Trajectory-based identification of cognitive-performance phenotypes across adulthood from psychophysiological testing

Multidimensional psychophysiological batteries reveal substantial inter-individual variation in processing speed, accuracy, memory, executive control, and visuospatial performance. Because the present sample is predominantly young to middle-aged, the analysis is framed as adult cognitive-performance phenotyping rather than identification of clinical cognitive-aging stages.

We analyzed a cross-sectional convenience sample of 1,117 adults (age 18-78 years; mean 31.4 ± 12.0 years; 65.8% women) assessed with a web-based psychophysiological battery. Forty-two cognitive, psychomotor, demographic, anthropometric, lifestyle, and self-reported health variables were standardized after singular-value-decomposition imputation of one missing BMI value. DANCo and local PCA estimated intrinsic dimensionalities of 5.83 and 3.82, respectively. Six principal components (39.24% cumulative variance; seventh-component increment 3.67%) were used to fit an elastic principal tree. Between-branch differences were evaluated by Kruskal-Wallis tests with Holm-adjusted Dunn comparisons or chi-square tests, with effect sizes. Assignment reproducibility was evaluated in 100 random 80% subsamples projected onto the fixed reference tree. The workflow is an unsupervised statistical/geometric structure-learning analysis rather than supervised prediction; no train/test predictive model or learned longitudinal dynamics are claimed.

Three connected branch profiles were identified: Cluster 0 (n = 781, 69.9%), a high-accuracy reference profile; Cluster 1 (n = 98, 8.8%), a rapid-processing profile with domain-specific visuospatial variability; and Cluster 2 (n = 238, 21.3%), a slower, lower-accuracy profile enriched for older age and positive insomnia/disease indicators. Moderate-to-large effects were observed for Stroop-4 speed (η² = 0.195), Stroop-4 accuracy (η² = 0.194), figure-shape accuracy (η² = 0.240), figure-shape-color accuracy (η² = 0.251), and figure-shape-position accuracy (η² = 0.253; all p < 0.0001). Fixed-tree projections reproduced assignments with mean adjusted Rand index, normalized mutual information, and raw agreement of 1.000 ± 0.000. In the additional sensitivity analyses, exclusion of the binary variables changed the solution from three to nine clusters (ARI = 0.1203), and reduction from six to four PCs changed it from three to seven clusters (ARI = 0.1736). All binary features in the dataset are directly or indirectly related to cognitive functions or to the interpretation of psychophysiological test results. Removing the complete binary-variable block therefore changed not only the numerical feature space but also the substantive interpretation of the analysis, particularly because this block included age-associated lifestyle and health factors such as smoking, alcohol use, insomnia, and disease status.

Elastic principal trees represent cognitive-performance profiles as connected branches and provide a graph-ordering coordinate unavailable from ordinary discrete clustering. The results are exploratory, cross-sectional, and non-diagnostic and should not be generalized to neurodegenerative aging without older, clinically characterized, longitudinal cohorts. Among the examined alternatives, the original full-feature six-PC solution was retained for interpretation because the alternative solutions produced additional very small clusters that could not be characterized reliably.