P3‐271: Automated index of semantic verbal fluency in mild cognitive impairment and Alzheimer's disease

Abstract

Background: Semantic verbal fluency (SVF), widely reported in AD and MCI, often show early and disproportionate decline relative to other language, attention, and executive abilities. Successful performance on the SVF test depends on how well conceptual information is organized into related clusters and whether the patient is able to use an efficient strategy to access these clusters. Current methods for clustering and switching behavior assessment are labor-intensive and subjective. We developed and tested an automated computational linguistic approach to cluster words generated on the SVF test. Methods: Participants were a random sample from the Mayo Clinic Alzheimer’s Disease Registry and the Mayo Clinic Study of Aging (20 probable AD, 21 MCI, 19 controls; controlled for age mean 72 y.o.). All participants underwent a cognitive assessment including the paper-andpencil SVF test subsequently converted to electronic form and a short test of mental status (STMS). The test was analyzed using a novel automatic clustering assessment tool based on semantic relatedness between pairs of words calculated using a variant of principal components analysis. Traditional SVF scores were compared to automatically computed mean cluster size (MCS) and a measure of cumulative relatedness between all pairs of words (CuRel) using logistic regression modeling to distinguish between diagnostic groups and linear regression to predict STMS scores. Results: Classification of MCI vs. controls with SVF score as the only predictor in logistic regression model resulted in smaller area under the curve (AUC) estimates for (MCI vs. controls AUC 1⁄4 0.70; AD vs. controls AUC 1⁄4 0.87) than classification with SVF/CuRel score and MCS as predictors (MCI vs. controls AUC 1⁄4 0.70 and 0.77 for CuRel; AD vs. controls AUC 1⁄4 0.97 and 0.90 for CuRel). Linear regression models that included the SVF score together with MCS resulted in significantly better fit with STMS scores than models with SVF score as the sole predictor (r 2 1⁄4 0.55 vs. r 2 1⁄4 0.38, respectively). Conclusions: Our preliminary study indicates that automatically assessed clustering behavior on the SVF test provides complementary information to traditional SVF scoring. Our computerized approach is objective, reproducible and easily scalable to large numbers of participants.