Saturday, March 20, 2021

Increased subjective age (personal aging rate perception) increases mortality very much; biggest factors in SubjAge are physical health and expections of sex life in 10 years

PsychoAge and SubjAge: development of deep markers of psychological and subjective age using artificial intelligence. Alex Zhavoronkov et al. Aging, Volume 12, Issue 23 pp 23548—23577, December 8, 2020. https://doi.org/10.18632/aging.202344

Rolf Degen's take: Higher subjective age doubles mortality risk, and one of the top concomitants of subjective age is how satisfying people expect their sex life to be in 10 years’ time

Abstract

Aging clocks that accurately predict human age based on various biodata types are among the most important recent advances in biogerontology. Since 2016 multiple deep learning solutions have been created to interpret facial photos, omics data, and clinical blood parameters in the context of aging. Some of them have been patented to be used in commercial settings. However, psychological changes occurring throughout the human lifespan have been overlooked in the field of “deep aging clocks”.

In this paper, we present two deep learning predictors trained on social and behavioral data from Midlife in the United States (MIDUS) study: (a) PsychoAge, which predicts chronological age, and (b) SubjAge, which describes personal aging rate perception. Using 50 distinct features from the MIDUS dataset these models have achieved a mean absolute error of 6.7 years for chronological age and 7.3 years for subjective age. We also show that both PsychoAge and SubjAge are predictive of all-cause mortality risk, with SubjAge being a more significant risk factor.

Both clocks contain actionable features that can be modified using social and behavioral interventions, which enables a variety of aging-related psychology experiment designs. The features used in these clocks are interpretable by human experts and may prove to be useful in shifting personal perception of aging towards a mindset that promotes productive and healthy behaviors.

Discussion

In this article, we present two novel aging clocks created within the deep learning paradigm — PsychoAge and SubjAge. Both these clocks use the same set of 50 psychosocial features to estimate human chronological age and subjective age, respectively. These clocks showed superior performance during CV in MIDUS 1 (MAEPsychoAge= 6.70 years; MAESubjAge= 7.32 years) and were verified in two other large data sets — MIDUS 2 and MIDUS Refresher (Table 1). In terms of epsilon accuracy, PsychoAge reached a score of 0.78 in MIDUS 1, and SubjAge — 0.74.

Having trained and verified the final models, we aimed to understand how PsychoAge and SubjAge see human aging and what features they pay the most attention to. With a tandem PFI-DFS approach we ranked all features according to their relative importance. Top-5 important features in both clocks were associated with health conditions (e.g. headache frequency) and relationship status (marital status, expectations from sex life in 10 years). Less significant features greatly differ in their relative importance for SubjAge and PsychoAge predictions. For example, top-20 PsychoAge features contain only one personality trait — neuroticism. Meanwhile, the only personality traits encountered among top-20 SubjAge features are — extraversion and openness.

These three personality traits, along with conscientiousness and agreeableness form “the big five traits”, which are commonly used in practice and scientific research to describe the human mental state landscape. High neuroticism is characteristic of emotional instability and common mental disorders, such as mood disorders, anxiety, and substance use disorders. Openness and extraversion, on the other hand, are considered more balanced traits, although their abnormally low scores are also related to social phobia and agoraphobia [27]. Positive orientation, seeking warmth, social interaction, and emotional stability may play an important role in psychological aging.

We hypothesized that the human mind evolves throughout the lifespan, which results in some traits, beliefs, or priorities shifting — not always in unison or at the same speed. At certain life stages, career-related priorities may rise, while at others they may fade and be replaced by different priorities. These lifelong progressions of the psyche eventually get recognized by the neural networks we constructed to let them build an image of psychological aging.

This idea of human mind progression is described in much more detail in the review of SST by Laura Carstensen. SST suggests that younger people are more goal-oriented, interested to obtain new knowledge and skills, while older people tend to value emotionally meaningful goals more.

To identify the psychosocial features that change while a person advances from one age group to another we trained separate DNNs on MIDUS 1 samples from three age groups (25-39, 40-64, 65-75 years). First, we defined the psychological aging core — variables that remain highly important (top-25) across all age groups (Table 2). The core contained not only strictly psychological features, however. To illustrate, marital status, hypertension medication, headaches, and body mass index were among the seven core features required for accurate chronological age prediction. Interestingly, neuroticism score also belonged to the same psychological aging core, as well as seeing the community as a source of comfort. Psychological traits within the subjective aging core contained aspirations scale, extraversion, openness, positive reappraisal prevalence, and two career-related variables — effort put into work now and work expectations in 10 years. In contrast to the first psychological core, which contained few psychological traits, the subjective core consisted almost exclusively of psychological features.

This highlights an important distinction between aging per se (as judged by PsychoAge) and our perception of it (as judged by SubjAge): subjective aging is mostly dependent on internal causes.

We also explored the uniquely important features for each age group — features that emerged only in one top-25 set. Since these features were recognized as important only in these groups, it may be assumed that they shift the most markedly during the corresponding life periods. To illustrate, young adults were not the only age group who responded affirmatively to the statement “Forceful describes you well”, but rather many of these people went through a transformation that affected their forcefulness. Detecting such a change was essential for a predictor to accurately predict whether a person was at the beginning or the end of this phase of life.

On their own, DNNs are unable to tell generational and age-related changes apart or tell the difference between pro-longevity and progeroid features. Thus, the results of the feature importance analysis should always be cautiously inspected and verified in more rigorous settings. Still, feature importance analysis is a powerful tool for hypothesis generation and the verification of overall biological relevance.

While neuroticism was identified as a part of the psychological aging core, it was also a uniquely important subjective aging feature in the elderly MIDUS 1 subsample. This may be interpreted as neuroticism progressing unnoticed by an individual until old age when it starts to affect the perception of age. Previous studies identified that neuroticism tends to cause low emotional differentiation, anxiety, and depression in old people [2829].

Other personality traits rendered important for subjective age estimation in the elderly were optimism, being outgoing, and content with life in general. These results indicate that these might be top-priority features to focus on while developing policies aimed to involve the elderly in social life. Several studies have shown the importance of a social and productive lifestyle during aging [3031]. psychologically important and active events may protect against aging diseases, such as dementia [32].

After establishing which variables are important in absolute terms, we aimed to measure the models’ response to changes in their values. Using mixed-effects linear models, we explored the monotonic trends between 50 variables, PsychoAge and SubjAge predictions (Supplementary Figure 1 and Supplementary Figure 2).

Once again, neuroticism showed unique behavior. Contrary to the other big five traits, neuroticism score was associated with higher SubjAge and lower PsychoAge. More specifically, people within the same PsychoAge group could have >5 years of SubjAge difference due to differences in neuroticism score alone. This verifies our previous conjecture that neuroticism is a key marker of subjective aging and may be used as a sensitive measure of emotional states and late-life depressive symptoms.

Other big five traits also had significantly large effects on both PsychoAge and SubjAge. For example, a person with the bottom openness score would feel 7.2 years older than their PsychoAge counterpart with the top score. In the meantime, a person with the bottom openness score would be 5.3 years younger, as measured by PsychoAge, than their SubjAge counterpart with the top score. Similar tendencies could be observed for most other personality traits, thus building a strong case for SST.

Interestingly, personal opinion on when middle age starts and ends was significantly associated with higher PsychoAge but does not affect SubjAge. We hypothesize that this is an indication of “time dilation” associated with aging. As people get older, they place “middle age” higher and higher, as if their lifetime dilates, while younger participants may have stereotypes about aging and place “middle age” lower. An excellent study on the topic of perception of age stereotypes and self-perception of aging has been written by Hummert [33].

Among the health-related features, the distinction between internal and external health locus of control is of utmost interest. Health locus of control is a set of personal beliefs and experiences that determine whether a person takes responsibility for their health (internal locus) or considers it to be outside of their power, fully dependent on external factors (external locus). Internal locus of control is associated with a problem-solving mindset, while external locus is tied to depression, anxiety, and suicidal thoughts, as well as maladaptive behaviors [34]. We demonstrated that external locus of health control is a rare feature that demonstrated a linearly positive effect on both PsychoAge and SubjAge. It was the only feature to offer no payoff in at least one aging dimension, except for “Taking prescription medications for blood pressure”. Internal control, per contra, did not display concordant linearly negative effect on. Instead, it decreased SubjAge and decreased PsychoAge, just as most other health-related variables.

While the external locus of control was a senopositive (higher values increase age predictions) feature in both aging dimensions, some features were identified as double senonegative (higher values decrease age predictions). Increasing values for the variables from the relationships category were associated with lower PsychoAge and SubjAge, thus favoring single people content with their sex life, who expected to remain sexually active in 10 years. In this case, it is difficult to conclude the cause-effect relation between psychological aging and sexuality. Is reduced libido a precondition to becoming subjectively old? Or does feeling old due to other factors make people less interested in the sexual aspect of life? Can more satisfying sex life prolong healthy longevity, or does PsychoAge simply see higher sex drive as a feature more frequently encounter in the youth? More thorough research is required to answer these questions as well as similar questions concerning other variables.

Although the effect of most variables on PsychoAge and SubjAge was shown to be discordant, the magnitude of their effects on these two measures of psychological aging is not equal. Since the target variable in the mixed-effects model is expressed in years, Table 3 can be used to approximate how a shift in a psycho-social parameter will affect PsychoAge or SubjAge, and which one of them will change more. For example, the variable “Rate health in 10 years” is a survey question that measures health expectation on a scale from 0 to 10, from worst to best. Each increment increases PsychoAge by 0.5 years but also decreases SubjAge by 1.0 years. This yields an “exchange rate” of 2 subjective years lost per 1 chronological age gained. Other features have their own exchange rates, which may be manipulated to accumulate “net profit” in both SubjAge and PsychoAge dimensions.

Other directional feature analysis methods may be more appropriate for navigating the psychological aging landscape since linear mixed effect models operate based on multiple assumptions and simplifications. More specifically, they treat all features independently and approximate the complex interrelations between PsychoAge and SubjAge that may be in place with a random intercept. Accumulated local effects or more sophisticated Shapley value analysis may handle the convoluted feature interrelations more efficiently.

To further validate PsychoAge and SubjAge we tested their prediction errors (delta) as all-cause mortality risk factors (Figure 6). SubjAge delta was proven to be a more powerful risk factor than PsychoAge. More specifically, the SubjAge delta beyond ±5 years was associated with roughly doubling or halving the mortality rate.

We also tested the 50 psychosocial markers of aging as risk factors. We identified significant mortality risks associated with certain factors among (i) health features (“Health compared to others your age”, “Rate current health”, “Shortness of breath while walking up a slight hill”), (ii) personality traits (“Conscientiousness personality trait”, “Agency personality trait”), (iii) psychological beliefs (“Live for today”, “Positive reappraisal”, “Lower aspirations”), (iv) well-being (“Satisfied with life at present”), and, (v) demographic factors (“Chronological age”) (Figure 7).

A problem frequently encountered even by psychologists is obtaining sufficiently detailed information about their patients while keeping the data collection process as short as possible to avoid survey fatigue. In this work, we propose a solution to this survey length-descriptiveness balance problem based on modern deep learning and biogerontological methods. The solution is a relatively short list of features that are both modifiable and provably important in the context of aging.

Some studies show that family history is a source of numerous highly important aging-related features [35]. For example, having long-lived parents and grandparents is strongly correlated with a longer lifespan. However, such factors are not easily modified, especially if the grandparents are no longer alive. Therefore, in this study, we have deliberately limited questions on non-modifiable historical factors to give our surveys more practical value. We demonstrated that variables related to health and closer personal relationships play a crucial role in chronological and subjective age prediction. Furthermore, images about life changes, for instance, when females or males enter middle age, demonstrate a strong predictive power. We suggest that modifying the behavior and the mindset via these variables may be a promising therapeutic concept.

The factors comprising the developed aging clocks can be used to create individual behavioral therapies that would make them feel and actually become biologically younger. For example, PsychoAge can be used to quantify the beneficial effects of daily vigorous-intensity activity on their rate of aging. SubjAge, in its turn, can be used to quantify the beneficial effects of physical activity on personal perception of age.

Focusing on such modifiable age-related features while being able to score lifestyle choices numerically offers interesting opportunities to both professional therapists and individuals seeking self-improvement. We believe that the described approach has a high potential to increase longevity conscience on a population level.

When the category of close relationships is considered, people with access to PsychoAge and SubjAge may choose to develop stronger bonds, get married, or stay married out of an egoistic incentive to prolong their healthspan. The beneficial effect of close relationships and intimacy on health was previously shown in multiple studies, and we believe that drawing people’s attention to such physical-mental health connections should not be neglected [36].

Extracting actionable items from human biological profiles, such as transcriptomic or proteomic profiles, is an actively researched subject. The profiles associated with human psychology can also be subjected to similar workflows to devise personal behavioral therapy plans. In this study, we have demonstrated how the combination of deep learning and aging clocks can be used to create psychological surveys that promote longevity consciousness and personal improvement. These tools and methods could be applied in a wide range of research areas, including psychiatry, longevity, psychology, and psychophysiology for the greater good of society.

No comments:

Post a Comment