This prospective study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines (Supplementary Table 2). Participants were recruited using convenience sampling from a tertiary general hospital in Guangdong Province, China, between January and November 2024. Ethical approval was obtained from the Institutional Review Board of the participating hospital (Approval No. SDFY B-2022-268) and all procedures complied with institutional and national ethical standards. Written informed consent was obtained from all participants prior to enrollment.

Sample size estimation was based on Kendall’s rule, which recommends 5~10 participants per variable for multivariate analyses (30). With 20 sociodemographic variables, Water Swallowing Test (WST), National Institutes of Health Stroke Scale (NIHSS), Barthel Index (BI), 10-item Connor–Davidson Resilience Scale (CD-RISC-10) and Social Support Rating (SSRS), a minimum of 125 to 250 participants was required. Allowing for a 10% drop-out rate, the final target sample size ranged from 139 to 278 participants. Furthermore, previous studies recommend a sample size of 300 to 1000 for LPA (31). The study ultimately included 502 participants.

Key independent variables, including sociodemographic data, disease-related characteristics, laboratory indicators, stroke severity, swallowing function, activities of daily living, resilience scale and social support scale were collected within one week of disease onset. Depression and anxiety were assessed via phone follow-up at 3 months post-stroke.

Trained research personnel explained the study procedures and obtained written informed consent at hospital admission. Baseline assessments, including demographics, clinical characteristics and laboratory indicators, were recorded. Swallowing function and NIHSS were evaluated by a team comprising trained researchers and clinicians. Instruct the patients to fill in the CD-RISC-10 and SSRS. Depression and anxiety assessments were conducted via telephone at 3 months (44). All data were independently entered and double-checked by two researchers using EpiData 3.1 to ensure accuracy.

Missing data were handled using multiple imputation. To assess potential common method bias, Harman’s single-factor test was performed. Descriptive statistics were conducted using SPSS version 26.0. Categorical variables were presented as frequencies and percentages; continuous variables as means ± standard deviations (M ± SD). The Kolmogorov–Smirnov test was used to assess normality, revealing non-normal distributions for all continuous variables.

LPA was performed in R version 4.4.3 using the tidyLPA package (45). Model fit was evaluated using the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), sample size adjusted BIC (aBIC), entropy and the Bootstrap Likelihood Ratio Test (BLRT) (46). Lower AIC, BIC and aBIC values indicated superior model fit. Entropy values ≥0.80 were considered acceptable and a significant BLRT P supported the inclusion of additional classes (47). Models were also required to include no latent class smaller than 5% of the sample (48).

Between-profile comparisons were conducted using the Kruskal–Wallis test for continuous variables and chi-square or Fisher’s exact test for categorical variables, as appropriate. Due to unequal class sizes, Firth’s penalized likelihood logistic regression was used for multivariable analysis (49), which reduces small-sample bias and addresses issues such as separation (50, 51). Variables significant in univariate analyses (P < 0.1) were screened for multicollinearity (52); no multicollinearity was detected. Final models reported odds ratios (OR) and 95% confidence intervals (CI). A P < 0.05 was considered statistically significant.

Out of the 551 individuals initially enrolled in the study, 49 were lost to follow-up or chose to discontinue participation. As a result, a total of 502 participants (91.11%) were included in the final analysis (Figure 1). The mean age of the participants was 67.19 years (SD = 0.48), with 65.54% being male and 97.61% married. A detailed breakdown of the demographic and clinical characteristics of the AIS patients is provided in Table 1.

Given the use of self-reported data, Harman’s single-factor test was conducted through exploratory factor analysis to assess potential common method bias. Eight factors with eigenvalues greater than 1 were extracted, explaining a total of 61.93% of the variance. The first factor accounted for only 12.17% of the variance, which is well below the commonly accepted threshold of 40% (53), indicating that common method bias was not a concern.

To further validate the analysis, multicollinearity diagnostics were performed. Tolerance values ranged from 0.12 to 0.95, and the variance inflation factors (VIFs) ranged from 1.06 to 8.32. As none of the VIFs exceeded the threshold of 10 (54), multicollinearity among the independent variables was not considered problematic (Supplementary Table 3).

LPA was performed using the PHQ-9 and GAD-7 scores as indicator variables. Models ranging from one to five classes were evaluated (Table 2). The BLRT yielded significant P for all models, and entropy values were above 0.9 for the two to five class solutions, indicating good classification precision. AIC, BIC and aBIC values decreased with each additional class. Although the four-class model exhibited an acceptable fit, it did not meet the minimum sample size requirement per class (n > 50) for model stability (55). Therefore, the three-class model was chosen as the optimal solution. The three latent classes were as follows: no depression-anxiety (Class 1: 67.93%), high-risk depression-anxiety (Class 2: 21.12%) and major depression-anxiety (Class 3: 10.95%). The mean PHQ-9 and GAD-7 scores for each class are presented in Figure 2. Item-level comparisons across classes are shown in Table 3. One-way ANOVA revealed significant differences in all item scores (P < 0.001). Post hoc Bonferroni tests indicated that Class 1 significantly differed from both Classes 2 and 3 across all items, while differences between Classes 2 and 3 were significant only for selected items.

Univariate analysis identified several factors significantly associated with latent depression-anxiety profiles (P < 0.1), including: personal traits: hypertension、WST、age、PNI、SII、NLR、PLR; behavioral characteristics: BI and NIHSS; interpersonal network: SSRS; work/living conditions: occupation and CD-RISC-10 (Table 1).

To further explore factors associated with latent class membership, Firth’s logistic regression was performed. Variables with P < 0.1 in the univariate analysis were included in the model, with no depression-anxiety class serving as the reference category. For high-risk depression-anxiety class, significant predictors included: occupation (OR = 0.61, 95% CI [0.40–0.93]), hypertension (OR = 2.37, 95% CI [1.29–4.35]), NIHSS (OR = 1.60, 95% CI [1.06–2.42]) and BI (OR = 1.67, 95% CI [1.27–2.19]). For major depression-anxiety class, significant predictors were: NIHSS (OR = 2.33, 95% CI [1.42–3.85]), BI (OR = 2.65, 95% CI [1.62–4.35]) and CD-RISC-10 (OR = 0.92, 95% CI [0.87–0.98]). Full regression results are summarized in Table 4.

In this study, LPA identified three distinct subgroups of psychological status among patients with AIS, highlighting substantial heterogeneity in co-occurring depression and anxiety symptoms. This person-centered analytic approach provides a more nuanced understanding of post-stroke psychological responses by classifying individuals based on symptom patterns, rather than relying on aggregated mean scores. Three latent classes were identified based on item response means: “no depression-anxiety” (Class 1: 67.93%), “high-risk depression-anxiety” (Class 2: 21.12%) and “major depression-anxiety” (Class 3: 10.95%). Class 1 demonstrated consistently low scores across all variables, indicating better overall psychological health. Class 2 is characterized by scores that fall between those of Class 1 and Class 3. Prior research has established a cutoff of 5 or higher on the PHQ-9 and GAD-7 as indicative of mild depression and anxiety, respectively (56, 57). In addition, other studies have identified thresholds for severe symptoms, defining major depression as a PHQ-9 score of 10 or above (58) and severe anxiety as a GAD-7 score of 8 or above (59). Given the mean PHQ-9 and GAD-7 scores in Class 2 (M_PHQ-9 = 8.56; M_GAD-7 = 6.83), this group comprises individuals exhibiting elevated levels of depressive and anxiety symptoms—levels that exceed the thresholds typically indicative of mild symptomatology and approach the severity observed in major depressive and anxiety disorders. These individuals are at increased risk of developing clinically significant psychological distress, justifying their classification as a distinct profile. Patients in Class 2, while not meeting diagnostic thresholds, occupy a vulnerable “borderline” state that may progress to more severe psychopathology if left unaddressed (58, 60). In contrast, Class 3 had significantly higher scores across all items, reflecting the presence of severe psychological symptoms. Notably, two items consistently yielded the highest scores in both Class 2 and Class 3: GAD-7_1 (“Feeling nervous, anxious, or on edge”) and PHQ-9_1 (“Little interest or pleasure in doing things”). These findings suggest that anhedonia and nervousness may represent core, early-emerging symptoms of post-stroke psychological distress. Future screening instruments should prioritize these sentinel items to facilitate early identification of at-risk individuals. Importantly, previous evidence indicates that screening alone is insufficient to improve patient outcomes (20). These findings highlight the critical need for interdisciplinary collaboration, integrating systematic detection with targeted psychological and rehabilitative interventions.

As a fundamental element of the HEM, individual characteristics play a pivotal role in shaping post-stroke psychological outcomes, with particular emphasis on hypertension and stroke severity. Hypertension, a well-established vascular risk factor (61), was independently associated with a 2.37-fold increased risk of comorbid PSD and PSA. A possible mechanism for this involves sympathetic nervous system overactivation in hypertensive individuals (62), leading to elevated catecholamine levels, which contribute to heightened arousal, anxiety and emotional dysregulation. Stroke severity, measured by NIHSS, was also a significant predictor of depression-anxiety profiles (Class 2: OR = 1.60; Class 3: OR = 2.33). More severe neurological impairments exacerbate emotional distress through various pathways, including physical disability, impaired communication and reduced autonomy, which hinder engagement in meaningful activities and increase the likelihood of social isolation (63–65).

In terms of behavioral characteristics, deficits in activities of daily living, as measured by the BI, were strongly associated with increased depression and anxiety risk (Class 2: OR = 1.67; Class 3: OR = 2.65). Reduced functional independence may undermine self-esteem and contribute to feelings of helplessness and identity loss (66). Moreover, the inability to participate in familial and social life may erode social support and contribute to psychological morbidity (67). Stroke survivors frequently encounter not only physical disabilities but also emotional instability and the redefinition of their social roles. These challenges play a critical role in escalating psychological distress. Resilience, defined as the ability to recover and adapt when confronted with adversity, emerged as an important protective factor for mental health (68). In this study, resilience was significantly linked to latent subgroup classification at three months post-stroke, serving as a buffer against psychological distress. This finding aligns with evidence from oncology research, where lower resilience is associated with a rapid escalation of negative emotional states, including depression and anxiety (69). Although many longitudinal studies have neglected the perspective of comorbidity, one investigation demonstrated that resilience independently predicts PSD across multiple time points within six months, reinforcing its role as a critical protective factor (70). Individuals with higher resilience tend to adopt more constructive perspectives when facing illness, which enhances engagement in rehabilitation and leads to better health outcomes (71–73). Therefore, early interventions designed to enhance resilience-related skills may substantially improve patients’ self-efficacy while concurrently reducing psychological distress, anxiety and depressive symptoms (74).

Regarding work and living conditions, the study suggests that engagement in occupational activities, including agricultural work, serves as a protective factor against both depression and anxiety in stroke survivors. Occupation plays a dual role, not only providing financial support but also forming a core element of personal identity, social integration and existential purpose (75). Regular involvement in work-related tasks enhances autonomy, perceived self-efficacy and a sense of accomplishment, which contributes to psychological well-being and improves overall quality of life (76, 77).

Turning to Interpersonal Networks, although univariate analysis identified a significant association between social support and psychological profiles, this effect was not sustained in the multivariable model. This may indicate that social support exerts indirect effects on emotional outcomes, possibly mediated through functional status or resilience (78). Alternatively, the impact of social support may be more significant during later stages of recovery than during the acute post-stroke period (79). Future studies employing longitudinal structural equation modeling could offer deeper insights into the mediating effects and temporal dynamics of social support on mental health.

Several limitations should be acknowledged in this study. First, the outcome measures were assessed at a single time point, three months post-AIS, which restricts our ability to examine the long-term trajectory of psychological profiles in AIS patients. Future research will address this limitation by examining the long-term trajectories of these psychological features. Second, the use of telephone follow-up to collect outcome data introduces potential reporting biases, which may affect the accuracy of the results. Thirdly, while the HEM provided a robust theoretical framework for identifying potential influencing factors, certain significant determinants were not incorporated into this study. Specifically, environmental, biological and psychosocial factors may have contributed to the observed associations. Future research should adopt a longitudinal design to track psychological outcomes over a longer period and utilize more objective data collection methods for follow-up assessments. Incorporating neuroimaging and other objective biomarkers, alongside the use of mediation analysis, would help clarify the underlying mechanisms. Furthermore, the potential mediating role of social support in the development of concurrent PSD and PSA warrants focused investigation.

This study identified three latent psychological profiles among AIS survivors: no depression-anxiety, high-risk depression-anxiety and major depression-anxiety, with clear inter-individual heterogeneity. Five key factors including occupation, hypertension, NIHSS, BI and resilience were significantly associated with profiles. Our findings support the integration of early psychological screening protocols into routine AIS care, particularly for individuals at high risk based on clinical and sociodemographic characteristics. Interventions designed to enhance resilience and facilitate occupational re-engagement should be prioritized to mitigate psychiatric morbidity, improve functional outcomes and promote holistic recovery in stroke patients.