We hypothesised that item-level Category Fluency (CF) scores would predict performance in other neuropsychological tests, and that semantic metrics would be better descriptors of CF word complexity than non-semantic metrics.

CF words from 266 participants (Table 1) with Mild Cognitive Impairment were scored as a function of 11 semantic (Box 1) and 5 non-semantic (Box 2) metrics (Table 2, Fig. 1).

Partitional k-means Cluster Analyses were carried out on these 16 variables to capture word complexity in a unitary way and clusterise each cohort in a data-driven manner.

k = 2 and k = 3 were identified as the best k-solutions, as prompted by the within-cluster sum of squares "elbow" point, and by the Silhouette method (Fig. 2). Non-parametric inferential tests (Wilcoxon Signed Rank Test for k = 2, and Kruskal Wallis Test for k = 3, conservatively thresholded at a p < 0.001) were run to compare, across the resulting clusters (Fig. 3), performance in 10 neuropsychological tests. Additional models were run to compare individual metrics.

Between-cluster differences were found, under both k-solutions, in quantitative CF scores and Letter Fluency (LF) performance (Table 3, Fig. 4). Differences across clusters were found in relation to both semantic and non-semantic metrics.

A large body of research has reported altered item-level CF scores in people with Mild Cognitive Impairment or dementia. This study shows that item-level scores can also predict quantitative LF performance in people with Mild Cognitive Impairment. These findings also indicate that both semantic and non-semantic metrics play a central role in characterising complexity of CF words.