and speech-recall data; CI status (n=574, 7.9%) reflected self-reported prior diagnosis of mild cognitive impairment, Alzheimer disease, or dementia. The majority class was undersampled to a balanced 1:1 sample (n=1148; train n=861, test n=287), and nested logistic regression models were compared by area under the receiver operating characteristic curve (AUC), with successive AUCs compared using the paired DeLong test.

Results:

Self-ECog12 alone discriminated CI from CU with an AUC of 0.791 in the held-out test set (n=287). Adding demographics raised discrimination to AUC=0.816, and further adding self-reported memory items raised it to AUC=0.836. Adding speech-recall features significantly improved discrimination beyond Self-ECog12, demographics, and self-reported memory items (AUC=0.861 vs 0.836, DeLong P=.02), corresponding to a sensitivity of 0.80 and specificity of 0.79 at the optimal cut point. The full model also significantly outperformed Self-ECog12 alone (AUC=0.861 vs 0.791, DeLong P=.002). In a smaller subgroup with APOE ε4 genotype available (n=138), adding APOE4 carrier status did not change discrimination (AUC=0.673 without vs 0.667 with), a negative finding.

Conclusions:

Self-ECog12 alone identified cognitive impairment with good accuracy, and adding automated speech-recall features provided statistically significant incremental discrimination beyond Self-ECog12, demographics, and self-reported memory items in this remote digital cohort. These findings support combining questionnaire and speech-based remote assessments as a scalable, low-burden adjunct for identifying older adults at risk of cognitive impairment, supporting earlier detection and referral for comprehensive evaluation. Because CI status reflected self-reported diagnostic history rather than independent clinical evaluation, these results should be confirmed against clinician-diagnosed outcomes before informing screening decisions.