Latent profiles and network dynamics of psychological frailty in older adults: a computational simulation study

Abstract

Background

Previous studies on psychological frailty in older adults have typically examined individual components or assumed population homogeneity, often overlooking latent heterogeneity and complex symptom interrelations. This study aimed to identify latent profiles of psychological frailty and to integrate computational simulation with network analysis to explore patterns of symptom associations.

Methods

This cross-sectional study included 4,735 adults aged 65 years or older. Psychological distress, cognitive function, physical vulnerability, and memory decline were assessed using established measures from the China Health and Retirement Longitudinal Study. Latent profile analysis was conducted in Mplus to identify subgroups, followed by Gaussian graphical model network analysis in R to examine symptom interrelations within each profile. Additionally, an in silico simulation approach using the NodeIdentifyR algorithm was applied to estimate changes in overall network burden under hypothetical symptom perturbations.

Results

Three latent profiles were identified: Relatively Healthy (57.8%), Moderate Distress (32.8%), and Severe Cognitive Deterioration (9.4%). Network structures differed significantly between the Moderate Distress and Severe Cognitive Deterioration profiles. The former showed higher centrality of orientation-related symptoms, whereas the latter was characterized by greater centrality of affective symptoms. In simulation analyses, hypothetical increases in powerlessness, depression, attentional problems, and functional limitations were associated with the largest estimated increases in overall network burden, whereas symptom-response patterns differed across latent profiles.

Conclusion

This cross-sectional study identified distinct latent profiles of psychological frailty and demonstrated differences in symptom network organization across profiles. The simulation findings indicate that certain symptoms may be associated with greater model-based changes in network activation. These results provide an exploratory framework for understanding heterogeneity in psychological frailty, but require validation in longitudinal and intervention studies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12877-026-08070-w.

Introduction

With the accelerating pace of global population aging, maintaining psychological well-being and functional independence among older adults has become a major public health concern. Traditionally, frailty has been conceptualized as a physical syndrome characterized by reduced strength, endurance, and physiological reserves. However, recent research has increasingly recognized frailty as a multidimensional construct encompassing not only physical but also psychological, cognitive, and social domains [1, 2]. This holistic view more accurately reflects the complexity of the aging process and has profound implications for comprehensive assessment and intervention strategies [3–5]. Initially introduced in the 1990s to explain heterogeneity in health status among individuals of the same age [6, 7], frailty is now recognized as a cornerstone of geriatric medicine. It describes a state of heightened vulnerability to stressors, significantly increasing the risk of adverse outcomes, including falls, disability, hospitalization, and mortality [8, 9]. Given its high prevalence and strong predictive value, addressing frailty remains a global public health priority [10]. While the physical model of frailty has long dominated the field, its narrow focus has increasingly come under criticism. Scholars now emphasize the need for a more holistic approach that integrates physical, psychological, and social components [11–13].

Psychological frailty, as an emerging concept for promoting healthy aging and resilience in older adults, has received increasing attention in public health and psychological research. However, despite its growing recognition, there remains no universally accepted definition or fully consistent operational framework. Several competing conceptualizations of psychological frailty have been proposed in the literature. Fitten defined psychological frailty as age-related brain alterations that deviate from normal aging but are not necessarily diseases, thereby contributing to reduced emotional and cognitive resilience when confronted with minor stressors and potentially leading to adverse health outcomes similar to those associated with physical frailty [14]. In contrast, other frameworks adopt a broader mind–body perspective that incorporates physical vulnerability as part of the construct, reflecting the close interrelationship between psychological and physical functioning in older adults [15, 16]. Within this broader conceptualization, the cumulative deficit model provides a useful theoretical basis for integrating heterogeneous health-related deficits into a unified measure of vulnerability [6]. Specifically, frailty is operationalized as the proportion of accumulated deficits relative to the total number of health variables assessed, enabling diverse physical, psychological, cognitive, and social deficits to be standardized and aggregated into a single continuous index.

In the present study, psychological frailty was operationalized as a multidimensional construct encompassing psychological distress, cognitive decline, physical vulnerability, and memory decline [17]. Factor analysis of the 15-item Psychological Frailty Index (PFI) supported this multidimensional structure, demonstrating that it comprises four distinct domains [17]. It should be noted, however, that this operationalization represents only one of several conceptualizations of psychological frailty reported in the literature, and alternative conceptual frameworks may define different operational boundaries. Previous variable-centered validation studies have generally relied on continuous index scores, assuming population homogeneity. This assumption may obscure clinically meaningful subgroups of older adults with distinct symptom configurations. Given that the PFI aggregates multidimensional deficits across emotional, cognitive, memory-related, and physical vulnerability domains, the total score alone cannot determine whether an individual’s vulnerability stems primarily from emotional distress, cognitive concerns, physical exhaustion, or different combinations of these domains. Therefore, a person-centered approach is needed to identify heterogeneous psychological frailty profiles and better characterize individual symptom configurations among older adults.

To address this limitation, the present study employed Latent Profile Analysis (LPA) to identify latent subgroups of older adults with similar psychological frailty profiles. LPA is a person-centered statistical method that classifies individuals into unobserved subgroups (latent profiles) based on their responses to a set of observed variables [18]. This approach is particularly suitable for capturing the heterogeneous manifestations of psychological frailty and identifying distinct symptom configurations among older adults. However, although LPA effectively identifies latent subgroups, it provides limited insight into the symptom interactions underlying each profile. Because individuals assigned to the same latent profile may still differ in the organization and interaction of their symptoms, examining symptom networks within each profile provides complementary analytical insights beyond profile membership alone. To bridge this gap, we additionally conducted network analysis, which conceptualizes psychological frailty as a system of interacting symptoms. In this framework, symptoms are represented as nodes and statistical associations as edges, enabling the visualization and quantification of symptom interaction patterns [19]. To further explore how these symptom networks may respond to hypothetical perturbations and to identify symptoms exerting greater influence on overall network functioning, we additionally applied an in silico network simulation approach. Unlike conventional network analysis, which primarily describes static network structure and centrality, in silico simulation enables the examination of hypothetical symptom dynamics under model-based assumptions, thereby facilitating the identification of highly influential symptoms that may serve as promising, profile-specific intervention targets.

In silico intervention approaches provide a promising extension of traditional network analysis. By simulating hypothetical changes to individual nodes within a network, these methods allow researchers to examine how alterations in one symptom may be associated with changes throughout the modeled system under network-based assumptions [20]. Compared with traditional intervention development, which often requires substantial time and resources [21, 22], computational simulations offer an efficient exploratory framework for generating hypotheses regarding potentially influential symptoms. Simulating both symptom alleviation and symptom aggravation allows the identification of nodes associated with relatively large network-wide changes within the modeled system, thereby providing preliminary insights for future prevention and intervention research [23, 24]. Although in silico simulation has been increasingly applied in psychopathology research, its application to psychological frailty among older adults remains limited. The NodeIdentifyR algorithm (NIRA) extends this approach by enabling in silico simulations of hypothetical symptom perturbations and estimating their influence on the modeled symptom network [25]. Importantly, these simulations are model-based and hypothesis-generating and should not be interpreted as evidence of causal mechanisms, intervention efficacy, or clinical effectiveness. Instead, they provide an exploratory framework for understanding symptom heterogeneity and generating hypotheses that can be evaluated in future longitudinal and experimental studies.

This study conducted a cross-sectional analysis using data from Wave 4 (2018) of the China Health and Retirement Longitudinal Study (CHARLS). By integrating network analysis with hypothetical symptom perturbation simulations across both the overall sample and the identified latent profiles, we aimed to determine which symptoms were associated with relatively larger model-based changes within the estimated network framework. These findings provide an exploratory basis for generating hypotheses regarding profile-specific symptom dynamics that can be evaluated in future longitudinal and intervention research. Furthermore, this study demonstrates the application of advanced computational methods in geriatric mental health research, providing new insights into the heterogeneity of psychological frailty among older adults. These findings may inform the development of profile-specific assessment and intervention strategies, thereby supporting more personalized care for healthy aging.

Materials and methods

Study sample

This study used a cross-sectional design based on data from Wave 4 (2018) of the CHARLS. CHARLS is a nationally representative survey of Chinese adults aged 45 years and older that uses standardized questionnaires to collect comprehensive information on demographic characteristics, socioeconomic factors, and health status. The 2018 survey covered 150 county- level units across 28 provinces, including 450 villages or urban communities, and enrolled 10,257 households comprising 19,816 participants [26]. Detailed methodological descriptions of CHARLS have been published previously, and all data are publicly available at https://charls.pku.edu.cn/. Participants from the 2018 wave were included if they (1) were aged 65 years or older; (2) had complete responses to all 15 items of the PFI, with participants responding “refuse to answer,” “not assessed,” or “do not know” to any PFI item excluded; and (3) had complete data for all study variables. Missing data were handled using a complete-case analysis approach. The final analytic sample comprised 4,735 older adults (Supplementary Fig. 1). The CHARLS study protocol was approved by the Institutional Review Board of Peking University (IRB00001052-11015). This study was conducted and reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement for cross-sectional studies. In addition, the reporting of network analysis followed current recommendations for psychological network analysis (Supplementary Table 1) [27].

Measures

Psychological frailty index

The 15-item PFI comprises four dimensions: psychological distress, cognitive decline, physical vulnerability, and memory decline [17]. All items were recoded to a 0–1 scale, with higher scores indicating greater symptom severity (Supplementary Table 2). The PFI demonstrated satisfactory internal consistency (Cronbach’s α = 0.764) and criterion validity (Spearman’s ρ = 0.806). For psychological distress, responses of “rarely or never” (< 1 day) and “sometimes” (1–2 days) were scored as 0, whereas “often” (3–4 days) and “most or all of the time” (5–7 days) were scored as 1 [28]. For physical vulnerability, responses indicating any level of difficulty (levels 2–4) were coded as 1, whereas “no difficulty” was coded as 0 [29]. For memory decline, response options were collapsed into “no decline” (levels 1–3) and “decline” (levels 4–5) [30]. To capture multidimensional symptom severity, LPA was conducted using domain-level scores, which were calculated as the sum of the original item-level data within each respective domain. Network analysis was performed using the original item-level data to examine pairwise symptom associations. Subsequently, the dichotomized (0/1) responses were used in the Ising model-based network analysis, which is designed for binary variables and enables in silico simulations of hypothetical symptom perturbations [24].

Covariates

The analyses adjusted for a range of sociodemographic and health-related covariates. Sociodemographic characteristics included age (65–74 years vs. ≥75 years), sex (female vs. male), education level (illiterate, primary school or below, secondary school, and high school or above), marital status (married vs. unmarried), and residence (rural vs. urban). Health-related variables included smoking, alcohol consumption, physical exercise, sleep duration, hypertension, diabetes, and history of falls. Smoking, alcohol consumption, physical exercise, hypertension, diabetes, and history of falls were treated as binary variables (no/yes). Sleep duration was categorized into three groups: <6 h, 6–8 h, and > 8 h per day.

Statistical analysis

Data analyses were performed using Mplus version 8.3, SPSS version 25.0, and R version 4.5.1. LPA and the robust three-step multinomial logistic regression (R3STEP) were conducted in Mplus. Descriptive statistics, chi-square tests, multivariate analysis of variance (MANOVA), and ANOVAs were performed using SPSS. Network estimation, network comparison, sensitivity analyses, stability assessments, and in-silico intervention simulations were conducted in R. Missing values were handled using complete-case analysis, and only participants with complete data on all study variables were included. The analytical workflow of the study is shown in Fig. 1.

Fig. 1.

Analytical workflow of the study

Latent profile and group difference analyses

Descriptive analyses were used to summarize sample characteristics across psychological frailty profiles, with continuous variables expressed as means ± standard deviations (M ± SD) and categorical variables as frequencies and percentages. Group differences were examined using chi-square tests, as appropriate. Domain-level scores were included as input variables in the LPA. Models ranging from one to five latent profiles were estimated sequentially. Specifically, model fit was evaluated using the Akaike information criterion (AIC), Bayesian information criterion (BIC), and sample-size adjusted BIC (aBIC), with lower values indicating better model fit [31]. The Lo-Mendell-Rubin test (LMR) and bootstrap likelihood ratio test (BLRT) were used to compare models with k and k − 1 profiles. Significant LMR and BLRT results indicated that the k-profile model provided a better fit than the corresponding k − 1 profile model [32]. Entropy was used to assess classification quality, with values closer to 1 indicating better separation between profiles. We examined the average posterior probability of profile membership. Values ≥ 0.80 indicated a good profile solution [33]. In addition, profiles with a proportion smaller than 0.05 were considered potentially unstable and less interpretable [34]. To facilitate interpretation and comparison of profile characteristics, domain scores were standardized into z-scores (M = 0, SD = 1). Differences between profiles were analysed by MANOVA and ANOVAs as a follow-up. After determining the optimal latent profile solution, the R3STEP procedure was used to include all demographic and health-related characteristics as covariates, identifying predictors of latent profile membership while accounting for classification uncertainty [35].

Network analysis

Network analysis was conducted using R packages bootnet, qgraph, NetworkComparisonTest, dplyr, and readxl. Networks were estimated using Gaussian graphical models (GGMs) with the Extended Bayesian Information Criterion graphical LASSO (EBICglasso) method in the bootnet package [36, 37]. Given the potential non-normal distribution of our dataset, Spearman’s rank correlations were computed prior to network estimation. The tuning parameter γ was set to 0.5, following common practice to balance network sparsity and goodness of fit [38]. Network density and mean absolute edge weight were calculated to characterize global connectivity for each profile [38]. Networks were visualized using the Fruchterman- Reingold layout algorithm.

Node strength centrality and expected influence (EI) were computed in the bootnet package [39]. Strength centrality quantifies a node’s overall connectivity regardless of edge sign, whereas EI captures both positive and negative associations, reflecting a node’s net contribution to the connectivity structure of the network. To facilitate comparison across nodes, these centrality metrics were standardized into Z-scores. Nodes with |Z| ≥ 1.0 were considered relatively prominent within the network [27]. Differences between Profiles 2 and 3 were tested using the NetworkComparisonTest package with 1,000 permutations, assessing both network structure (i.e., the overall patterns of connections) and global strength (GS; the absolute sum of all edge weights) [40]. Statistical significance was evaluated using permutation-based p-values (α = 0.05). To examine robustness to potential confounding, we conducted a sensitivity analysis using residualization. Within each profile, each node variable was regressed onto age, gender, education level, sleep duration, and household registration status. The residuals were then used to re-estimate the network, thereby removing variance attributable to these covariates [41].

To ensure the robustness of the estimated network, we used the package bootnet to assess centrality accuracy and stability [19]. Edge-weight accuracy was evaluated using non-parametric bootstrapping with 1,000 resamples to compute 95% confidence intervals for edge weights. Centrality stability was assessed via case-dropping bootstrapping with 1,000 iterations. Stability was quantified using the correlation stability (CS) coefficient, representing the maximum proportion of cases that can be dropped while maintaining a correlation of ≥ 0.7 between original and subset node strength estimates. A CS-coefficient above 0.25 is considered acceptable, and values above 0.50 indicate good stability [42]. In addition, bootstrapped difference tests were conducted to examine pairwise differences in node strength centrality.

In-silico simulation analysis

To explore the projected network-wide impact of modifying individual psychological frailty symptoms, in-silico intervention simulations were conducted using the NIRA [24]. All psychological frailty indicators were dichotomized (0 = absent, 1 = present) prior to network estimation to satisfy the requirements of the Ising model. A baseline Ising network was estimated using the IsingFit package. Symptom-specific perturbations were then introduced by systematically modifying the threshold parameter of each node while keeping all other network parameters unchanged. The magnitude of the threshold perturbation determines the intensity of the simulated intervention, and the intervention effect size was set at two standard deviations [24]. Two simulation conditions were specified: a decrease condition, in which node thresholds were reduced by two standard deviations, and an increase condition, in which thresholds were increased by two standard deviations [43].

For each perturbed network, 5,000 observations were generated using the IsingSampler package and the Metropolis–Hastings algorithm. Mean total network activation under each perturbation condition was compared with that of the baseline network. Differences between baseline and perturbed conditions were evaluated using independent-samples t-tests. Effect sizes were calculated using Cohen’s d. To account for multiple testing across all node-specific simulations, p-values were adjusted using the False Discovery Rate (FDR) procedure [44]. Nodes with larger effect sizes and greater changes in network activation were identified as showing higher sensitivity to hypothetical perturbations.

Results

Latent profile analysis of psychological frailty

Table 1 presents the fit indices for the one- to five-profile latent profile analysis models. Although the four- and five-profile models showed lower AIC, BIC, and aBIC values, they were not retained due to concerns regarding model stability and classification quality. Specifically, the four-profile model included a class with a proportion below 5%, and the five-profile model showed suboptimal classification accuracy, with an average posterior probability of 0.769 for one class, below the recommended threshold of 0.80. In contrast, the three-profile model demonstrated acceptable model fit, significant LMR and BLRT results (p < 0.001), high entropy (0.976), and excellent average posterior probabilities (0.982–1.000), indicating clear and reliable class separation. Therefore, considering model fit, statistical evidence, classification quality, stability, and interpretability, the three-profile model was selected as the optimal solution (see Table 1).

Table 1.

Fit indices for one- to five-profile latent profile analysis models

Model	AIC	BIC	aBIC	Entropy	LMR(p)	BLRT(p)
1	46934.771	46986.473	46961.052	—	—	—
2	44513.592	44597.608	44556.299	0.949	< 0.001	< 0.001
3	43179.686	43296.015	43238.818	0.976	< 0.001	< 0.001
4	42551.706	42700.349	42627.264	0.992	< 0.001	< 0.001
5	38039.224	38220.180	38131.206	0.930	< 0.001	< 0.001

To guide the interpretation and labeling of the three latent profiles, standardized domain scores and effect sizes were examined across profiles (Supplementary Table 3). As shown in Fig. 2, Profile 1 (n = 2,736, 57.8%) displayed consistently lower standardized scores across all four psychological frailty dimensions (range: −0.695 to − 0.075) and was labeled the “Relatively Healthy Group.” Profile 2 (n = 1,553, 32.8%) showed moderate elevations across all dimensions (range: 0.060–0.532) without any single dominant deficit and was labeled the “Moderate Distress Group.” Profile 3 (n = 446, 9.4%) was distinguished by a disproportionately elevated cognitive decline score (standardized score = 2.411) relative to the other dimensions (range: 0.251–0.410) and was therefore labeled the “Severe Cognitive Deterioration Group.” A MANOVA confirmed significant overall differences among the three profiles across all four psychological frailty dimensions (Pillai’s Trace = 0.922, F(8, 9460) = 1,010.37, p < 0.001, partial η² = 0.461). Among the four dimensions, cognitive decline showed the largest between-profile effect size (partial η² = 0.919), reinforcing its role as the primary distinguishing feature of Profile 3. The remaining dimensions—psychological distress (partial η² = 0.030), physical vulnerability (partial η² = 0.037), and memory decline (partial η² = 0.010)—also differed significantly across profiles but with substantially smaller effect sizes.

Fig. 2.

Profile patterns of indicators in three latent profiles

Characteristics of older adults by psychological frailty profiles

Among the 4,735 older adults included in the study, 57.8% were classified into Profile 1, 32.8% into Profile 2, and 9.4% into Profile 3. The majority of participants were aged 65–74 years (74.8%) and were male (53.8%). Significant differences across latent profiles were observed for most demographic and health-related characteristics. Detailed characteristics of the study participants by psychological frailty profiles are presented in Table 2. Multinomial logistic regression using the R3STEP procedure indicated that, compared with the Relatively Healthy Group (Profile 1), older age (≥ 75 years), urban residence, and long sleep duration (> 8 h) were associated with increased odds of membership in both the Moderate Distress Group (Profile 2) and the Severe Cognitive Deterioration Group (Profile 3) (all p < 0.01). In contrast, male sex, being married, higher education levels, and regular physical exercise were associated with lower odds of membership in these two higher-risk profiles (all p < 0.05). Notably, several predictors showed profile-specific associations in the adjusted model. Smoking was associated only with higher odds of Profile 2 membership (OR = 1.301, p = 0.012), whereas diabetes was associated only with lower odds of Profile 3 membership (OR = 0.574, p < 0.001). No significant associations were observed for alcohol consumption, short sleep duration (< 6 h), history of falls, or hypertension (Supplementary Table 4).

Table 2.

Characteristics of older adults by psychological frailty profiles

Variables	n (%)	Psychological frailty			P
		Profile 1 (2736, 57.8%)	Profile 2 (1553, 32.8%)	Profile 3 (446, 9.4%)
Age (years)					< 0.001
65–74	3540 (74.8)	2127 (77.7)	1116 (71.9)	297 (66.6)
≥ 75	1195 (25.2)	609 (22.3)	437 (28.1)	149 (33.4)
Sex					< 0.001
Female	2188 (46.2)	997 (36.4)	815 (52.5)	376 (84.3)
male	2547 (53.8)	1739 (63.6)	738 (47.5)	70 (15.7)
Education level					< 0.001
Illiterate	2187 (46.2)	771 (28.2)	1013 (65.2)	403 (90.4)
Primary school or below	1378 (29.1)	976 (35.7)	365 (23.5)	37 (8.3)
Secondary school	745 (15.7)	613 (22.4)	126 (8.1)	6 (1.3)
High school or above	425 (9.0)	376 (13.7)	49 (3.2)	0 (0.0)
Marital status					< 0.001
Unmarried	982 (20.7)	435 (15.9)	395 (25.4)	152 (34.1)
Married	3753 (79.3)	2301 (84.1)	1158 (74.6)	294 (65.9)
Residence type					< 0.001
Non-urban	2017 (42.6)	1386 (50.7)	524 (33.7)	107 (24.0)
Urban	2718 (57.4)	1350 (49.3)	1029 (66.3)	339 (76.0)
Smoking					< 0.001
No	3484 (73.6)	1960 (71.6)	1133 (73.0)	391 (87.7)
Yes	1251 (26.4)	776 (28.4)	420 (27.0)	55 (12.3)
Alcohol consumption					< 0.001
No	3164 (66.8)	1696 (62.0)	1111 (71.5)	357 (80.0)
Yes	1571 (33.2)	1040 (38.0)	442 (28.5)	89 (20.0)
Physical exercise					< 0.001
No	525 (11.1)	244 (8.9)	194 (12.5)	87 (19.5)
Yes	4210 (88.9)	2492 (91.1)	1359 (87.5)	359 (80.5)
Hypertension					0.830
No	2403 (50.7)	1380 (50.4)	798 (51.4)	225 (50.4)
Yes	2332 (49.3)	1356 (49.6)	755 (48.6)	221 (49.6)
Diabetes					< 0.001
No	3958 (83.6)	2243 (82.0)	1320 (85.0)	395 (88.6)
Yes	777 (16.4)	493 (18.0)	233 (15.0)	51 (11.4)
Sleep duration (hours/day)					< 0.001
< 6	1747 (36.9)	960 (35.1)	597 (38.4)	190 (42.6)
6–8	2508 (53.0)	1570 (57.4)	752 (48.4)	186 (41.7)
> 8	480 (10.1)	206 (7.5)	204 (13.1)	70 (15.7)
History of falls					0.002
No	3716 (78.5)	2172 (79.4)	1223 (78.8)	321 (72.0)
Yes	1019 (21.5)	564 (20.6)	330 (21.2)	125 (28.0)

Network analysis of latent profiles

Figure 3 presents the estimated networks for Profile 2 and Profile 3. Profile 1 was excluded from network analysis due to zero variance in four disorientation variables (dc007_w4, dc001_w4, dc008_w4, dc002_w4), which precluded estimation of the covariance matrix. Therefore, network analysis was conducted separately for Profile 2 and Profile 3.

Fig. 3.

Network structure of Profile 2 and Profile 3

The networks comprised 15 nodes representing frailty indicators. Profile 2 contained 52 non-zero edges (density = 0.495), whereas Profile 3 contained 38 non-zero edges (density = 0.362), indicating a denser network structure in Profile 2. The mean absolute weight of non-zero edges was comparable between Profile 2 (0.116) and Profile 3 (0.113). The strongest positive regularized partial correlation was observed between dd012_w4 and dd013_w4 in both Profile 2 (0.544) and Profile 3 (0.463). Regarding centrality, dc011 exhibited higher strength and EI in Profile 3 (Z-strength = 2.32, Z-EI = 1.50) than in Profile 2 (Z-strength = 1.02, Z-EI = 0.91). In contrast, dc001_w4 and dc002_w4 showed higher strength centrality in Profile 2 (Z-strength = 1.79 and 1.90, respectively) but lower values in Profile 3 (Z-strength = − 1.40 and − 1.21, respectively), indicating differences in their relative importance across networks (Fig. 4).

Fig. 4.

Standardized strength and expected influence (Z-Scores) of network nodes in Profile 2 and Profile 3

The Network Comparison Test (NCT) revealed significant differences in network structure (M = 0.41, p < 0.001) and global strength (S = 5.87, p = 0.027) between Profile 2 and Profile 3. Sensitivity analyses indicated high similarity between covariate-adjusted and original networks, with edge-weight correlations of r = 0.997 (Profile 2) and r = 0.998 (Profile 3), and minimal changes in global strength. The adjusted NCT remained significant for network structure (M = 0.430, p = 0.001), whereas the difference in global strength was no longer statistically significant (S = 1.763, p = 0.054). Bootstrap analyses showed narrow 95% confidence intervals for most edge weights, indicating good estimation accuracy (Supplementary Fig. 2). Centrality stability was acceptable in both profiles, with CS coefficients of 0.516 (95% CI: 0.439–0.594) for Profile 2 and 0.749 (95% CI: 0.673-1.000) for Profile 3, both exceeding the recommended threshold of 0.50 (Supplementary Fig. 3). Full results of centrality difference tests are provided in Supplementary Fig. 4.

Simulation of alleviating and aggravating interventions

To examine the network-level response to hypothetical changes in individual symptoms, in silico intervention analyses were conducted for each node in the overall network and in the two latent profiles (Profile 2 and Profile 3). Profile 1 was excluded due to insufficient variability in several indicators, which may compromise the stability of network estimation and simulation results.

Figure 5 presents the simulation results for the overall symptom network in the total sample. Under the alleviating condition, reductions in all 15 symptoms were associated with significant decreases in total symptom burden after FDR correction (all adjusted p < 0.05). The largest model-based decreases were observed for db001, dc012, and db004, indicating greater network-level sensitivity to perturbations in these nodes. Under the aggravating condition, increases in all symptoms were associated with significant increases in total symptom burden (all adjusted p < 0.05). The largest model-based increases were observed for dc012, dc011, and dc010, suggesting these nodes were associated with the greatest increases in network activation under simulated worsening conditions.

Fig. 5.

Simulation of alleviating (A) and aggravating (B) interventions in the total sample

For Profile 2, under the alleviating condition, reductions in most nodes were associated with significant decreases in total network activation after FDR correction. However, dc007_w4, dc008_w4, and dc002_w4 did not reach statistical significance. The largest model-based decreases were observed for db001, dc012, and db004. Under the aggravating condition, increases in most nodes were associated with significant increases in total network activation. The strongest effects were observed for db007, dc011, and dc017 (Fig. 6).

Fig. 6.

Simulation of alleviating (A) and aggravating (B) interventions in Profile 2

For Profile 3, which was characterized by higher baseline symptom burden, alleviating interventions were associated with reductions across all 15 nodes, with the largest model-based decreases observed for dc012, dc011, and db001. Under the aggravating condition, increases in symptom burden were observed for most nodes; however, dc007_w4, dc002_w4, and dc008_w4 did not reach statistical significance after FDR correction. The largest model-based increases were observed for db007, dc011, dc009, and dc010 (Fig. 7).

Fig. 7.

Simulation of alleviating (A) and aggravating (B) interventions in Profile 3

Discussion

The present study identified three distinct psychological frailty profiles among older adults in China through latent profile analysis: the “Relatively Healthy Group” (57.8%), the “Moderate Distress Group” (32.8%), and the “Severe Cognitive Deterioration Group” (9.4%). Network analysis further revealed differences in network structure and connectivity patterns across profiles, suggesting heterogeneity in symptom interrelationships beyond differences in overall symptom severity. In addition, in silico simulation analyses were conducted to examine potential network-level responses to hypothetical changes in individual symptoms, offering additional insights into profile-specific variations in symptom activation within the estimated model. Overall, these findings indicate that psychological frailty subtypes differ not only in symptom severity but also in the organization of symptom interactions.

Heterogeneity of psychological frailty profiles

The LPA identified three distinct psychological frailty profiles among older adults, indicating substantial heterogeneity in symptom severity and distribution. The three-profile solution demonstrated optimal model fit and satisfactory classification quality, suggesting that psychological frailty is more appropriately conceptualized as comprising qualitatively distinct subgroups rather than a homogeneous construct. Profile 1 was characterized by consistently low levels of impairment across all psychological domains, particularly cognitive functioning, indicating relatively preserved psychological health. This profile may reflect a subgroup with stable and resilient functioning. Profile 2 exhibited moderate and relatively balanced impairments across cognitive, emotional, and vulnerability domains, suggesting a broadly distributed symptom burden without a clearly dominant dimension. This pattern implies that integrated, multi-domain assessment and intervention strategies may be more appropriate for this subgroup. Profile 3 was distinguished by markedly elevated impairments across all domains, with particularly pronounced cognitive deficits, indicating a subgroup with severe global dysfunction. These individuals may represent a high-risk group requiring prioritized clinical attention and monitoring. Overall, these findings are consistent with previous research emphasizing the heterogeneity of psychological frailty and the complex interplay among cognitive, emotional, and physical domains, as well as the protective roles of psychological resilience and social support in mitigating frailty risk [45, 46].

Significant differences in sociodemographic characteristics, lifestyle behaviors, and health conditions were observed across the three psychological frailty profiles. Individuals in Profile 1 were more likely to be younger, male, married, and have higher educational attainment, which is in line with evidence linking these factors to lower levels of frailty [47]. Higher educational attainment may facilitate better access to health-related knowledge and social participation opportunities, which in turn may help preserve psychological functioning in later life [48]. Regarding lifestyle behaviors, regular exercise was associated with lower odds of belonging to Profiles 2 and 3, supporting its well-established role in maintaining physical and psychological health [49]. For health conditions, no significant differences were observed for hypertension across profiles, whereas diabetes was associated with lower odds of membership in Profile 3. This finding may suggest that diabetes is more closely related to physical aspects of frailty than to the psychological frailty patterns captured in the present study [50]. However, this interpretation should be made cautiously given the complexity of multimorbidity in older adults. Overall, these findings indicate that relying solely on the presence or number of chronic diseases may be insufficient to explain psychological frailty profiles. Future research should consider a broader spectrum of health conditions and examine how specific disease patterns contribute to distinct dimensions of psychological frailty. In addition, qualitative approaches may provide valuable insights into the subjective experiences and psychological burdens associated with chronic illness in later life.

Symptom network characteristics across psychological frailty profiles

Network analysis indicated distinct symptom organizations across profiles. Orientation-related indicators occupied relatively central positions in Profile 2, whereas depressive symptoms emerged as the most central nodes in Profile 3, suggesting a topological shift from cognitive orientation processes toward affective symptoms with increasing psychological frailty severity. This interpretation is broadly consistent with longitudinal evidence from community-dwelling older adults without overt cognitive impairment, where latent cognitive profiles, including orientation-related deficits, were associated with increased risks of depression, anxiety, insomnia, falls, and functional decline [51]. These findings support the possibility that cognitive-related symptoms may be more characteristic of lower-burden stages of psychological frailty. Future longitudinal research is needed to examine whether these cross-sectional differences reflect a potential severity-related shift in symptom organization, whereby orientation-related symptoms become less central as affective symptoms gain prominence over time.

Covariate-adjusted analyses showed minimal changes in network structure within each profile, with high correlations between adjusted and unadjusted edge weights, indicating robust and stable symptom associations. Differences in network structure between profiles persisted; however, they were attenuated in terms of global connectivity after adjustment for socio-demographic factors, suggesting that profile differences may be more closely related to overall network connectivity than to specific pairwise symptom interactions. Overall, these findings indicate that psychological frailty may involve differences not only in symptom severity but also in symptom organization. However, given the exploratory nature of the analyses and the relatively lower stability of centrality estimates in Profile 2, these findings should be interpreted with caution. Longitudinal research remains necessary to assess the temporal stability of these network patterns and to clarify their potential relevance for understanding trajectories of frailty progression.

Insights from network-based simulation analyses

Using simulated symptom aggravation and alleviation scenarios, we identified depression, powerlessness, attentional problems, and functional limitations as symptoms associated with relatively larger model-based changes in the psychological frailty network. Among these, powerlessness exhibited the strongest bidirectional modeled influence on overall network burden, generating the hypothesis that perceptions of control may represent an important feature within the symptom network. This interpretation is broadly consistent with qualitative evidence suggesting that perceived control and self-efficacy are integral to the experience of frailty [52]. Older adults and caregivers have described frailty not only in terms of physical decline but also as involving concerns about control, coping, and adaptation, whereas higher self-efficacy has been associated with greater resilience and adaptation to frailty-related challenges. If these simulated patterns are confirmed in prospective studies, they may suggest potential mechanical pathways for future empirical evaluation, such as investigating how variations in self-efficacy relate to network density during active engagement in health management and social participation.

Subgroup analyses further suggested that the modeled symptom dynamics differed across the identified psychological frailty profiles. In Profile 2, functional limitations and limited physical activity exhibited greater responsiveness under both simulated alleviating and aggravating conditions, suggesting a functionally oriented symptom configuration. These findings suggest that functional capacity and physical activity may occupy relatively important positions within the modeled symptom network of this subgroup. Previous evidence has identified exercise as an effective strategy for mitigating frailty-related functional decline [53]. Our simulation results generate the hypothesis that functional limitations and low physical activity may be symptom nodes associated with relatively larger model-based changes within this subgroup’s network. If future longitudinal observations targeting functional capacity demonstrate measurable changes, they may provide an opportunity to test whether these observed symptom-level fluctuations propagate through the network in ways consistent with the model-based predictions generated by the present simulations.

In Profile 3, depressive symptoms demonstrated greater responsiveness under both simulated conditions, suggesting a comparatively affect-centered symptom configuration. This pattern highlights the potential importance of affective symptoms within this subgroup. Previous randomized studies have demonstrated the efficacy and acceptability of both cognitive behavioral therapy and emotion-focused therapy for reducing depressive symptoms [54]. Our simulation findings do not speak to the effectiveness of these or any other interventions. Rather, they generate the hypothesis that depressive symptoms may represent a relatively highly responsive node within this subgroup’s network. Future prospective longitudinal studies could test this hypothesis by examining whether naturally occurring fluctuations in depressive symptoms are accompanied by network-level changes that are consistent with the patterns predicted by the model-based simulations.

Limitations and future directions

Several limitations should be noted. First, given the cross-sectional design, we were able to characterize concurrent symptom burdens and their interconnections; however, causal inferences cannot be established. Longitudinal studies are needed to examine the temporal stability of these network structures and to assess the robustness of the model-based simulation results over time. Second, the study sample consisted solely of Chinese older adults, and cultural and contextual factors may influence both the expression and interrelationships of psychological frailty indicators. Future studies in diverse populations are needed to evaluate the generalizability and replicability of the observed network structures and subgroup patterns. Third, although the in silico simulation analyses were data-driven, their model-based nature warrants caution in interpretation. Real-world intervention studies are required to validate whether targeting the identified symptoms yields comparable effects in practice. Future work may also consider approaches such as target trial emulation to strengthen causal inference and enhance translational relevance. Fourth, although several key covariates were included in sensitivity analyses, the possibility of residual confounding due to unmeasured or unavailable factors (e.g., medication use, social support, and detailed socioeconomic conditions) cannot be excluded. Future research incorporating more comprehensive clinical and social determinants is needed to further validate these findings.

Conclusion

By integrating LPA with network simulations, this study provides preliminary insights into both generalized and subgroup-specific characteristics of psychological frailty in older adults. Given the cross-sectional, observational nature of the data, these model-based associations should not be interpreted as causal mechanisms, nor as evidence to support clinical intervention decisions. Rather, these exploratory findings offer a conceptual basis for hypothesis generation, which may inform future longitudinal research on tailored, profile-specific symptom dynamics.

Abbreviations

LPA

Latent profile analysis

PFI

Psychological frailty index

GGM

Gaussian graphical model

EBICglasso

Extended Bayesian Information Criterion Graphical LASSO

NIRA

NodeIdentifyR algorithm

CHARLS

China Health and Retirement Longitudinal Study

NCT

Network comparison test

CS

Correlation stability

FDR

False discovery rate

AIC

Akaike information criterion

BIC

Bayesian information criterion

aBIC

Sample-size adjusted BIC

LMR

Lo-Mendell-Rubin test

BLRT

Bootstrap likelihood ratio test

R3STEP

Robust three-step multinomial logistic regression

Authors’ contributions

YB and ZJ conceptualized and designed the study. FW and YB carried out the initial analyses, drafted the initial manuscript and reviewed and revised the manuscript. MF, ZW and JL critically reviewed and revised the manuscript; and all authors approved the final manuscript as submitted.

Funding

This study was supported by grants from the Medicine and Health Science and Technology Plan Project of Zhejiang Province (Grant No. 2023KY837).

Data availability

The data used in this study are accessible through the China Health and Retirement Longitudinal Study (CHARLS) upon registration and application. More information is available at https://charls.pku.edu.cn/.

Declarations

Ethics approval and consent to participate

The studies involving human participants were reviewed and approved by Peking University (IRB00001052-11015). The studies were conducted in accordance with local legislation and institutional requirements. Written informed consent was obtained from all participants to participate in this study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.