Adaptive testing for smartphone‐based cognitive screening: Reducing patient burden while maintaining classification accuracy
Stephanie Ruth Young, Julia Yoshino Benavente, Elizabeth McManus Dworak, Michael S. Wolf, Cindy J. NowinskiAbstract
INTRODUCTION
Remote cognitive screening for primary care must balance accuracy with patient burden. We evaluated whether adaptive task administration could reduce testing while preserving classification accuracy.
METHODS
Using data from older adults ( N = 277; 100 mild cognitive impairment [MCI]; 177 normal cognitive aging) who self‐administered MyCog Mobile, we developed a stepped protocol in which the final task is skipped when preceding tasks classify the patient with confidence. Leave‐one‐out cross‐validation (LOO‐CV) assessed agreement with the full‐battery classification and accuracy relative to the reference diagnosis.
RESULTS
Thirty‐three percent of participants were classified at the initial decision point, and only 17% required the full battery. LOO‐CV showed a negligible change in area under the curve (ΔAUC = 0.008; bootstrapped 95% confidence interval [−0.004, 0.021]) using the adaptive battery.
DISCUSSION
Approximately one third of older adults can be accurately classified as with MCI using only two MyCog Mobile tasks under an adaptive approach. Prospective validation, in which the adaptive battery is administered in real time, is warranted.