Several processes may account for the declines associated with cognitive aging (Zelinski et al., 2011). Two of them, atrophy and neuroplasticity, are in direct theoretical opposition. The concept of brain atrophy is intrinsically associated with vascular factors (Borja et al., 2020) and inflammation (Alkasir et al., 2017), dietary habits (Seetharaman, 2016), metabolic disorders (Komuro et al., 2020), and dysbiosis (Alkasir et al., 2017). These factors provide important clues to dementia-related mechanisms and support the concepts regarding the management of cognitive impairment through the modulation of these factors. In contrast, the concept of brain neuroplasticity suggests that behaviors such as aerobic exercise and direct cognitive training can serve as compensatory factors that may potentially reduce the appearance of cognitive aging and hinder the signs of dementia (Zelinski et al., 2011).

Declines vary by cognitive domain. Many of them are subject to deterioration, whereas others, such as language, remain stable. In general, age-related changes in reaction speed underlie changes in a number of cognitive domains (Salthouse et al., 2003). The decline in information processing affects executive functioning (EF), a cognitive domain that accounts for individual goal-directed behavior. This may be explained in two ways. The first is that relevant operations cannot be executed within the required (limited) period. The second is that slow processing reduces the amount of simultaneously available information and that the higher-level processing shuts down without the information supply (Salthouse, 1996). For this reason, EF deficit has the potential to affect performance in a wide variety of cognitive variables and may serve as a potential mediator of age-related cognitive decline (Salthouse et al., 2003).

Studies suggest that there are several interdependent cognitive aging mechanisms. Despite the common pattern of findings for different speed-based tasks, cognitive slowing is not known to be a general factor of decline in abilities (Zelinski et al., 2011; Anderson and Craik, 2017). For example, working memory is required for recalling a task. It correlates positively with reasoning. Logically, its impairment has substantial ramifications for cognitive performance in older adults.

Executive Functioning (EF) and Cognitive Control. Cognitive control is the ability to orchestrate our thoughts and actions per internal goals. EF is commonly used as an umbrella term that covers both EF and cognitive control and applies to a set of higher-order (cognitive) processes involved in organizing intentional behavior in a novel situation (Miyake and Friedman, 2012). EF comprises three subdomains: inhibitory control that involves the preclusion of irrelevant information and the prohibition of prepotent responses; task switching that is the ability to switch flexibly between mental sets; and updating that is the constant monitoring and rapid addition/deletion of working memory contents. Decreased inhibitory control is a feature of age-related cognitive decline (Maldonado et al., 2020). Switching between different tasks requires working memory if a cue is absent, and a decline of the memory domain across time may impact this EF (Aschenbrenner and Balota, 2015; Moreira et al., 2018).

Attention is typically divided into two global subdomains: selective attention (e.g., concentration) and sustained attention or vigilance (e.g., divided attention) (Harvey, 2019). The elderly have a reduced ability to pay attention to selected stimuli. Age-related declines in selective attention interact with other cognitive domains and other known age-related changes (Zanto and Gazzaley, 2017).

There are other attentional subsystems. For example, it was shown that aging affects the balance between goal-guided and habitual spatial attention (Twedell et al., 2017). By the habitual spatial attention, authors of this study understand attention to the target that attracts notice more frequently, i.e., a space with the high-probability of triggering stimuli. Older adults suffer from disrupted goal-guided attention, but they are relatively unimpaired in deploying spatial attention through incidental habit-based learning. Hence, training older adults to develop good search habits may help compensate for a decline in some other attentional subdomains. All the attentional skills have EF components, the age-related decline of which impacts attention (Harvey, 2019). Older adults have a deficit in inhibiting irrelevant information, presumably because of changes in the prefrontal cortex (Zanto and Gazzaley, 2017).

Memory, as a cognitive domain, may be affected by multiple processes including speed, working memory, executive control operations, and sensory declines. Memory declines in aging may stem from a deficient ability to dynamically allocate attention and switch between functional brain networks. Because of structural and functional brain changes, older adults do not ignore distraction but, rather, co-encode relevant and irrelevant information. This leads to the overload of the limited cognitive resources (Zanto and Gazzaley, 2017).

Despite the well-established fact that memory performance declines with age, not all aspects of memory are impaired equally (Balota et al., 2000). The procedural memory is manifested without the direct recollection of the previous events, whereas the declarative memory is revealed through intentional retrieval of previous experience. It encompasses the episodic memories on some events and the semantic memories that reflect our general knowledge of facts and word meanings. In general, older adults have the most significant memory deficits that appear in long-term episodic memory because of the major attention demand. The semantic, sensory, and procedural memory have a minimal demand for attention, and they produce relatively little age-related change in performance (Balota et al., 2000; Zanto and Gazzaley, 2017).

The mediation of memory with other cognitive domains is described with the concept that semantic memory is an organized data storage of words or concepts - “nodes” - connected to other nodes via associative pathways. When an individual by directing attention activates a node, the activation spreads through the network from the node to other related nodes for subsequent processing (Balota et al., 2000).

Working memory as a temporary buffer for cognitive processing undergoes such age-related changes as a reduction of its capacity and speed of processing. The reasons for the changes are a decline in the processing speed and/or a breakdown in the basic control processes (resistance to interference, task coordination, memory updating, binding, and/or top-down control as inferred from neuroimaging data). Working memory is more age-sensitive regarding spatial rather than verbal material (Verhaeghen, 2018).

The age-related sensory and perceptual deficits exacerbate cognitive decline and complicate the performance of cognitive tasks. Declines in these domains indicate a reduction of the ability to detect a stimulus that occurs in one of the five sensory modalities, to process and integrate the acquired information. Clinically, the elderly may experience a variety of challenges in the identification of objects, sounds, tastes, smells, and tactile sensations. These include agnosia, the inability to recognize previously identifiable objects (Harvey, 2019). Cognitive aging in sensation and perception domains impacts other domains in turn. For example, at any age, hearing deficits lead to auditory memory problems and may affect language comprehension (Zelinski et al., 2011). There are strategies to compensate for declines in perceptual processes. The “posterior-to-anterior shift in aging” model by Davis describes the increased prefrontal activity as a means to overcome declines in perceptual processes that occur in occipital regions. This accounts for the task performance comparable in both older and younger adults, although it utilizes different neural mechanisms (Zanto and Gazzaley, 2017).

Motor skills (the basic elements of motor activity) and construction (the ability to either copy or produce drawings of common objects, as in clock drawing paradigms) may be impaired in severe cases of dementia, damage to the nondominant hemisphere, or lesions to the parietal cortex (Harvey, 2019).

Language. Some cognitive processes associated with language remain stable or improve into the mid-70s (Zelinski et al., 2011). Language deficits may be associated with a deficit in EF (e.g., the ability to access semantic storage successfully) or with a slowed processing speed (Harvey, 2019).

A general slowness of a neurologic response is the major factor in comprehension impairment. Additionally, attentional changes and memory decline impact speech comprehension (Obler and Albert, 1981).

Inseparability and intercorrelation of cognitive domains. Cognitive domains should not be viewed as lacking validity if they are intercorrelated. There is considerable evidence that in many patient populations, including those with schizophrenia and bipolar disorder, conventional domains of cognitive dysfunction are not truly separable (Harvey, 2019).

To estimate the psychophysiological status, we used a battery of neurophysiological tests that involve such cognitive domains and subdomains as attention, working memory, information processing speed, task switching and inhibitory control, and executive control (Statsenko and Charykova, 2010; Statsenko et al., 2019, 2020). The tests and their dependent output variables are listed below.

We analyzed the data from a publicly available dataset named after the title of the project Psychophysiological Outcomes of Brain Atrophy (POBA). The dataset contains about 100 features reflecting the overall psychophysiological status of 231 people of different age (4–83 years; 134 women, 97 men). Written patient consent or parental consent with assent from minors for testing and scanning was obtained in each case. The examinees were scanned to exclude brain pathology. Some of them suffered from periodic headaches. The others started their professional sports career. Not to impact the study outcomes, the examinees were not paid for participation or being tested.

The inclusion criterion was literacy, i.e., only those adults who indicated that they did at least professional courses after finishing general education took part in the study.

The exclusion criteria were as follows: organic brain pathology, mental disorders, and recent or past injury to the head. All the participants were examined by a qualified neurologist. They were found healthy based on the clinical examination and the negative results of MRI which was conducted to rule out any underlying pathology. In this way, we followed the recent practice parameters of the American Academy of Neurology that recommend neuroimaging (e.g., MRI) to detect neurodegenerative conditions in early stages (Fuller et al., 2019). Supplementary Table 1 shows the distribution of the subjects across age and sex groups. The dataset is provided on demand (see section 9). Additional details on the study design and description of all the features are available in our recent studies (Statsenko et al., 2020, 2021a).

To find the appropriate solution for the first objective, we analyzed the structure of the complex visual-motor reaction (CVMR), which is also commonly called the choice reaction. Similar to the simple visual-motor reaction, the choice reaction includes such components as sensory acquisition (visual perception) and motor responding. Additionally, CVMR encompasses the decision-making component required to process an inhibitory condition present in the task (see Figure 4). This processing causes a time delay (DMT). In our recent study, we analyzed the age-related variability of DMT. In this study, we want to concentrate on the ratio between DMT and SVMR time, because they reflect different cognitive functions: the switching and inhibitory control estimate vs. the information processing speed estimate.

We also found the analysis of RMO_mean variable relevant to this study. In the RMO task, the examinee is supposed to respond subsequently to a set of expected events accurately at the anticipated moments when the events occur. To perform the test, the individual employs such cognitive domains as attention, inhibitory control, and task switching. Thus, the dependent variable of the reaction to a moving object test (RMO_mean) reflects the cooperative involvement of different cognitive domains.

The second objective was to find the optimal number k of distinct nonoverlapping subgroups with regard to age. For this, we used a heuristic approach based on the elbow method. We utilized the K-means algorithm and evaluated the distortion score or the sum of squared distances from each point to its assigned centroid as a performance metric of each chosen k. The optimal cutoff k value corresponded to the minimal distortion score. When the optimal k was set, we utilized the K-means clustering method to assess the separability measure with regard to the age group by finding a centroid for each cluster and counting the number of relevant and irrelevant data points in the groups.

To address the third objective, we analyzed the charts that describe age-related changes of the aforementioned indices and dependent variables. Using the bootstrap method with regard to age, we built the linear regression model fit and 95% CI for the regression estimate. Then we analyzed the progression or dynamics of the proposed indices. To compare lifelong changes of the variables in different age-groups, we used descriptive statistics and the Kruskal-Wallis test (Habuza et al., 2021c). To figure out whether there is statistically significant differences between the data for two sexes, we resorted to statistical hypothesis tests, specifically t-test.

The fourth objective was multifold. We hypothesized that the newly developed indices may reflect age-related changes of the psychophysiological status of the examinees. To look for the possible associations of the age, indices, and other psychophysiological tests, we fed these data to conventional ML models, trained with a 10 fold cross-validation technique, and compared their performance.

In the first part of the fourth objective, we used ML classification algorithms to predict the age group from an individual performance in PTs. The threshold value was 40 years of age. Our recent study justifies this age to be a cut-off value for cognitive decline that can be identified from test performance (Statsenko et al., 2021a). In this study, we utilized classification models. For this, we added a single newly proposed index to the list of predictors used in the recent study (see the left column in Supplementary Table 2). We employed conventional ML classification models with default architecture from Python scikit-learn v. 0.24.2 library. The classifiers we used are listed in comments to Table 5. We trained all the models in the stratified k-fold cross-validation technique until convergence as per scikit-learn default settings. The results predicted in each fold were merged and then averaged to report the final accuracy of the models. By comparing the model performance metrics, we tested the informative input of the indices to the prediction of the age. To evaluate the performance of the predictive models, we generated a receiver operating characteristic (ROC) curve averaged over 10 times. We also calculated the mean sensitivity, specificity, accuracy (ACC), and area under the curve (AUC) values. These performance measures were suitable as the dataset was balanced across the age attribute. Then, we averaged the accuracy of the models by calculating Mean ± SD for each index used as a predictor (see the left columns of Table 5). Finally, with the Kruskal-Wallis test we tested whether the variance of the accuracy differs significantly (p > 0.05) for an index compared to other indices.

Working on the second part of objective four, we figured out whether the proposed indices provide a summary of the findings obtained while testing individuals. To address the task, we used ML regression models forecasting values of the newly proposed indices. As predictors, we utilized the features comprising POBA dataset with an exception of those from which the values of the indices can be calculated. Specifically, to predict the values of ISD and ISDA indices we used the variables listed in the middle column of Supplementary Table 2. To forecast the value of ISCA index, we used the same list of predictors with an exception of “SVMR_IES” as the prognosis can be calculated from this feature. To evaluate the quality of regressor outcome, we employed mean absolute error (MAE), root mean squared error (RMSE), and a proportion of MAE to the range of values (see the right columns of Table 5).

All the experiments were conducted with the Linux Ubuntu 18.04 workstation with 24 CPU cores and two NVIDIA GeForce GTX 1080 Ti GPU with 11 GB GDDR5X memory each using programming language Python, and its libraries for data processing, ML, and data visualization, such as scikit-learn, NumPy, Pandas, Matplotlib, Seaborn, and Yellowbrick.

Based on the analysis described in section 3.3, we came up with the index of simple reaction time to decision-making time (ISD). It provided us with the ratio of processing speed to decision-making time (see Equation 5). Both the time estimates were vulnerable to age-related neurocognitive slowing; however, there are no clear data supporting the assumption that the pace of decline is equal in diverse cognitive domains. Therefore, the derivative variable may serve as a marker of their disproportional decline.

ISD index considers two indicators that constitute the reaction time of the visual-motor task with the switching condition in it. The weak point is that it does not take the performance accuracy into consideration. For this reason, we proposed an additional derivate variable, which was the index of simple reaction time to decision-making time with the accuracy performance (ISDA). If compared with the previous one, the index includes the percentage of correct responses in the denominator (see Equation 6).

Another way to combine the speed and accuracy estimates of SVMR and CVMR is to calculate the IES score for each of the tests separately and then find the ratio between them. This solution leads us to the index of performance in simple and complex visual-motor reaction with account for accuracy (ISCA) calculated as it is seen in Equation 7.

In this research, we did not intend to test all the possible ratios of the performance metrics in different cognitive tasks. Our goal was to show the utility of the approach when cognitive testing covers several domains and estimates their interrelated divergent changes.

In the cluster analysis, we used two variables (age and a proposed index) and employed K-means method to segregate the data points between groups. We utilized the unsupervised learning clustering K-means method to find the optimal number of homogeneous groups of data points. With the elbow method, we found the most appropriate number of clusters. It was justified by the separability measure based on a distortion score. The score is the sum of squared distances from each point to its assigned centroid. All centroids are obtained iteratively by minimizing the intracluster proximity while maximizing the distance between clusters. We applied the elbow method to two-dimensional data points composed of the age of the subjects and corresponding index values, where the index was set to ISD, ISDA, or ISCA.

The knee point detection algorithm (Satopaa et al., 2011) returns the optimal value of clusters equal to four for each proposed indices. In Figure 5, built for the data points (age, ISCA), the optimal choice is annotated with a black dashed line. The blue line on the graph is plotted from the values of distortion scores with regard to the number of clusters, whereas the green dashed line displays the amount of time needed to train the clustering model per k.

The participants in our sample were not uniformly distributed over age; however, with the proper choice of bin width, the age histogram can become very close to the uniform one. Another concern is related to the number of subjects who were over the age of 75 years old, a relatively small subgroup, in contrast to the number of participants who were under 15. To get a relatively equal number of participants in each group, we decided to count the 20 years long intervals from birth rather than from the age of the youngest examinee.

We assessed the performance of the clustering method by looking at the values of the aforementioned indices as predictors of age-groups. We plotted the points returned by the clustering method with their centroids and thoroughly analyzed their age values. The centroids were almost uniformly scattered on the age axis with the step of approximately 20 years (see Table 2). The centroid coordinates were calculated as the average value of all the points in the corresponding cluster; therefore, we used the obtained age granulation to compose our groups. In such a manner, we segregated four groups in our sample: Adolescent ∈ [0, 20), Young adults ∈ [20, 40), Midlife adults ∈ [40, 60), and Older adults ≥60. With this division in Table 2, we presented the number of points correctly identified by the clustering method (performance column) vs. misclassified points (misclassified column). The best performance was obtained on the ISCA. Only five cases of Midlife adults were misclassified as Young adults. So, the ISCA index reflects the age-related psycho-physiological changes reliably.

We built a pairwise distribution of each proposed index with age (see Figures 6, 7). The linear horizontal trendlines with a 95% CI for the linear regression model estimates represent tendencies toward sustaining a balance between cognitive functions corresponding to diverse interrelated domains.

To compare the distribution of the indices over the age-groups, we applied statistical significance tests. Because none of our indices data underwent Gaussian distribution according to the Shapiro-Wilk test for normality (p < 0.05), we utilized nonparametric statistics. First, we checked the hypothesis that the population medians of all the groups were equal to the Kruskal-Wallis test, which showed significant changes in the distribution of the four groups (p < 0.05).

To study which groups differed in their medians, we ran post hoc Dunn test with the step-down method. We used Bonferroni adjustments (Holm's step-down procedure) to control the family-wise error rate. Regarding the indices values, only the median of the Adolescents group differed from those of the other three groups in the indices values. The three remaining groups shared a similar distribution pattern (p > 0.05). After a period of neurodevelopmental changes and maturation, the indices preserve almost constant values with a slight trend toward functional decline.

In the RMO test, however, there is another tendency. The RMO_mean values pairwise comparisons show that the Midlife adults group median significantly differs from the remaining three age subsamples (p < 0.01). No general trend in age-related changes of this dependent variable is observed (see Figure 8).

Finally, we investigated the variances of the proposed indices and selected psychophysiological attribute values for different groups. Levene's test reveals no significant changes (p > 0.05) in the variances among age-groups for the aforementioned values. Homoscedasticity also supports our assumption about stable linear dependency between the observed features and age.

Our idea was to inspect possible relationship of ISD, ISDA, and ISCA both with the general psychophysiological status of the examinee and the age of the individual. For getting an insight into this, we resorted to machine learning approach.

To preselect and justify the age group ranges, we resorted to a heuristic approach and clustering. Then, we did a review to elucidate the biological changes that may underpin the choice of such subcohorts. The group boundaries correspond to the time points between the end of neurodevelopment, the appearance, and the acceleration of the cognitive decline. The following observations of other authors evidence this.

Healthy educated adults start developing age-related cognitive changes in their 20 and 30s (Salthouse, 2009). Reasonably, the age period before 20s is the time of massive neurodevelopment, acquiring skills, and possessing knowledge and intelligence.

In the next 20 years of life, cognitive functions undergo conflicting changes. Basic physiological cognitive functions decline early in life. Because of this, young adults may exhibit early signs of fluid intelligence, memory, and processing speed. At the same time, crystallized intelligence increases (Zimprich and Mascherek, 2010). In a study on simple RT, authors reported an increase in consistency of response with age from 8 to about 30, after which a decline starts. They observed the fastest response at the age of over 20 years, but the most consistent responding in terms of time variance comes at over 30 years (Pierson and Montoye, 1958). A roughly similar timeline of changes was received in another study: The shortest RT comes in the examinees' mid-20s (Rabbitt, 1993).

The summary volume of the brain white matter rises up until the early middle-age adulthood (aged over 35 year) (Ferreira et al., 2014; Nilsson et al., 2014). Then comes a period of stability in terms of WM volume and cognitive performance (Ferreira et al., 2014; Nilsson et al., 2014). At this period of life, neurocognitive slowing may impair cognitive abilities of midlife adults. Neuroplasticity stimulated by physical and mental exercising may attenuate the changes and ameliorate the cognitive status (Bauermeister and Bunce, 2016; Haynes et al., 2017). Nonetheless, cognitive decline is already evident in the middle-aged population (Singh-Manoux et al., 2012). But its more accurate onset remains an issue of debate (Singh-Manoux et al., 2012).

Accelerated cognitive decline starts only after the late middle age (55–60 years) (Ferreira et al., 2014; Nilsson et al., 2014). It is marked with a massive WM volume reduction, while gray matter volume follows a steady rate of reduction throughout life.

Our findings suggest that there are proportional age-related changes in the time estimates of inhibitory control in task switching (e.g., DMT) and information processing speed (e.g., SVMR_mean). We claim that this is a feature of normal brain aging because we did a cross-sectional study of the healthy participants. In accelerated brain aging, the pattern of the changes may differ, in which case the proposed indices may serve as a screening tool for detecting such conditions. To verify this, a comparison study of the cohort of patients with dementia is required.

We do not see a significant accuracy decline in the relative CI of the performance of RMO technique across the lifespan. This is consistent with the origins of the Cognitive Aging Theory according to which intelligence does not decline with age. At that time, this pushed neuroscientists to correct the intelligence test data for the speed of processing (Anderson and Craik, 2017). Analyses revealed that most of the age differences in the cognitive performance were attributable to a slower speed of executing a relevant operation (Salthouse and Fristoe, 1995). Our recent studies conducted with the same battery of tests also showed that the age-related decline in reaction time in the attention study test is the most prominent feature of neurocognitive decline.

Many studies of the speed of processing have failed to develop unconditioned criteria that could help to distinguish normal aging from abnormal brain changes. Age-related declines in measures of cognitive functioning have been known as relatively large. They begin in early adulthood, manifesting themselves in several different types of cognitive abilities (Salthouse, 2004). To gain a better insight into the issue, scientists should investigate the interrelated changes of the cognitive abilities. A study of the proportionality of changes in the interdependent cognitive domains may supply the missing evidence to build up a reliable screening test for patients with dementia.

We hypothesized that a single index (e.g., ISCA) may substitute the dependent variables of the battery of PTs. In this way, it may serve as a marker of the psychophysiological status. If so, machine learning algorithms are capable to calculate its value out of other PT results. The potential of PTs to predict the values of ISCA is illustrated in Table 5. The performance metrics are good; the proportion of MAE to the range of values is low (3.49 ± 0.14%). All the created regression models are quite reliable. Random Forest regressor showed the best performance (3.36%). The accuracy of the models is higher than that of the models that we built recently for predicting IES score out of PT data (3.36 – 3.77% vs. 3.37 – 5.15%) (Statsenko et al., 2020).

The predictive accuracy is almost equal in all the age-groups studied. In contrast to this, the performance of predicting IES varies regarding the age: It is maximal for Adolescents and a bit lower for Older adults (Statsenko et al., 2020). This reduces the reliability of IES and makes ISCA the most suitable index for assessing psychophysiological performance and comparing the results irrespective of the age of examinees. As it is hard to segregate between the normal vs. accelerated aging, the index which is not vulnerable to aging may improve the currently existing strategies for the early detection of dementia.