Clinical Validation and Machine Learning Optimization of MyCog: A Self-Administered Cognitive Screener
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Importance
Early detection of cognitive impairment in older adults is crucial for identifying reversible causes, managing symptoms, and enabling future planning; yet primary care settings face significant barriers to implementing effective cognitive screening.
Objective
To clinically validate an innovated self-administered, tablet-based cognitive screening tool (“MyCog”, adapted from the NIH Toolbox Cognitive Battery) and to develop an optimized scoring algorithm for detecting cognitive impairment in older adults during primary care visits.
Design
Cross-sectional clinical validation study conducted between summer 2021 and fall 2024, comparing cognitive performance between impaired and non-impaired participants using multiple machine learning approaches to identify optimal predictors of impairment.
Setting
Academic medical centers and community-based general population recruitment.
Participants
248 adults aged 65 and older, including 70 participants with clinically diagnosed cognitive impairment (MCI or dementia), 81 participants confirmed to be cognitively intact from clinical settings, and 97 participants from the general population. Participants were predominantly White (88%) with varied educational backgrounds and a mean age of 76.3 years ( SD = 6.9).
Main Outcomes and Measures
Sensitivity, specificity, and area under the receiver operating characteristic curve (ROC AUC) were the primary metrics used for evaluating model fit in classifying cognitive impairment of known groups from the MyCog Picture Sequence Memory and Dimensional Change Card Sort tasks.
Results
Across nine machine learning model classes, Picture Sequence Memory exact match score and Dimensional Change Card Sort rate-correct score were identified as the most robust predictors of cognitive impairment. The final optimized model achieved a sensitivity of 82.5% (bootstrapped 95% CI [0.733, 0.913]) and a specificity of 76.8% (bootstrapped 95% CI [0.709, 0.828]) with an ROC AUC of 0.892, comparable to traditional, long form cognitive screening measures while requiring less than 7 minutes to self-administer.
Conclusions and Relevance
MyCog demonstrates strong diagnostic accuracy for detecting cognitive impairment in older adults with a parsimonious model focusing on episodic memory and executive function. As a brief, self-administered screening tool that can be integrated into electronic health records, MyCog addresses key barriers to cognitive screening in primary care settings, which may help increase screening and early detection rates.
