With advancing age, people tend to spend more time indoors; interior environmental factors significantly affect older adults' physical and mental health and are indispensable for supporting healthy aging (34). The scope of this study primarily concerns modifiable built-environment features in interior settings, such as lighting, layout, accessibility design, and furniture.

Prior research has shown that general physical health is associated with ventilation/air quality and lighting (34). Lighting has become the most comprehensively studied interior factor, particularly with respect to its effects on sleep, mood, and daily functioning. Older adults in nursing homes may have fewer opportunities for exposure to natural light than community-dwelling older adults (35). Appropriate lighting design affects visual function and decision-making and may mitigate depressive and anxious tendencies (36–38). Accordingly, increasing daylighting through windows and optimizing lighting design to ensure a bright and comfortable visual environment are critical to promoting health. Existing studies have extended dynamic correlated color temperature schemes or illuminance recommendations to daily activities (39, 40). Low levels of ventilation/air quality are associated with elevated indoor CO₂ concentrations and increased cardiovascular strain; indoor plants are a practical and low-cost strategy to improve air quality (41).

In layout-related studies, bathroom configuration and bedroom privacy most clearly demonstrate key roles in supporting older adults' health (34). Related factors include the completeness of amenities within functional spaces, spatial connectivity between functional zones, and the degree of control over personal rooms (35). Visibility, perceived openness, and accessibility of space are critical for older adults; adequate space facilitates the use of walkers and enhances social interaction, whereas overly narrow spaces may increase depressive symptoms and extreme affect among older adults with dementia (42). Broad sightlines and the ability to view natural scenery from indoors can also substantially enhance positive perceptions (43).

As a key enabler of mobility and daily functioning in older adults, accessibility design is increasingly integrated with multisensory visual strategies (44, 45). This involves not only installing infrastructure such as call buttons in living spaces (46) but also emphasizing furniture, door and window design, material and color schemes, and visual access to exterior landscapes. For example, warm, bright colors can enhance spatial awareness (47), and older adults with dementia are more likely to exhibit relaxed, calm, and pleasant affect in environments with soft color palettes (48). In addition, personalized door decorations can aid wayfinding memory among older adults with cognitive impairment (38); high color-contrast handrails against background walls, floors, and doors made of materials with low reflectance, and LED guide lights are also strategies to improve visibility and reduce fear of falling (36, 49). Prior research has shown that when wood coverage in activity rooms for older adults is about 30%, subjective responses are most positive, whereas 90% coverage produces a significant calming effect (50). In recent healthcare settings, the discussion of furniture has expanded to deeper health dimensions—such as cognitive performance, loneliness, and satisfaction (36, 38)—reflecting a shift in furniture design from purely functional considerations toward affective and symbolic roles. For instance, open kitchens or dining spaces may encourage residents to gather and engage in daily activities, providing a supportive social environment (51).

Research over recent decades has demonstrated that outdoor environments can serve as resources for recovery and rehabilitation, leading to the establishment of gardens in healthcare settings (52). The health benefits of outdoor activities for older adults have been extensively discussed, with their psychological, cognitive, emotional, mobility, and physical function linked to diverse outdoor spaces (53–55), including improved mood, sleep, wellbeing, and quality of life (56–58). Despite the well-established positive impact of outdoor natural landscapes on health and wellbeing, gaps remain in understanding their effects on individuals with dementia residing in long-term care (LTC) facilities (59).

For residents with dementia living in nursing homes, the value of outdoor natural landscapes within the healthcare system is associated with characteristics of dementia-friendly environments, including orientation, accessibility, socialization, meaningful activities (meaningful engagement), sensory stimulation, reminiscence, safety, and sustainability, in order to meet their needs and maximize the therapeutic potential of outdoor natural landscapes (59). Previous studies have identified the following drivers of outdoor space utilization in nursing homes: garden planting and architectural features, safety concerns and staffing, adequate shelter, the design of main-building entrances, and social activities (56). Therefore, providing residents with opportunities for plant cultivation in gardens, equipping spaces with interactive facilities, weather-protected seating, and accessible entrances, and organizing social activities are crucial for facilitating their engagement with outdoor environments.

Expansive, aesthetically pleasing outdoor landscapes and greenery have restorative effects on older adults, helping them maintain a positive mood (59). Biophilic experiences in walking environments—such as green spaces, vegetation, and natural visual simulations—provide positive sensory stimulation, significantly reducing participants' stress and anxiety and offering measurable therapeutic benefits (55). Therefore, providing sufficient pathways and walkable surfaces is crucial to enable safe movement (56). Installing handrails in outdoor walking areas to assist residents with mobility and balance is recognized as enhancing a sense of safety (60). Furthermore, outdoor spaces and green spaces should offer ample opportunities for social interaction among residents, fostering engagement and strengthening a sense of belonging (61). Adequate seating that is easily accessible, comfortable, and not overly hard is necessary for residents to enjoy garden spaces (56). Heavy or locked doors and high thresholds or doorsteps present barriers to independent outdoor access for older adults, particularly those using walkers (62). Several studies also emphasize the need for adequate lighting in outdoor spaces, especially at night (56, 63). Some nursing-home residents desire views beyond garden boundaries to observe surrounding neighborhoods or businesses, fostering a connection to life outside the facility (64).

Although assisted living environments primarily focus on physical safety and visual supports, integrating cultural and psychological design elements is also crucial (61). The cultural characteristics of nursing homes, and their connections with the broader society, support nursing-home residents' continued sense of belonging and engagement in the local community (65). This specific culture arises not only from the words and actions of fellow residents and staff but is also manifested in details such as decor and furnishings. The provision of personalized care enhances residents' physical safety and emotional wellbeing in later life (34). Specific dimensions and indicators used for MLLM description are shown in Table 1.

This study aims to identify the belief structures held by nursing homes regarding the built environment for chronic disease rehabilitation care, thereby revealing their value orientations in rehabilitation practices and environmental design. To this end, a hybrid “semantic-cognitive” framework integrating multimodal language models (MLLM), natural language processing (NLP), and Q-methodology was constructed to systematically explore these relationships (see Figure 1).

Q methodology is regarded as a research approach that investigates individual perceptions via the application of inverted factor analysis (66). While it can characterize holistic responses and aggregate them into stable configurations of viewpoints (67), its primary limitations lie in preparatory information gathering and Q-statement construction: the selection of contextual materials and wording is influenced by researchers' choices, making it difficult to fully eliminate the interaction between the subjectivities of the observer and the observed (67). Therefore, given the high accessibility of information on web portals, the environmental images and supporting materials that institutions select and publish on these portals provide explicit signals of their spatial orientations and care philosophies. Compared with template-based, highly stylized texts, images that directly present spatial elements and service contexts better reflect underlying beliefs. These beliefs can be viewed as the externalization of practical knowledge—transformed from accumulated experience—by institutional managers, frontline caregivers, designers, and other stakeholders through long-term work with older adults, information that has yet to be systematically explored.

To limit discretion and enhance transparency in the chain of evidence, this study employs MLLM and NLP as upfront preprocessing to provide data-constrained sources for the statements. First, MLLM automatically extracts spatial semantics and care-context information from images scraped from web portals, converting visual representations of the built environment into analyzable textual descriptions. Subsequently, NLP vectorizes and clusters these semantic texts to construct a data-constrained Q-statement set, and a semantic similarity model computes an institution-by-statement scoring matrix that serves as the objective structural basis for subsequent Q-sorting and factor analysis. The methodological process includes defining and constructing the concourse, developing the Q-set, selecting the P-set, and conducting Q-sorting, analysis, and interpretation (66, 67). The following subsections detail the Q-methodological steps of this study.

The concourse is the entirety of discourse concerning a specific research question (68) and is constructed by gathering statements that represent a broad range of subjective perspectives on that issue. Subsequently, a final set of statements—the Q sample—must be selected. It must “broadly represent the opinion domain” and should “demonstrate good coverage relevant to the research question.” The collection process includes both structured and unstructured methods (69). Q methodology does not define a specific procedure for generating statements. In practice, Q sets are created based on scientific literature, policy documents, mass media, structured interviews, and informal discussions with experts (70). Following the unstructured approach, this section derives “data-constrained” Q-sample statements by integrating MLLM with NLP, with the process steps illustrated in Figure 2.

Step 1: Identify dimensions and acquire semantic data. The concourse is constructed based on existing research on the relationship between the built environment of nursing homes and chronic disease rehabilitation, from which core dimensions highly relevant to the research theme are distilled, specifically including: for indoor environments—“Daylighting and Lighting conditions,” “Ventilation and Air Quality,” “Spatial Layout and Privacy,” and “Safety and Assistive features”; for outdoor environments—“Greening and Natural elements,” “Paths and Surface conditions,” “Main-building design,” and “Social Opportunities.” In addition, a “free dimension” is set to capture unspecified features (see Supplementary materials for detailed instruction configuration). These dimensions are used to instruct a multimodal language model (MLLM) to objectively audit environmental images self-selected and uploaded by institutions to nursing-home web portals, thereby generating descriptive text of the images. MLLMs have been shown to achieve high accuracy and stability in the semantic understanding of images and the recognition of scene elements. This study employed GPT-5.0 (71, 72), the latest iteration of such models.

Step 2: Verification and Reliability Assessment. To verify the usability and consistency of the model-generated text, this study randomly selected ~10% of the samples (358 images in total). Two experts with backgrounds in environmental design or elderly care assessed the consistency of elements between the original images and the generated text. The assessment results showed that the two experts classified 95.35 and 94.69% of the generated text as fully usable and 1.96 and 3.07% as partially usable, respectively. The proportion of unusable text was 2.52%. The kappa coefficient calculated based on the above classification results was 0.822 (95% CI: 0.658–0.945, P < 0.001). These results support a high level of consistency between assessors and demonstrate that the GPT output has good reliability in key element identification and semantic integrity.

Step 3: Embedding, Dimensionality Reduction, and Cluster Specification. These qualitative texts extensively cover nursing homes' statements regarding perceptions of the built environment for chronic disease rehabilitation. Information extraction in this study is achieved through inductive analysis based on natural language processing (NLP). After sentence segmentation and normalization, semantic embeddings are generated with a multilingual SBERT model, followed by principal component analysis (PCA) for dimensionality reduction to mitigate noise and redundancy. K-means clustering is then applied to the semantic vector matrix to identify latent thematic structures. To determine the number of clusters K, an elbow curve was computed over a grid with K ∈[2,60] to delimit the candidate range; as shown in Figure 3, the curve exhibits diminishing marginal improvements for K ∈[10,20], indicating the candidate interval. To further refine the choice of K, the silhouette score and the Davies-Bouldin index (DBI) were jointly used for validity assessment, where larger silhouette values indicate greater within-cluster compactness and better between-cluster separation, and smaller DBI values are preferred (73). Results showed that K = 4 achieved the best overall ranking (Silhouette = 0.097, DBI = 2.571). All computations were performed in PyCharm 2023.2.3.

Step 4: Statement Screening and Refinement. Based on the clustering results, the research team first selected three statements from each partition that were closest to the centroid and had a certain degree of discriminative power, resulting in 42 representative statements. Several near-optimal clustering solutions were then cross-referenced to check for any semantic categories marginalized by the optimal solution. Next, two researchers reviewed the initially selected statements, eliminating entries that were highly repetitive within the same or different partitions, covered multiple environmental dimensions, or were vaguely worded, making it difficult to maintain a consistent understanding. Through discussion, statements that both summarized the core meaning of each partition and were distinguishable were retained, ultimately determining 18 partitions and their corresponding 54 statements. Based on this, wording was standardized and redundancy was eliminated. Table 2 provides an example statement for each category. A complete list of statements and categories can be found in Supplementary Table S1.

The P set refers to respondents who are as heterogeneous as possible in Q sorting and are theoretically relevant to the research topic, and they are also called Q participants. A qualified study requires 25–75 participants (24). This study uses nursing homes as the P sample unit. Data on established nursing homes in mainland China were collected through portals such as “Nursing Home Network” (https://www.yanglao.com.cn/). The collected data include basic information about the institutions and environmental images uploaded by the institutions themselves to the website. These environmental images are considered as explicit representations of value orientations and spatial concepts, and the set constitutes the overall statement. Therefore, the Q sorting in this study describes the environmental orientation at the organizational level, rather than the daily experience of residents or staff. The research design using institutions as the P set aims to classify and benchmark the types of nursing home environmental orientations at the institutional level for service governance and public communication scenarios, rather than as a direct auditing tool for daily environmental quality.

Using “nursing homes” as the search keyword, a total of 765 institutions of this type were found on the website. Subsequently, the open-source online data collection tool “Octopus” was used to systematically crawl the website data. After data collection, this study adopted a two-step data cleaning strategy that combined code screening and manual review. The script first completed duplicate detection and quality and relevance screening, and then manual sampling and correction were carried out according to the preset specifications to ensure the validity and consistency of the samples. After cleaning and removing invalid data, 389 valid sample institutions were retained, corresponding to 3,578 images. Basic information about these institutions is shown in Table 3. The P-set comes from 78 cities, with a wide geographical distribution of the participants and diverse economic regions. The included institutions also cover different types and sizes, thus meeting the heterogeneity requirements for sample selection.

Q sorting involves individual participants recording their subjective judgments of Q statements on a Q-sort grid. Unlike previous related studies, this research does not employ the traditional approach of collecting perspectives through expert or public self-reports. Instead, it calculates the semantic similarity between the image-derived texts of nursing homes selected as the P sample and each Q statement to derive their levels of agreement across statements. Statements are then assigned to the Q-sort grid based on these agreement levels, ranging from “least important” (−5) to “neutral” (0) to “most important” (+5). As shown in Figure 4, the Q-sort employs a fixed quasi-normal distribution, allocating scores across 11 levels. Following this process, each selected individual is represented by a Q-sort (74).

In Q methodology, factor analysis is an iterative process in which researchers identify and evaluate multiple factor solutions to find those interpretable as meaningful perspectives (68). However, identifying significant factors requires more than statistical measures; it also entails determining theoretical insights and qualitative considerations consistent with the Q design and dataset (75, 76). Data analysis was conducted using the embedded statistical functions of the online Ken-Q Analysis v2.01 web application (77). First, by-person correlations were computed to construct a correlation matrix; high pairwise correlations indicate overlapping organizational understandings of chronic disease-oriented built-environment practices, with each shared-meaning cluster represented by a distinct factor. Second, principal component analysis (PCA) was implemented on the correlation matrix for factor extraction, aiming to reduce dimensionality while maximizing explained variance (24). The matrix was then rotated (using Varimax as the baseline, with small-angle manual rotation when necessary) to achieve optimal loadings of Q sorts on the factors, thereby enhancing interpretability. Varimax rotation uses statistical criteria to maximize the total variance explained by the factors. Nevertheless, in Q methodology the optimal mathematical solution is not always the most meaningful one—i.e., the solution that best explains and illuminates diverse perspectives. Accordingly, it is recommended to apply Varimax first, followed by manual rotation informed by substantive knowledge obtained during the Q analysis (68). The rotated factor loadings indicate how well each individual/Q-sort aligns with each perspective, while minimizing the number of perspectives with high loadings for the same individual. Subsequently, scores for each statement within each factor are calculated to interpret meanings based on the statements most relevant to each perspective (74). Finally, the perspective framework is refined and clarified based on the z-scores for each factor and the resulting factor arrays. These z-scores and factor arrays reveal the statements ranked highest and lowest within each framework, enabling the naming and interpretation of each perspective. It is important to emphasize that the framework does not represent any single individual's opinion but rather an idealized position collectively presented by individuals representative of that factor (70). The aim is to synthesize dispersed individual rankings into a comparable shared discourse of action, thereby identifying differences and consensus across groups. On this basis, a typology of viewpoints can be constructed, systematically characterizing core insights into chronic disease-oriented built-environment practices according to the statements occupying high and low positions within each type.

The specific factor extraction procedure followed previous research (78), employing PCA to extract and examine one- to eight-factor solutions. Among the candidate solutions, the Kaiser-Guttman criterion was first used to assist in identifying the optimal solution; this statistical standard requires eigenvalues exceeding 1.0 as the threshold for factor inclusion, and each factor must have at least two Q sorts to achieve significant loading (66). The significance threshold (P < 0.01) was determined by dividing 2.58 by the square root of the total number of statements in the Q set (66). With 54 statements in this study, this yielded a threshold of 2.58/√54, ~0.351. Thus, only Q sorts with absolute loadings exceeding 0.35 on a given factor were considered statistically significant. According to the embedded functionality in Ken-Q Analysis v2.01, a Q sort must account for the majority of its communality on the target factor to be considered a defining Q sort for that factor. In addition, the inflection point on the screen plot in Figure 5 was jointly considered. Factor 4 represents a clear inflection point; beyond it, factor eigenvalues decline gradually and contribute little to the cumulative percentage. Applying these screening rules yielded a four-factor solution explaining 86% of the variance in the data (see Table 4).

Varimax rotation was evaluated through a preliminary interpretation of the four-factor solution, examining all Q sorts exhibiting significant loadings on one or more factors. Qualitative data subsequently revealed that three Q sorts—Q sorts 4, 317, and 375—with significant cross-loadings on two factors (borderline cases) actually aligned more closely with Factor 4. Therefore, following the rotation procedure outlined by Watts and Stenner (78), the research team performed a manual adjustment: rotating Factors 1 and 4 counterclockwise by two degrees. This increased the number of defining Q sorts in Factor 4 to five. Beyond these three Q sorts, the manual adjustment had no further impact on the factor structure. Thus, in the final four-factor solution, all 389 Q sorts loaded significantly on one or more factors, with 360 being defining Q sorts—those loading significantly on a factor and accounting for the majority share of their communality. Factor 1 had the highest number of defining Q sorts (180), followed by Factor 2 (135) and Factor 3 (40) (see Table 5).

To deepen understanding of the four perspectives, this study interprets two types of statements: high- and low-ranked statements (the extreme columns of the factor array, per this study's forced distribution of ±4/±5), which delineate what each perspective most emphasizes and most opposes regarding chronic disease-oriented built-environment practices; and distinguishing statements, which identify items ranked significantly higher or lower than in other perspectives, thereby pinpointing key divergences (see Figure 6). Complete factor scores and significant factor loadings for each statement are presented in Supplementary Tables S2, S3.