Healthy memory aging - the benefits of regular daily activities increase with age

Selection of daily activities

Feature selection analyses identified 17 daily activities out of 33 that were significantly associated with changes in cognition, for at least one age group, by at least one method. These 17 variables represented cognitively challenging, physical, or social activities. Table 1 lists the selected variables with corresponding p-values for each age category. Notably, no single daily activity was significantly associated with all age groups. Only one daily activity (use computer) was identified as associated with changes in memory in four age groups. Only four daily activities (do activities with grandchildren and do word games, meet up with children, and speak on phone with other family members) were associated with changes in memory in two age groups. The rest of daily activities were significantly associated with changes in memory in only one age group.

Importance of individual daily activities

The relative importance of daily activities at predicting memory changes was quantified through sensitivity analyses for each age group. Figure 1 shows the relative importance of individual daily activities for each age group. No clear pattern of age-related changes in the relative importance of individual daily activities for predicting cognition changes was identified.

Figure 1. Relative importance of daily activities for predicting changes in memory, as estimated from sensitivity analysis. Error bars represent standard errors across repetitions in the sensitivity analysis.

The combined effect of daily activities on memory

The relative importance of daily activities from Figure 1 were combined for each age group. Figure 2 shows the relative importance of these combined daily activities compared with education and baseline cognition for each age group, demonstrating their power in predicting memory changes. The relative importance of the combined daily activities was computed as an arithmetic sum of the individual relative importance. Although finding regular patterns in the predictive power of individual daily activities was elusive, clear trends could be observed by considering their combined effects.

Figure 2. The relative importance of all life activities, education, and baseline memory for predicting changes in memory, as estimated from sensitivity analysis. In contrast to Figure 1, we show the combined importance of life activities (total sum). Note that on average age-related changes are characterized as an increasing trend for the combined effect of life activities on changes in cognitive health, whereas effects of education and baseline cognition are decreasing.

The importance of daily activities increases with age

As can be seen from Figure 2, the relative importance of combined daily activities on memory increased with age, which was confirmed by ANOVA (p < 0.001). Furthermore, pairwise comparisons confirmed significant differences in the relative importance of combined daily activities between age groups with p < 0.0002 for all pairs. In contrast to the combined effects of daily activities, the importance of education and baseline cognition on memory decreased with age (see Figure 2), which was confirmed by ANOVA (p < 0.001). Pairwise comparisons also showed significant differences between age groups (p < 0.001) except the 75–79 and 80–84 age groups for baseline cognition and 80–84 and 85–89 age groups for education were not significantly different.

Future research and study limitations

While the results of this study are encouraging, the findings are based on a statistical model that uses historical observations relating to daily activities, education and cognition baseline as predicting variables. In addition, the study had two other significant limitations: i) The data set did not include other potentially influential factors such as diet or socioeconomic factors. ii) Causal relationships were not considered. For example, part of the memory baseline may be explained by education, which may ultimately affect the quantitative assessment of relative importance. Similarly, engagement in one daily activity may be directly affected by the engagement in another activity. This relationship between daily activities needs to be understood further to develop causal relationship models in the future. Despite these limitations, and because the considered set of daily activities is unlikely to be highly interconnected, we feel that the results of this present study are valid, but additional studies need to be conducted to confirm our findings substantially.

Data source

Data obtained through interviews and surveys were taken from the 2008 and 2014 releases of the Health and Retirement Study (HRS), a longitudinal cohort study on health, retirement, and aging [19]. In particular, the study contained information from individuals with a broad range of ages on cognition and daily activities.

Sample selection

Three thousand five hundred seventy participants were selected based on: i) aged between 65 and 89 years (2014 release); ii) education record (2014 release); iii) memory assessment (2008 and 2014 release); iv) at least one daily activity (2014 release; see next section); and v) having gone through typical testing procedures (i.e., no interruptions, or problems with hearing). The participants were then reduced to 3,210 based by removing those with more than 15% missing data across all variables. Any remaining missing data were replaced with the median across participants, separately for each variable. The selected 3,210 participants were then grouped into five age ranges: 65–69 years (757); 70–74 years (840); 75–79 years (800); 80–84 years (518); and 85–89 years (295). It is worth noting that the upper age limit for the third age category corresponded to the USA’s life expectancy, which is about 79 years of age.

Measures

Data analyses

Taking into account education, machine learning was used to predict age-related changes in memory from participation in daily activities. Given the growing body of evidence indicating non-linear relationships between daily activities and physical health outcomes, no assumptions were made regarding the relationship between daily activities and potential cognitive outcomes. To that end, a two-step approach was followed to identify and quantify the relationship between daily activities and memory function.

Step 1. Feature selection to identify relevant predictive variables

In step one, a combination of non-linear techniques (Distance Correlation [28] and Random Forest [29]) was used to identify the most relevant variables with the goal of dimensionality reduction and support vector regression (SVR) for constructing predictive models. It should be noted that the use of SVR was not to focus on predictive accuracy but rather to investigate the importance of each predicting variable in explaining the target variable and exploring how a variable’s importance changes as a function of age.

Still, we estimated the predictive accuracy by randomly splitting dataset into training and testing subsets, with 80% used for training, and 20% used for testing for each age category, and further computing the coefficient of determination for the testing set. We repeated this estimation for 100 times, with mean coefficient of determination being 0.51. Non-linear techniques were considered most appropriate due to their versatile ability to capture functional relationships between variables of interest and that linearity between variables could not be assumed. Besides, non-linear techniques are better at accurately capturing strong associations between non-linearly dependent variables, while retaining the ability to capture strong associations between linearly dependent variables [30].

Distance Correlation is a measure of dependence between two random variables, not necessarily linearly related, which is zero only when variables are independent [28]. The distance correlation between memory changes and the daily activities was computed separately for each age group. The significance of correlations was tested with a bootstrap procedure, based on ~300 samples, as implemented in the R statistical toolbox ‘Energy’ [31].

Random forest is an ensemble learning algorithm that is an extension of decision trees. An algorithm progressively learns to predict the value of a target variable based on several input variables. Random forests are a way of averaging multiple decision trees, trained on subsamples of the same training data, to reduce the variance of the trained model [29]. One advantage of using random forests is the ability to identify, as a batch, groups of predicting variables that are related to the target variable. Random forest regression was used to predict changes in memory based on daily activities, education, and memory baseline for each age group. Each feature’s statistical significance was estimated for each age group using a permutation test based on ~300 samples, as implemented in the R package ‘pRF’ [32]. For further analysis, a subset of daily activities, which were found to be associated with memory changes at the 95% confidence interval (at least in one age category and by one method), was identified.

Step 2. Sensitivity analysis to quantify the relative importance of daily activities

The next step was to predict changes in memory using a sub-set of daily activities, which were identified as significant in the previous analysis, as well as education level and baseline cognition from the 2008 HRS release. Each age category was analyzed separately to quantify the relative importance of each predicting variable in its ability to explain the variability of the target variable. This procedure was performed using an SVR (based on a Gaussian kernel with a 10-range grid search for hyper-parameter optimization [33]) in combination with sensitivity analysis. Sensitivity analysis is based on the idea that changes in a relevant feature should yield substantial changes in the target variable. Features (variables) with low sensitivity are considered less important, whereas those with high sensitivity are considered more valuable. A random selection of several training sub-samples of the data with the corresponding responses of the target variable, taking into account the physiological well-being of an individual, was implemented in the R statistical toolbox ‘rminer’ [34]. The measure of importance was defined as a relative measure of dispersion in the target variable caused by changes in a given feature. Therefore, the importance of different features represents a unified measure that can be subsequently summed up.

The sensitivity analysis was repeated 200 times for each age group. As a result, each predicting variable was associated with a distribution of its important values, describing its ability to cause changes in the predicting variable. Analysis of variance (ANOVA - a generalization of the two-sample t-test to more than two groups with the null hypothesis that the group means are equal) was applied to test for significance of the differences in a variable’s importance between age groups. To further compare the variable’s importance between age groups, a series of two-sample Tukey Honest Significant Differences tests [35] was performed on a pairwise basis.

Consent for publication

Corresponding Author has the right to grant on behalf of all authors and does grant on behalf of all authors, an exclusive licence (or nonexclusive for government employees) on a worldwide basis to the BMJ Publishing Group Ltd to permit this article (if accepted) to be published in BMJ editions and any other BMJPGL products and sublicences such use and exploit all subsidiary rights, as set out in our licence.