Background:

Incorporation of cognitive screening into the busy primary care will require the development of highly efficient screening tools. We report the convergence validity of a very brief, self-administered, computerized assessment protocol against one of the most extensively used, clinician-administered instruments—the Montreal Cognitive Assessment (MoCA).

Method:

Two hundred six participants (mean age = 67.44, standard deviation [SD] = 11.63) completed the MoCA and the computerized test. Three machine learning algorithms (ie, Support Vector Machine, Random Forest, and Gradient Boosting Trees) were trained to classify participants according to the clinical cutoff score of the MoCA (ie, < 26) from participant performance on 25 features of the computerized test. Analysis employed Synthetic Minority Oversampling TEchnic to correct the sample for class imbalance.

Results:

Gradient Boosting Trees achieved the highest performance (accuracy = 0.81, specificity = 0.88, sensitivity = 0.74, F1 score = 0.79, and area under the curve = 0.81). A subsequent K-means clustering of the prediction features yielded 3 categories that corresponded to the unimpaired (mean = 26.98, SD = 2.35), mildly impaired (mean = 23.58, SD = 3.19), and moderately impaired (mean = 17.24, SD = 4.23) ranges of MoCA score (F = 222.36, P < .00). In addition, compared to the MoCA, the computerized test correlated more strongly with age in unimpaired participants (ie, MoCA ≥26, n = 165), suggesting greater sensitivity to age-related changes in cognitive functioning.

Conclusion:

Future studies should examine ways to improve the sensitivity of the computerized test by expanding the cognitive domains it measures without compromising its efficiency.