Predicting Age From Behavioral Test Performance for Screening Early Onset of Cognitive Decline

Abstract

Background: Neuronal reactions and cognitive processes slow down during aging. The onset, rate, and extent of changes vary considerably from individual to individual. Assessing the changes throughout the lifespan is a challenging task. No existing test covers all domains, and batteries of tests are administered. The best strategy is to study each functional domain separately by applying different behavioral tasks whereby the tests reflect the conceptual structure of cognition. Such an approach has limitations that are described in the article. Objective: Our aim was to improve the diagnosis of early cognitive decline. We estimated the onset of cognitive decline in a healthy population, using behavioral tests, and predicted the age group of an individual. The comparison between the predicted ("cognitive") and chronological age will contribute to the early diagnosis of accelerated aging. Materials and Methods: We used publicly available datasets (POBA, SSCT) and Pearson correlation coefficients to assess the relationship between age and tests results, Kruskal-Wallis test to compare distribution, clustering methods to find an onset of cognitive decline, feature selection to enhance performance of the clustering algorithms, and classification methods to predict an age group from cognitive tests results. Results: The major results of the psychophysiological tests followed a U-shape function across the lifespan, which reflected the known inverted function of white matter volume changes. Optimal values were observed in those aged over 35 years, with a period of stability and accelerated decline after 55-60 years of age. The shape of the age-related variance of the performance of major cognitive tests was linear, which followed the trend of lifespan gray matter volume changes starting from adolescence. There was no significant sex difference in lifelong dynamics of major tests estimates. The performance of the classification model for identifying subject age groups was high. Conclusions: ML models can be designed and utilized as computer-aided detectors of neurocognitive decline. Our study demonstrated great promise for the utility of classification models to predict age-related changes. These findings encourage further explorations combining several tests from the cognitive and psychophysiological test battery to derive the most reliable set of tests toward the development of a highly-accurate ML model.

Keywords: aging; biological age; cognitive decline; cognitive impairment; executive functioning; machine learning; neurodegeneration; psychophysiological tests.