Explainable Artificial Intelligence on Social Support, Poor Self-Rated Health and Dementia

1. Introduction

Dementia can be defined as “deterioration in cognitive function beyond usual consequences of biological ageing” [1]. It comes from various conditions having major or minor effects on the brain. Its prevalence is more than 55 million in the world now, whereas its new incidences are almost 10 million on the globe every year [1]. Its global prevalence is expected to reach 153 million for the year 2050, while its global female-to-male ratio is projected to stay stable between 1.69 in 2019 and 1.67 in 2050 [2]. This global trend agrees with its local counterpart in Korea. The prevalence of elders’ dementia in the nation registered a growth from 5.9% to 7.3% (588,000) during 2015-2019 [3]. The direct, indirect and total cost of dementia in the nation expanded by 56%, 36% and 46% during 2015-2019, i.e., 1400 to 2191, 1485 to 2026, and 2885 to 4218 USD million, respectively [3].

On the other hand, self-rated health can be conceptualized as “the overall perception of one’s own health” [4,5,6,7,8,9]. It is usually assessed in a single item, where participants rate their current health. It is considered to be an effective indicator of health status on its own right but it is expected to affect other health outcomes such as mortality based on a variety of indirect channels, e.g., socioeconomic conditions, biological, physical and mental functions, as well as health-related behaviors [4,5,6,7,8,9]. For example, a previous study on self-rated health in Japan and Korea [9] reported positive effects from socioeconomic status, daily activity and physical exercise, as well as negative impacts from chronic disease, overweight/obesity and smoking. In a similar context, existing literature confirms a strong association between poor self-rated health and the risk of dementia [10,11,12,13,14].

However, little study has been done on social support, dementia and self-rated health. A previous study [15] made a rare contribution for this line of research to use a machine learning model for testing: whether social determinants are major determinants of the association among cerebrovascular disease, hearing loss and cognitive impairment in a middle-aged or older population (hypothesis 1); and whether the association among the three diseases is very strong in the middle-aged or older population (hypothesis 2). Data came from the Korean Longitudinal Study of Aging (2014-2016), with 6060 participants aged ≥ 53 years. The findings of this study supported the two hypotheses, highlighting the importance of preventive measures, family support, socioeconomic status and friendship activity for managing the three diseases. But this study did not examine whether there exists a variation among different diseases regarding the importance of social support. In this vein, this study evaluates the following hypotheses based on machine learning models:

Hypothesis 1.

The predictors of social support are major predictors of dementia, poor self-rated health and their comorbidity in the middle-aged or old

Hypothesis 2.

There exist some variations among dementia, poor self-rated health and their comorbidity in the middle-aged or old in terms of the importance of social support

2. Methods

2.1. Participants and Variables

The data source of this study was the Korean Longitudinal Study of Ageing (KLoSA) (2016-2018). This study did not require either the approval of the ethics committee or the informed consent of human subjects given that (1) data were publicly available (https://survey.keis.or.kr/eng/klosa/klosa01.jsp) and (2) data were de-identified (patient anonymity was preserved). This study used the recoded version of data for immediate analysis (“light version”). But there were still missing values in some independent variables hence they were replaced with their median values (median imputation). The final sample of this study consisted of 5961 subjects aged 56 or more. The dependent variables were: (1) poor subjective or self-rated health (SRH) with 5 categories (very good, good, middle, poor, very poor); (2) dementia (no, yes); and (3) their comorbidity with 4 categories (SRH very good, good or middle, dementia no; SRH very good, good or middle, dementia yes; SRH poor or very poor, dementia no; SRH poor or very poor, dementia yes). The KLoSA questions on dementia and poor self-rated health in 2016 and 2018 were “Since the last survey, have you ever been diagnosed by a doctor dementia? 1. Yes. 5. No.” and “How do you evaluate your health condition? 1. Very Good. 2. Good. 3. Middle. 4. Poor. 5. Very Poor.”.

The independent variables were the following 49 predictors in 2016: (1) 5-category SRH (very good, good, middle, poor, very poor); (2) demographic information including age, gender, marital status (married, separated, divorced, widowed, unmarried); (3) socioeconomic conditions such as educational level (elementary school or below, junior high school, senior high school, college or above), income (monthly, normalized between 0 and 1), health insurance (Medicare, Medicaid), economic activity (employed, unemployed); (4) social support, that is, friendship meeting, social activities (religious, friendship, leisure, association), grandchildren (alive, raised, taken care), children (alive, cohabiting, meeting, contact, proximity), brothers/sisters (alive, cohabiting), parents (alive, cohabiting, meeting, contact, proximity); (5) health-related information, i.e., body mass index, smoker (non, former, current), drinker (non, former, current); and (6) other determinants including region (big urban, small urban, rural), religion (non, Protestant, Catholic, Buddhist, Won-Buddhist, other), residential type (apartment, other), subjective class (high-A, high-B, middle-A, middle-B, low-A, low-B), life satisfaction - health (0-100), life satisfaction - economic (0-100) and life satisfaction - overall (0-100).

2.2. Analysis

Logistic regression and the random forest were compared for the prediction of dementia, SRH and their comorbidity. For the random forest, the number of trees was 100, the criterion of split was GINI and the max of depth was not determined. Data on 5961 participants were divided into training and validation sets with a 75:25 ratio (4471 vs. 1490 observations). Criteria for validating the models trained were (1) accuracy, a ratio of correct predictions among 1490 observations, and (2) the area under the receiver-operating-characteristic curve (AUC), the plot of sensitivity over 1 - specificity. Here, the AUC measures how much sensitivity can be secured in case the threshold of sensitivity increases from 0 to 1 (that is, specificity increases from 0 to 1). Variable importance from the random forest, a GINI gap between a complete model and a model excluding a certain variable, was used for testing the two hypotheses of this study. The evaluation of Hypothesis 1 was based on whether the predictors of social support are top-20 predictors of dementia, SRH and their comorbidity across board. The evaluation of Hypothesis 2 was based on whether some variations are found among dementia, SRH and their comorbidity in terms of the variable importance ranking of social support. Finally, the Shapley Additive Explanation (SHAP) values were calculated to analyze the direction of association between a major predictor and its dependent variable in the model (random forest). The SHAP value of a particular predictor for a particular observation measures a difference between what the model (the random forest) predicts for the probability of the dependent variable for the observation with and without the predictor (https://github.com/shap/shap). Python 3.52 (Centrum voor Wiskunde en Informatica, Amsterdam, Netherlands) was employed for the analysis on January 2023.

3. Results

Descriptive statistics for participants’ categorical and continuous variables are shown in Table 1 and Table 2, respectively. The proportions of those with SRH poor/very poor and dementia in 2018 were 28.7% (1712) and 1.5% (89), respectively. The proportion of those with either or both disease in 2018 was 28.8% (1717). The proportion of those with SRH poor/very poor in 2016 was 26.8% (1598). The average numbers of family members or the average frequencies of social activity in a month were 4 (grandchildren alive), 3 (friendship meeting), 3 (brothers/sisters alive), 3 (children alive), 2 (social activity - friendship) and 2 (children in weekly contact). In the case of median imputation for missing values, logistic regression and the random forest registered similar AUCs with the range of 0.80-0.87 (Table 3). These performance measures were higher than those in the case of piece-wise deletion for missing values, i.e., 0.27-0.84.

Based on random forest variable importance (Table 4), 18 out of top-20 predictors were identical for SRH, dementia and their comorbidity: previous SRH, life satisfaction - health, age, body mass index, income, chronic diseases, life satisfaction - economic, life satisfaction - overall, grandchildren alive, friendship meeting, brothers/sisters alive, class identity, children alive, social activity - friendship, education, children in weekly contact, religion and drinker. This result supports Hypothesis 1: The predictors of social support are major predictors of dementia, poor self-rated health and their comorbidity in the middle-aged or old. However, some variations were found among the three dependent variables, i.e., previous SRH for comorbidity (1st) and SRH (1st) compared to dementia (11th), grandchildren alive for SRH (6th) and dementia (4th) compared to comorbidity (9th), chronic diseases for comorbidity (6th) compared to SRH (10th) and dementia (10th). This finding supports Hypothesis 2: There exist some variations among dementia, poor self-rated health and their comorbidity in the middle-aged or old in terms of the importance of social support. In terms of SHAP, the probability of dementia is expected to decrease by 21% in case a predictor grandchildren alive is included to the model (Table 5 and Figure 1). There exists a strong negative association between the predictor grandchildren alive and the dependent variable dementia.

4. Discussion

Existing literature reports inconsistent results on social support and dementia [16,17,18]. A retrospective cohort study adopted 355 elderly Americans and linear regression to find that a positive association exists between perceived emotional support and the Repeatable Battery for the Assessment of Neuropsychological Status score (cognitive function) in women only [16]. A retrospective cohort study, which employed 5852 elderly Koreans and Cox proportional hazard regression, reached a similar conclusion with its U.S. counterpart, i.e., emotional support has its protective effect on incident dementia among women only [17]. On the contrary, a prospective cohort study employed 14,088 elderly Japanese and Cox proportional hazard regression to discover that the protective effect of family support on incident dementia exists among men only [18]. The SHAP results in Table 5 of this study show that there exists, in general, a strong negative association between the predictor social support and the dependent variable dementia among all participants. In terms of SHAP, for example, the probability of dementia is expected to decrease by 21% in case a predictor grandchildren alive is included to the model. Artificial intelligence is a data-driven approach hence more examination is needed for more conclusive findings on social support and dementia.

On the other hand, existing literature reports that poor self-rated health results in the increased risk of mortality [19,20]. In a retrospective cross-sectional study with 2819 breast cancer patients in the United States, poor self-rated health was found to have a positive association with cancer mortality with the hazard ratio of 3.05 [19]. In a retrospective cohort study with 710 elderly men in Europe, poor self-rated health was also discovered to have a positive relationship with cancer mortality with the hazard ratio of 2.41 [20]. In other words, poor self-rated health requests due attention to its fatal role in the risks of mortality. Moreover, poor self-rated health in the previous period was the most important predictor for the comorbidity of dementia and poor self-rated health in the current period based on the findings of this study. Based on random forest variable importance (Table 4) in this study, 18 out of top-20 predictors were identical for SRH, dementia and their comorbidity, and six of these top predictors represented social support i.e., grandchildren alive, friendship meeting, brothers/sisters alive, children alive, social activity - friendship and children in weekly contact. Little literature is available and more investigation is needed on dynamic interactions among social support, poor self-rated health and dementia in a variety of study populations in different environments. The SHAP value can be considered to be the weighted average of all possible real-world scenarios [21] hence explainable artificial intelligence based on the SHAP including this study can be a very attractive option for this line research.

This study had some limitations. Firstly, this study had limited memory capacity and used a weak version of the longitudinal design with data in two waves. Poor self-rated health, dementia and their comorbidity in 2018 (Wave 7) served as the dependent variable of the models, while, the three diseases in 2016 (Wave 6) and the demographic, socioeconomic and health-related factors in 2016 (Wave 6) served as the independent variables of the models. Expanding memory capacity and employing a strong version of the longitudinal design with data in all seven waves would strengthen the performance of explainable artificial intelligence. Secondly, this study kept outliers to protect the sample size. Comparing different outlier management strategies would further the horizon of research on this topic. Finally, sub-group analysis would make a good contribution for this line of research, e.g., 55-64, 65-74 and 75+ in age.

In conclusion, this study highlights the importance of social support for the prediction of poor self-rated health, dementia and their comorbidity. Specifically, the promotion of family support and friendship activity for elders would be vital for the prevention and management of their self-rated health and dementia.