Development of frailty subtypes and their associated risk factors among the community-dwelling elderly population

Participant characteristics

Among all 5,341 elderly people surveyed, more than half were women (61.1%), and almost half were 60 to 70-years-old (46.8%). Most (56.9%) of the elderly people had received only primary education. The vast majority were married (74.3%), as shown in Table 1.

Latent class model fitting results

The study extracted one to seven potential class models, and the fitting results are shown in Table 2. As model classification increased from one to seven, the AIC, BIC, and aBIC continued to decrease, while the values of the Lo–Mendell–Rubin (LMR) test and the bootstrap-based likelihood ratio test (BLRT) reached significant levels (P < 0.01) when five categories were retained, and the LMR was no longer significant when the sixth classification was made. According to the LMR and BLRT indicators, the model with five sub-categories was significantly better than the model with four sub-categories, and the model with six sub-categories was not as well fitted as the model with five sub-categories. Therefore, the classification using five latent classes (Class 1, Class 2, Class 3, Class 4, Class 5) was finally selected.

Naming of latent classes

According to the conditional probability distribution differences and characteristics of 33 items (Supplementary Table 1), each latent class could be named. Compared with the other four classes, class 1 had the highest probability of answering “no” to almost all items, indicating that the group’s overall health was good, so class 1 was named “relatively healthy.” Conversely, class 2 had the highest probability of answering “yes” to most of the items, indicating that the group’s overall health status was poor, so class 2 was named “multi-frail”. The conditional probabilities of class 3, class 4 and class 5 answering “yes” to most of the items were between those of class 1 and class 2. Class 3 had a higher probability of performing poorly on items representing functional activities (items 16–21) and cognitive function (item 33), so class 3 was named “cognitive and functionally frail”. Class 4 had a higher probability of performing poorly on items representing mental state (items 26–31), so class 4 was named “psychologically frail”, and this group had the highest probability of physical pain and poor sleep. Class 5 was named “physiologically frail”, because the health problems of this group were mainly concentrated on the components representing general health status and symptoms.

The number of elderly people in the five classes was 2,817 (52.7%), 176 (3.3%), 508 (9.5%), 638 (11.9%) and 1,202 (22.5%), respectively. The FI score of all participants was calculated. The “relatively healthy” group had the lowest average FI score (0.07±0.04), the “multi-frail” group had the largest (0.61±0.15), and the average FI scores of the “cognitive and functionally frail” group, “psychologically frail” group and “physiologically frail” group were all between those of the above two groups.

Analysis of factors associated with different frailty subtypes

Multinomial logistic regression was used to evaluate the associations between the 12 baseline characteristics and the frailty subtypes. “Relatively healthy” was used as a reference class for the other four classes.

As shown in Table 3, the results showed that compared with people in the 60–70-year age group, the risk of having one of the four frailty subtypes in over 80-year-olds was higher. Elderly individuals who had not received education were also more likely to have one of the frailty subtypes than those who had received junior high school education or above. Those who were widowed, divorced or unmarried were more likely to be multi-frail, cognitive and functionally frail and psychologically frail compared to those who were married, and males had a lower risk of being cognitive and functionally frail and psychologically frail than females. Elderly people who had harmonious relationships with their family had a lower risk of being multi-frail than others; the fewer times the elderly were visited by their children, the more likely they were to be psychologically frail. The elderly who were dissatisfied with their housing had a higher risk of being cognitive and functionally frail, psychologically frail and physiologically frail, and those for whom pensions were their main source of income had a higher risk of being cognitive and functionally frail and psychologically frail than others. Individuals who exercised every day were less likely to be multi-frail and cognitive and functionally frail than those who never exercised, and those with a lower level of physical activity were more likely to have one of the four frailty subtypes. However, number of children and smoking status showed no significant effect on any of the frailty subtypes.

Limitations and strengths

The main strength of this study is our consideration of the heterogeneity of frailty in the analysis. We expanded the commonly used model of frailty deficit accumulation, combined categorical variables with latent class models, and explored the interactions between functional components using latent class analysis. Our discovery of four frailty subtypes verified the complexity and heterogeneity of frailty. We further found that several frailty subtypes had different associated risk factors, which implies that targeted management and care interventions are required. Another strength of this study was the choice of assessment tool, which was well suited to the study participants. Considering the geographical and cultural differences between China and other countries, when compared with the translated version of the foreign scale, the FI scale developed by the project team and used in this study was more suitable for the Chinese elderly population. Finally, the large sample, combined with the full range of domains of functioning made the current research and discoveries valuable.

Our research also inevitably had some limitations. First, LCA is a person-centered approach to identify unobserved groups of similar individuals [14]. The selected research population may have an impact on the final latent classes. Although we conducted tracked home visits and telephone interviews for elderly people who were not able to participate in the medical examination, there were inevitably some elderly people who refused to be interviewed for various reasons. It is unclear whether the absence of these individuals affected the final results. Second, our research was only an exploration of frailty subtypes, it did not provide an operational definition for each subtype, and the validity of these subtypes requires further testing. Finally, frailty is a dynamic process [40]; our cross-sectional analysis did not include frailty trajectories. In this regard, it would be beneficial to explore whether a frailty subtype eventually transfers to another frailty subtype and how the trajectory of individuals progresses.

Data and participants

We selected 21 communities (villages) in Dalang Town, Dongguan City, China, using a random cluster sampling method, and conducted cross-sectional surveys of all eligible seniors in these communities (villages). Elderly individuals aged 60 years and over and who were conscious were included; those diagnosed with Alzheimer's disease, a disability causing them to be bedridden, unable to communicate adequately, or unwilling to be investigated were excluded. The majority of participants were investigated during on-site physical examinations in Dalang Town from October 2017 to June 2019. In response to the national pension service policy, this medical examination was carried out free of charge for residents over 60 years old in all 28 communities (villages). We applied a questionnaire in the form of face-to-face interviews for the elderly who participated in the medical examination. In order to ensure the integrity of the sample, after completing the scheduled medical examinations, we immediately contacted all those elderly persons who did not present for the on-site physical examination. We conducted surveys face-to-face at these participants’ homes or via telephone.

A total of 5397 questionnaires were distributed. After excluding 56 unqualified questionnaires, 5341 samples were finally included in the analysis; the effective response rate was 98.96%.

Composition of the questionnaire

Frailty assessment tool

The FI scale developed by the project team at an early stage was used in this research, containing seven dimensions of a total of 33 health problems. General health status contained six items (e.g. Are you in poor health now? Has your health deteriorated compared to 1 year ago?) Activities of daily living contained nine items (e.g., In the past month, did you need help to complete the following activities? Bathing, Dressing, etc.). Functional activities contained six items (e.g. In the past month, did you need help to complete the following activities? Going up or down stairs, shopping, etc.). Symptoms contained four items (e.g. Have you had any physical pain in the past month? Is your vision impaired?). Mental state contained six items (e.g. Have you had the following feelings in the past week? Hard to concentrate; Feeling sad or depressed, etc.). Social support contained one item (Are you living alone?). For the above 32 questions, the respondent only needed to answer "Yes" or "No" according to their own circumstances. In addition, cognitive function was assessed using the Mini-Mental State Examination (MMSE). The investigator evaluated whether respondents’ cognitive function was poor based on their final scores. A score of 17 points or less was marked as "Yes", and a score of more than 17 points was marked as "No" [41, 42].

From our previous studies, the FI scale developed was based on a frailty index model. After debugging among the elderly population in China, the scale has been shown to fully reflect the frailty condition of Chinese elderly, and its reliability and validity has been proven to be good [25]. The number of health problems (answering "yes") of each elderly person was divided by the total number of 33 health problems, to obtain the FI score; the score ranged from 0 to 1, with a higher score indicating a higher level of frailty.

Baseline characteristics

A total of 12 variables of four aspects were included in the questionnaire. Sociodemographic characteristics: sex (male; female), age (60–; 70–; 80–), educational level (illiteracy; primary school; junior high school and above), marital status (married; widowed/divorced/ unmarried). Family relationship: number of children (0–1; 2–3; 4–), number of times children visit per month (0–2; 3–4; ≥5), relationship with family (harmonious; not harmonious). Economic status: whether pensions are the main source of income (yes; no), housing satisfaction (dissatisfied; general; satisfied). Behavioral factors: smoking status (yes; no), exercise frequency (every day; less than five times a week; less than 2 times a week; not exercising), and physical activity level (low level; medium level; high level), which was measured using the short version of the International Physical Activity Questionnaire (IPAQ).

Statistical analysis

Data preparation and descriptive analysis

EpiData 3.1 software (The EpiData Association, Odense, Denmark) was used to establish the database; Microsoft Excel 2013 version was used for data collation and error correction. We generated descriptive statistics for the total sample, giving frequencies and percentages for the sociodemographic variables.

Latent class analysis (LCA)

First, we used latent class analysis (LCA) to explore the latent classes of frailty. LCA is a statistical technique for exploring the categorical latent variables behind the statistically related categorical observed variables; it combines latent variable theory with categorical variables. Based on the observed variables, the purpose of LCA is to find the best class solution, that is, to explain the association among a set of observed variables with the least number of latent classes, and further to cluster similar individuals [43]. The observed variables we used in the LCA were the 33 variables mentioned above for the frailty assessment.

In this study, the maximum likelihood (ML) method was used for parameter estimation, and the expectation-maximum (EM) method was used in the iterative process. We fitted 1–7 latent class models. For model evaluation, the indicators included the Akaike information criterion (AIC), Bayesian information criterion (BIC), and adjusted BIC (aBIC). In general, the smaller the numerical values of these indicators, the better the model fits. The entropy index was used to evaluate the accuracy of the classification. A value greater than or equal to 0.80 in this index shows an accuracy of classification of greater than 90% [44]. The Lo–Mendell–Rubin (LMR) and the bootstrap-based likelihood ratio test (BLRT) were used to compare the differences in fit of the nested latent class models; if the P values of the two values achieved a significant level (P < 0.01), this indicated that the model with K classes was significantly better than the model with K–1 classes [45]. We determined the final classification on the basis of comprehensive consideration of the above indicators. The LCA analysis was done using Mplus7.4.

Naming of latent classes

First, based on the final classes selected, the probability of each latent class and the conditional probability of each observed variable under the latent classes were estimated by the EM method. Next, according to the differences in the conditional probability distribution and characteristics of the observed variables, we interpreted and named each latent class. Finally, we calculated the FI score of each individual and described the FI score distribution $(\overline{x}\pm s)$ for each class.

Analysis of factors associated with different frailty subtypes

In order to explore the factors associated with different frailty subtypes (latent classes), we conducted a multinomial logistic regression with frailty type of the elderly as the dependent variable and 12 possible associated variables as the independent variables. We used one category of each independent variable as a reference category to obtain the adjusted odds ratios (OR) and their 95% confidence intervals (CI) for the other categories; P ≤ 0.05 was considered statistically significant. This analysis was conducted using SPSS software (IBM SPSS Statistics 25.0, IBM Corporation).

Ethics approval

Informed written consent was obtained from all participants included in this study. The study was approved by the Research Ethics Committee of Guangdong Medical University (YJYS2018046).