How smart companions reshape psychological well-being in later life: the role of artificial intelligence technology in mental health during senior years

Abstract

Against the dual backdrop of global population ageing and digital transformation, how to leverage smart eldercare technologies to bridge the “digital divide” and enhance older adults’ psychological well-being has become a pressing issue. Grounded in the Technology Acceptance Model (TAM) and the Unified Theory of Acceptance and Use of Technology (UTAUT), this study innovatively develops a deep hybrid analytical framework that integrates Structural Equation Modeling (SEM) with Artificial Neural Networks (ANN). SEM is employed to validate the linear causal pathways among variables, while ANN is introduced to uncover nonlinear relationships and predictive weights in the data. The SEM results confirm a clear transmission pathway of “technology perception → demand satisfaction → enhancement of well-being,” in which “usage demand” plays a key mediating role. The nonlinear sensitivity analysis based on ANN further reveals that perceived usefulness and usage demand are the primary driving factors for predicting older adults’ well-being, indicating that their value orientation toward technology has shifted from “usable” to “truly useful.” The findings suggest that future product design should move away from mere functional stacking toward precise value alignment and zero–cognitive-load interaction, thereby providing scientific evidence for achieving high-quality, proactive healthy ageing.

Introduction

Against the backdrop of accelerating global population ageing, “smart eldercare” has evolved from a forward-looking concept into a national strategy for addressing structural challenges in society [1]. This transformation concerns not only material care, but also more profoundly touches upon the inner world of older adults. Traditional research has predominantly focused on physical health and social support in later life; however, mental health and subjective well-being, as core dimensions for evaluating the quality of life in old age, are undergoing a paradigmatic reconfiguration in the era of deeply embedded intelligent technologies. In the process of modernization, shrinking family size, increasing population mobility, and rapid social change are placing unprecedented strain on traditional support systems, rendering loneliness, social isolation, and cognitive decline the “three specters” haunting later life and severely eroding older adults’ psychological well-being. Consequently, the search for innovative and scalable solutions to maintain and enhance the psychological welfare of older adults has become an urgent issue in global public health. As AI-driven companionship, monitoring, and services become normalized, a central academic question emerges: through what specific pathways do intelligent technologies influence and reshape older adults’ psychological states and sense of well-being? Clarifying this complex mechanism is crucial for optimizing technology design, formulating precise eldercare policies, and achieving a genuine integration of technological empowerment and humanistic care.

In recent years, rapid advances in artificial intelligence (AI) have opened up new avenues for smart eldercare. Smart companions (such as smart speakers, virtual assistants, and social robots), as key AI applications in the eldercare domain, are gradually entering older adults’ everyday lives [2]. Compared with traditional physical care and social support, smart companions not only provide instrumental assistance [3] and health monitoring [4], but also exhibit unique advantages in emotional companionship [5], cognitive stimulation [6], and social interaction [7]. Existing studies indicate that, through voice interaction, emotion recognition, and personalized services, smart companions can alleviate loneliness and enhance older adults’ subjective well-being [8]. However, there is still no systematic consensus in the scholarly community regarding how smart companions specifically affect older adults’ mental health, including their underlying mechanisms, pathways of influence, and boundary conditions.

The existing literature mainly concentrates on three aspects: first, the acceptance and usage intention of intelligent technologies, examining older adults’ attitudes toward AI products and their determinants [9]; second, the positive role of smart companions in mitigating loneliness and promoting social participation [10]; and third, the measurement and intervention pathways related to mental health and subjective well-being [11]. Nevertheless, systematic literature reviews reveal that current research is largely confined to technical or macro-level descriptions and lacks a complete analytical chain tracing how older adults’ initial perceptions of technology, under the influence of external environmental drivers, translate into the satisfaction of specific usage needs and ultimately into experiential outcomes such as life satisfaction and psychological security. This study seeks to extend this field by constructing and empirically validating a multi-dimensional, multi-level mechanism model. Moreover, structural changes in contemporary society have weakened traditional social support networks, exacerbating problems of loneliness, social isolation, and cognitive decline among older adults. Against this backdrop, intelligent technologies are widely expected to provide new solutions to the mental health challenges faced by older adults. Yet it remains unclear through which mechanisms these technologies influence older adults’ subjective well-being; existing studies lack multi-dimensional, mechanism-oriented empirical analyses. In light of the above research progress and practical needs, this paper addresses the following research questions: (1) Which antecedent factors in the adoption of intelligent technologies significantly affect older adults’ subjective well-being? (2) Through which mediating mechanisms are these effects realized, and are there nonlinear or interaction effects? (3) How can a multi-dimensional, multi-level mechanism model be constructed and validated to reveal the pathways through which intelligent technologies shape older adults’ mental health?

The innovations of this study are mainly reflected in the following aspects. First, it moves beyond prior analytic paradigms centered on single variables or purely linear paths, and develops a multi-dimensional theoretical model of how smart companions influence older adults’ psychological well-being. By introducing new perspectives such as “relational technology” and “cognitive enhancement medium,” it enriches the theoretical landscape at the intersection of AI and mental health. Second, the study adopts a combined approach of structural equation modeling (SEM) and artificial neural networks (ANN), simultaneously testing theoretical hypotheses and uncovering nonlinear relationships and key sensitivity factors among variables. While SEM elucidates direct and indirect causal relationships, ANN helps capture complex dynamics that traditional models struggle to detect, thereby providing a more fine-grained analytical tool for mechanism-oriented research. Third, drawing on empirical data, the paper proposes tailored intervention recommendations for different subgroups of older adults, offering both theoretical grounding and practical guidance for smart eldercare product design and policy formulation. Through detailed analysis of the pathways of influence, this study provides a novel analytical lens and methodological framework for understanding how older adults’ mental health is being reshaped in the AI era.

Literature review and research hypotheses

Smart elderly care

The global population structure is undergoing profound transformation, with aging becoming one of the most significant social trends of the 21st century. Along with this trend, mental health problems among older adults have become increasingly salient, with social isolation and loneliness widely recognized as “silent epidemics” that undermine the quality of later life [12]. Research has confirmed that prolonged loneliness is not only strongly associated with mental disorders such as depression and anxiety, but also substantially increases the risk of cognitive decline, cardiovascular diseases, and all-cause mortality [13]. Traditional family support and social care networks are challenged by the processes of modernization, thus making the search for innovative and scalable solutions to safeguard and enhance the mental well-being of older adults an urgent global public health priority [14]. Against this backdrop, Smart Elderly Care has emerged, aiming to leverage emerging technologies to support older adults’ health, independence, and social engagement [15]. Early technological applications largely focused on functional support such as physiological monitoring, safety alerts, and daily assistance [16]. However, in recent years, scholarly attention has gradually shifted toward the potential of technology to meet older adults’ emotional and social needs [17]. In particular, the rise of Artificial Intelligence (AI) has opened new possibilities for providing personalized and interactive emotional companionship [18]. Intelligent technologies are no longer perceived merely as tools, but increasingly as social entities capable of partially assuming the role of “companions [19].”

In the realm of conversational AI, systematic reviews in healthcare contexts show that dialogue agents can provide psychoeducation, emotional support, and behavioral change guidance, demonstrating notable potential in managing symptoms of depression and anxiety [20]. Although randomized controlled trials specifically targeting older populations remain limited, voice-based interactions and multimodal inputs have lowered the barriers of digital accessibility, suggesting considerable transferability and scalability for promoting mental health in later life [21, 22]. Meanwhile, “Ambient Assisted Living” (AAL) systems—built upon environmental sensing and adaptive learning—continuously monitor activity patterns and identify anomalies, thereby enhancing safety and independence, while indirectly alleviating anxiety and caregiver burden [23]. Related reviews indicate that activity recognition, context inference, and personalized reminders constitute the core technological pathways for achieving “invisible companionship [24].” From a technology adoption perspective, critical antecedents include perceived usefulness, usability, perceptions of privacy and autonomy, and the social normative environment, all of which are closely tied to the psychological benefits experienced by older adults [25]. Broader evidence from ICT interventions has further shown that digital social connection can moderately reduce social isolation and loneliness, thus providing important external validation for the mental health value of AI companionship [26].

Therefore, this study seeks to bridge the aforementioned gaps by combining quantitative and qualitative approaches to investigate in depth how long-term use of AI companionship technologies (particularly smart elderly care platforms) within the home environment reshapes the psychological well-being of ordinary older adults. Specifically, it focuses on analyzing the mechanisms and subjective experiences through which these technologies alleviate loneliness and enhance well-being, aiming to offer deeper theoretical insights and practical guidance for the future design and application of smart elderly care solutions.

Psychological well-being in older adulthood

Subjective well-being (SWB) in older adulthood is a key psychological indicator for assessing quality of life in later life. It typically encompasses life satisfaction, the presence of positive affect, and the relative absence of negative affect [11]. With the accelerating pace of global population ageing, enhancing the psychological well-being of older adults has become an important objective in public health and social governance. A large body of empirical research has demonstrated that psychological well-being not only affects older adults’ physical and mental health, but is also closely associated with their cognitive functioning, social participation, and longevity [27]. Traditional studies have mainly examined the determinants of late-life SWB from the perspectives of social support, economic status, physical health, and family relationships. Social support is regarded as a core resource for enhancing older adults’’ well-being, effectively alleviating loneliness and social isolation [28]. Physical health is positively associated with well-being, whereas chronic illnesses and functional impairments significantly diminish older adults’ subjective well-being [29]. In addition, economic security, harmonious family relationships, and active social participation have all been shown to exert important positive effects on older adults’ well-being [30].

With the rise of artificial intelligence and digital health technologies, a growing number of studies have begun to focus on the impact of technological interventions on older adults’ psychological well-being. Koh et al. have pointed out that social robots and virtual companions can substantially enhance older adults’ subjective well-being and life satisfaction through emotional companionship, cognitive stimulation, and social interaction [31]. Abdi et al. further found that intelligent companions not only alleviate loneliness, but also foster older adults’ self-identity and sense of social belonging via personalized dialogue and emotion recognition [32]. Jung and colleagues reported that improving digital literacy among older adults can effectively enhance their life satisfaction and subjective well-being [33].

Although existing research has provided preliminary evidence for the positive role of intelligent companions in promoting older adults’ psychological well-being, the underlying mechanisms have not yet been systematically elucidated. The current literature predominantly focuses on technology acceptance and usage intention, and lacks integrative analyses of multidimensional mechanisms. Moreover, important issues remain underexplored, including the differential effects of various types of intelligent companions and the dynamic relationships between technology adoption and psychological well-being. Therefore, there is an urgent need to construct a multidimensional, multilevel theoretical model that systematically reveals the pathways through which intelligent companions reshape older adults’ psychological well-being, and to combine longitudinal data with multi-method empirical approaches to advance theoretical and practical innovation in this field.

Technology acceptance model (TAM) and psychological well-being

To understand how artificial intelligence (AI) companion technologies affect older adults’ psychological well-being, this study goes beyond examining whether technologies are “adopted” and further focuses on how technology use shapes individuals’ psychological states and digital well-being. The Technology Acceptance Model (TAM), proposed by Davis (1989), is a classic framework for predicting users’ behavioral intentions [34]. It posits that behavioral intention is primarily determined by two key beliefs: perceived usefulness (PU), that is, the extent to which an individual believes that using a particular technology will improve their performance or quality of life; and perceived ease of use (PEOU), that is, the extent to which an individual believes that learning and using the technology will be effortless. With its concise structure and strong explanatory power, TAM has been widely applied in fields such as healthcare [35], e-government, and e-commerce [36]. However, traditional research has largely remained at the level of the “efficiency improvement → usage intention” pathway, and has rarely examined in a direct manner the mechanisms through which technology use influences psychological well-being.

In the context of older adults, the connotation of “usefulness” needs to be expanded in light of their developmental tasks. Socioemotional selectivity theory suggests that as people age, their goals gradually shift from knowledge acquisition to emotion regulation and affective fulfillment. When older adults perceive that AI companion technologies can provide emotional support, health monitoring, and assistance with daily living, such “usefulness” not only reflects instrumental utility, but also constitutes a form of compensation for socioemotional resources and self-care capabilities. This, in turn, helps enhance their sense of security, autonomy, and control over life, thereby improving psychological well-being. Accordingly, we propose:

• H1: The perceived usefulness of smart eldercare platforms is positively associated with older adults’ psychological well-being.

For older adults who are widely affected by the “digital divide” and technological anxiety, perceived ease of use refers not only to operational simplicity, but also serves as a crucial source of self-efficacy and perceived environmental control. Declines in cognitive capacity and past experiences of failure with technology often undermine older adults’ confidence when facing new technologies, thereby reducing their life satisfaction [37]. When older adults believe that they can “learn” and “successfully use” a smart eldercare platform, a smooth human–computer interaction experience can strengthen their sense of control and competence with respect to their environment, which is highly consistent with the “environmental mastery” dimension in Ryff’s model of psychological well-being [38]. Ease of use reduces the frustration and helplessness associated with complex operations, enabling older adults to experience positive emotions and enhanced self-esteem through the feeling that “I am capable of mastering modern technology.” Therefore, perceived ease of use should not be viewed merely as an intermediary variable leading to usage behavior; it is itself a psychological resource that can reduce technological anxiety, enhance perceived control, and directly improve psychological well-being. Accordingly, we propose:

• H2: The perceived ease of use of smart eldercare platforms is positively associated with older adults’ psychological well-being.

Unified theory of acceptance and use of technology (UTAUT) and psychological well-being

The Unified Theory of Acceptance and Use of Technology (UTAUT), which integrates multiple technology acceptance theories such as TAM and TPB, has been widely applied in fields including health information systems, telecare, and gerontechnology to explain the adoption and continued use of digital technologies among different populations [9]. In the context of older adults, UTAUT has been used not only to predict behavioral intention and actual use, but has also been increasingly extended to outcomes related to “digital well-being,” such as subjective well-being, life satisfaction, and loneliness. UTAUT comprises four core constructs: performance expectancy, effort expectancy, social influence, and facilitating conditions. Conceptually, performance expectancy and effort expectancy closely correspond to perceived usefulness (PU) and perceived ease of use (PEOU) in TAM. Given that PU, PEOU, and their pathways have already been systematically elaborated in the preceding section based on TAM, this study, in order to avoid conceptual and measurement redundancy, retains in the UTAUT framework only two contextual factors that are more directly related to older adults’ psychological states—social influence and facilitating conditions—and introduces “usage needs” as a key psychological mediator, thereby constructing an integrated explanatory framework.

Social influence refers to the expectations and evaluations of important others (e.g., family members, peers, caregivers) that individuals perceive regarding their use of a particular technology. UTAUT posits that social influence significantly shapes individuals’ adoption intentions [39]. Research on older adults shows that recommendations, demonstrations, and ongoing encouragement from family members and caregivers, on the one hand, enhance older adults’ trust in and identification with smart eldercare platforms, and on the other hand, convey social signals of “being cared for” and “being valued.” These signals help alleviate negative feelings such as “being forgotten” or “being a burden,” strengthen social connectedness and sense of belonging, and thereby improve psychological well-being [40]. At the same time, positive social norms can reduce shame and feelings of deviance when older adults use AI companionship technologies, which is conducive to maintaining self-esteem and self-worth [41]. Thus, social influence not only improves well-being “indirectly” by promoting technology use, but also constitutes an important socioemotional resource that can “directly” enhance older adults’ psychological well-being. Accordingly, we propose:

• H3: The social influence of smart eldercare platforms is positively associated with older adults’ psychological well-being.

Facilitating conditions refer to the degree of organizational and technical support available to individuals when using a technology, such as training resources, hardware maintenance, help channels, and accessible design. Research in environmental gerontology and gerontechnology indicates that adequate external support can alleviate the mismatch between “environmental demands” and “individual capabilities,” reduce feelings of helplessness and frustration, and enhance perceived control, self-efficacy, and independence, thereby contributing to psychological well-being [42]. In the context of smart eldercare platforms, when communities, eldercare institutions, or families provide systematic training, timely technical assistance, and stable network and device conditions, older adults are better able to overcome technological anxiety arising from “not knowing how to use” or “being afraid to use.” During the learning and usage process, they can experience a sense of competence—“I can master new things”—and perceive institutional and environmental safety nets [43]. This not only reduces frustration and isolation caused by technological barriers, but also enhances digital inclusion, thereby promoting participation in online social interactions, health management, and leisure activities, which in turn improves overall psychological well-being. Based on this, we propose:

• H4: The facilitating conditions of smart eldercare platforms are positively associated with older adults’ psychological well-being.

However, social influence and facilitating conditions do not translate into well-being in a direct or mechanical manner; their effects are realized through older adults’ subjective evaluations of whether their “usage needs” are being met. Studies on digital well-being among older adults suggest that technologies can generate sustained positive welfare effects only when they achieve a good “needs–technology fit” with individuals’ key life needs—such as safety, health management, emotional companionship, social connection, and self-realization [44]. Drawing on need satisfaction theory and the “uses and gratifications” perspective, this study defines “usage needs” as older adults’ overall perception of the extent to which a smart eldercare platform can fulfill their multi-level usage purposes. When family members and caregivers actively recommend and assist in using the platform (high social influence), and when communities or institutions provide sufficient training and technical support (favorable facilitating conditions), older adults are more likely to recognize, in practice, the platform’s concrete value in areas such as safety alerts, health follow-up, and emotional comfort, and thus form the cognition that “this platform genuinely meets my actual needs.” This experience of need satisfaction not only increases the perceived meaning and value of use, but also directly enhances psychological well-being by fulfilling needs for safety, belonging, and self-efficacy. Therefore, the effects of social influence and facilitating conditions on psychological well-being are expected to be partly realized through increases in older adults’ subjective perception of “the degree to which their usage needs are fulfilled” by the platform. Based on this, we propose:

• H5: Usage needs play a mediating role in the relationship between smart eldercare platforms and older adults’ psychological well-being.

• H5a: Usage needs mediate the relationship between perceived usefulness of smart eldercare platforms and older adults’ psychological well-being.

• H5b: Usage needs mediate the relationship between perceived ease of use of smart eldercare platforms and older adults’ psychological well-being.

• H5c: Usage needs mediate the relationship between social influence of smart eldercare platforms and older adults’ psychological well-being.

• H5d: Usage needs mediate the relationship between facilitating conditions of smart eldercare platforms and older adults’ psychological well-being.

In sum, by integrating TAM and UTAUT, this study highlights two key social–environmental factors—social influence and facilitating conditions—and, through the psychological mechanism of “usage needs,” links technology acceptance research in the information systems field with interdisciplinary literature on gerontechnology, digital well-being, and positive psychology. As shown in Fig. 1, the theoretical model of this study aims to elucidate the complete pathway through which smart eldercare platforms move from “supported use” to “needed use,” and ultimately to “enhanced psychological well-being.”

Fig. 1.

Research hypothesis diagram

Research methodology and data collection

Research method

This study adopts a mixed research approach that combines Covariance-Based Structural Equation Modeling (CB-SEM) and Artificial Neural Networks (ANN), with the aim of examining, from the dual perspectives of “linear relational explanation” and “nonlinear predictive ranking,” how older adults’ use of smart eldercare platforms influences their well-being. The integration of these two methods allows their respective strengths to be complementary, thereby providing a more comprehensive and accurate depiction of the complex relationships among variables, as illustrated in Fig. 2.

Fig. 2.

Data analysis procedure

In the first stage, this study employs Covariance-Based Structural Equation Modeling (CB-SEM) to test the proposed hypotheses. CB-SEM, rather than Partial Least Squares Structural Equation Modeling (PLS-SEM), is chosen for the following reasons. First, the primary objective of this study is to examine a theoretically grounded model and its associated path hypotheses derived from technology acceptance theory and well-being theory. The emphasis is thus on theory testing and model fit, rather than purely on prediction, which aligns more closely with the theory-driven and confirmatory nature of CB-SEM [45]. Second, all measurement variables in this study are modeled as reflective indicators, making CB-SEM appropriate for the joint estimation of the measurement and structural models. Third, the sample size satisfies the medium-to-large sample requirements of CB-SEM, thereby supporting stable estimation of overall model fit and parameters in a covariance-based SEM framework. By contrast, PLS-SEM is more prediction-oriented and variance-explanatory, and is generally advantageous in situations with small sample sizes or highly exploratory models. Accordingly, CB-SEM is prioritized for the structural equation component of this study, with ANN subsequently used to strengthen predictive performance. CB-SEM analyses are conducted using IBM SPSS Statistics 26.0 and IBM SPSS Amos 26.0.

Although CB-SEM performs well in theory testing, it typically assumes linear relationships among variables and may fail to capture the nonlinear patterns that are common in real-world settings. To address this limitation, the present study introduces Artificial Neural Networks (ANN) as a complementary analytical tool and uses SPSS 26.0 to construct a Multilayer Perceptron (MLP) model. As a powerful data-driven and machine learning algorithm, ANN simulates the information-processing mechanisms of human neurons and excels at uncovering complex, nonlinear relationships in data, particularly in predictive tasks [46]. In the ANN model, perceived ease of use, perceived usefulness, social influence, facilitating conditions, and usage demand are specified as input-layer neurons, while well-being serves as the output-layer neuron, thus forming an MLP architecture. By partitioning the dataset (typically into training, validation, and test sets), the neural network is trained and optimized to develop a model capable of accurately predicting older users’ well-being.

In summary, this study first applies CB-SEM to test and validate the theory-based research framework and to clarify the causal pathways among variables. It then employs an ANN model to further probe and predict these relationships and to rank the relative importance of the influencing factors. By combining the “explanatory power” of CB-SEM with the “predictive power” of ANN, the study offers a deeper and more robust understanding of the underlying mechanisms through which smart eldercare platforms affect the well-being of older users.

Scale variables and measurement items

To ensure the scientific rigor and validity of the measurement instruments in the specific context of smart eldercare, this study followed a stringent scale development procedure consisting of “theoretical review – content validity assessment – pilot testing.” First, the initial questionnaire was primarily constructed on the basis of classical theoretical frameworks, including the Technology Acceptance Model (TAM), the Unified Theory of Acceptance and Use of Technology (UTAUT), and Subjective Well-being (SWB). Perceived Ease of Use (PEOU) and Perceived Usefulness (PU) were adapted from the scales developed by Davis [34]; Social Influence (SI) and Facilitating Conditions (FC) were drawn from the UTAUT model proposed by Venkatesh [47]; Usage Requirements (UR) were designed by reference to functional categorizations of smart eldercare services. The construct of “Well-being (WB)” was informed by Diener’s [48] Satisfaction With Life Scale (SWLS) and Ryff’s [38] dimensions of psychological well-being.

Given that the target respondents in this study are older adults and that the overall number of questionnaire items is relatively large, we sought to reduce cognitive load and avoid fatigue effects. Therefore, instead of adopting full-length, multidimensional scales, we employed a “core-indicator-focused” strategy. Specifically, we selected three key dimensions most directly affected by smart eldercare technologies—life satisfaction (cognitive evaluation), psychological security (emotional experience), and global subjective evaluation (overall perception)—with the aim of precisely capturing the enhancement of a technology-empowered form of well-being, rather than a broad, general sense of happiness. The main part of the questionnaire adopted a five-point Likert scale, ranging from 1 (“strongly disagree”) to 5 (“strongly agree”).

To establish the content validity of the scales, we implemented the following steps. First, three professors in geriatric psychology and two experts in information systems were invited to review the initial draft of the questionnaire. The expert panel examined the semantic clarity of each item and its correspondence with the underlying construct, removed ambiguously worded items, and thereby ensured the face validity of the instrument. Second, prior to the formal survey, a small-scale pilot test was conducted with 50 older users. Feedback from the pilot indicated that some abstract descriptions related to “well-being” were difficult for respondents to answer. In response, we revised these items to emphasize concrete feelings “specifically attributable to the use of the platform” (as shown in Table 1).

Table 1.

Questionnaire design

Variable Dimension	Explicit Variable	Measurement Item
Perceived Ease of Use (PEOU)	Simple operation	Using the smart senior care platform is simple and easy to understand.
	Standalone use	Even without assistance, I can operate the smart senior care platform independently.
	User-friendly interface	I find the interface design of the smart senior care platform user-friendly and easy to navigate.
Perceived Usefulness(PU)	Improve health management	I believe using a smart senior care platform can improve my health management.
	Enhance quality of life	Using a smart senior care platform can enhance my daily quality of life.
	Communicate with caregivers	A smart senior care platform can help me communicate more effectively with doctors or caregivers.
	Monitor health data	It allows for more efficient monitoring of my health data (such as heart rate and blood pressure).
Social Influence(SI)	Family encouragement	My family encourages me to use smart senior care platforms.
	Peer recommendations	My friends or peers recommend that I use smart senior care platforms.
	Surrounding usage	The experiences of people around me (such as neighbors or community members) using smart senior ca. re platforms influence my perception of these services.
Facilitating Conditions (FC)	Possess the necessary equipment	I possess the necessary devices (such as smartphones, tablets, etc.) to use the smart elderly care platform.
	Possess the knowledge and skills	I believe I have sufficient knowledge and skills to use the smart elderly care platform.
	Internet support	I have reliable internet connectivity available to me, facilitating my use of the smart elderly care platform.
Usage Requirements (UR)	Online Medical Consultation	I hope the smart senior care platform can connect to hospitals or doctors to provide online consultations or appointment scheduling services.
	Health Monitoring Devices	I hope the smart senior care platform can remind me to take my medication on time or schedule health checkups.
	Smart Home Devices	I hope the smart senior care platform can help me monitor smart home devices.
	Social Entertainment Apps	I need the smart senior care platform to offer social features that help me stay connected with family and friends.
Well-being (WB)	Life Satisfaction	After using the smart senior care platform, life feels more convenient and fulfilling.
	Psychological Safety	After using the smart senior care platform, it has sparked positive anticipation for future living.
	Subjective Well-Being	After using the smart senior care platform, the subjective assessment of overall happiness in current life.

Survey participants and data collection

Sampling rationale and inclusion criteria

This study defines the target population as older adults aged 60 and above residing in mainland China who have experience using, or express a clear intention to use, smart eldercare platforms. In view of the specific characteristics of the older population and the practical difficulties in accessing them, a non-probability sampling strategy combining convenience sampling and snowball sampling was adopted. To ensure the validity and internal consistency of the sample, we established stringent inclusion criteria: (1) age ≥ 60 years; (2) basic reading and comprehension ability, with the capacity to complete the questionnaire independently or with minimal assistance; (3) actual operational experience with a smart eldercare platform within the past six months, or a clear understanding of its functions; and (4) informed and voluntary participation. Individuals with severe cognitive impairment or acute illnesses that prevented them from cooperating with the survey were excluded.

To mitigate potential representativeness bias arising from non-probability sampling, this study implemented a “multi-dimensional stratified quota control” mechanism. Rather than relying on a single recruitment channel, we constructed a heterogeneous recruitment network encompassing community-based eldercare centers, Universities of the Third Age (U3A), outpatient departments of tertiary hospitals, and vertical online communities for older adults. During data collection, the research team continuously adjusted quotas in real time based on demographic characteristics (e.g., gender, age group) and social-structural characteristics (e.g., living arrangement, city tier). Specifically, we sought to maintain a balanced sample structure across dimensions such as younger-old (60–69 years) versus oldest-old (≥ 80 years), and living alone/empty nest versus co-residing with family, in order to approximate the distribution of the overall population as closely as possible and enhance the external validity of the study.

In addition, we consulted precedents from mainstream survey research to justify our sample size. Prior studies commonly report sample sizes in the range of 250–500 participants: for example, Jain et al. surveyed 242 undergraduates, postgraduates, and faculty members regarding their use of generative AI [49]; Adu-Gyamfi et al. surveyed 405 Chinese electric vehicle users [50]; Tang et al. surveyed 400 rural tourists [51]; Asamoah et al. surveyed 469 respondents in China [52]; and another study by Tang et al. conducted an online survey of 392 consumers with live-stream shopping experience [53]. In light of these benchmarks, the 321 valid responses used in this study are deemed sufficient for accurate prediction and statistical inference.

Data collection and sample characteristics

Data collection for this study was conducted between January and June 2025. To address the commonly observed “survivor bias” in ageing research (i.e., samples skewed toward older adults with high levels of digital literacy) and the broader “digital divide” problem, we innovatively adopted a mixed-mode data collection strategy that combined “online self-administration” with “offline assisted administration.” All participants provided written informed consent, and strict anonymization procedures were followed throughout the data collection process, as illustrated in Fig. 3.

Fig. 3.

Flowchart of the data collection process

• Online pathway: A digital questionnaire was created using the “Wenjuanxing” platform and distributed in a targeted manner to relevant communities via community key opinion leaders (KOLs) and managers of eldercare institutions, with the aim of capturing data from active users with relatively high levels of digital literacy.

• Offline pathway: A professionally trained fieldwork team was dispatched to communities and eldercare service centers to support older adults facing digital access barriers. For these participants, a “face-to-face structured interview” approach was used to assist with questionnaire completion. This procedure not only ensured the authenticity of the data, but also effectively incorporated vulnerable groups who are often marginalized in conventional online surveys.

After rigorous data cleaning and exclusion of invalid responses, a total of 321 valid questionnaires were retained for analysis. The demographic profile of the sample indicates good heterogeneity and ecological validity: gender distribution was relatively balanced (56% male, 44% female); the age structure exhibited a typical pyramidal pattern (60–69 years: 61%; 70–79 years: 30%; ≥80 years: 9%), consistent with the current demographic reality in China where the “younger-old” constitute the majority. In terms of living arrangements, 48% of respondents were living alone or in “empty-nest” households, while 52% were co-residing with family members (spouses and/or children), reflecting both the trend toward smaller family structures and the diversification of living arrangements in an ageing society. These characteristics provide a representative empirical basis for the subsequent model validation.

SEM analysis and results

Measurement model assessment

The reliability and convergent validity results presented in Table 2 indicate that the factor loadings of all observed variables corresponding to the latent constructs (PEOU, PU, SI, FC, UR, WB) range from 0.897 to 0.943, all exceeding the critical threshold of 0.70. The Cronbach’s α coefficients (0.928–0.968) and composite reliability (CR) values (0.928–0.968) are all above 0.90, demonstrating excellent internal consistency and strong reliability of the measurement instruments. The average variance extracted (AVE) values lie between 0.812 and 0.883, well above the acceptable benchmark of 0.50, indicating that each latent construct explains a substantial proportion of the variance in its observed indicators. In sum, the measurement model in this study exhibits outstanding reliability and convergent validity, thereby satisfying the prerequisite conditions for subsequent empirical analyses.

Table 2.

Reliability and convergent validity test results of each latent variable

Construct	Item	Estimate	Cronbach’s α	CR	AVE
PEOU	PEOU 1	0.908	0.881	0.928	0.812
	PEOU 2	0.898
	PEOU 3	0.897
PU	PU 1	0.912	0.917	0.955	0.840
	PU 2	0.920
	PU 3	0.909
	PU 4	0.925
SI	SI 1	0.929	0.909	0.947	0.856
	SI 2	0.918
	SI 3	0.929
FC	FC 1	0.923	0.909	0.942	0.844
	FC 2	0.914
	FC 3	0.919
UR	UR 1	0.936	0.945
	UR 2	0.943		0.968	0.883
	UR 3	0.943
	UR 4	0.937
WB	WB 1	0.935	0.915	0.945	0.851
	WB 2	0.904
	WB 3	0.928

PEOU Perceived Ease of Use, PU Perceived Usefulness, SI Social Influence, FC Facilitating Conditions, UR Usage Requirements, WB Well-Being

Table 3 shows that the KMO value is 0.900, far exceeding the critical threshold of 0.70, indicating that the observed variables share sufficient common factors and that the data are suitable for factor analysis. The Bartlett’s test of sphericity yields a chi-square statistic of 4988.409 with 190 degrees of freedom and a p-value of 0.000 (p < 0.001), leading to the rejection of the null hypothesis that the correlation matrix is an identity matrix and confirming significant correlations among the variables. Therefore, the data used in this study exhibit good suitability and meet the prerequisite conditions for conducting factor analysis.

Table 3.

KMO and Bartlett’s inspection

KMO		0.900
Bartlett’s sphericity	spherical test	4988.409
	df-value	190
	p-value	0.000

Common method bias

In terms of statistical assessment, this study employed Harman’s single-factor test to evaluate the severity of potential common method bias (CMB), a classical approach for detecting CMB. Specifically, all measurement items in the questionnaire were subjected to exploratory factor analysis (EFA), using unrotated principal component analysis to extract common factors, in order to determine whether serious common method bias was present in the collected data. The principal component analysis extracted a total of 20 principal factors, with the first factor accounting for 36.844% of the variance. This value is well below the 50% critical threshold proposed by Podsakoff et al. [54], indicating that no single factor dominates the explanation of variance in the data. These results provide preliminary evidence that the risk of common method bias in this study is relatively low.

Assessment of discriminant validity

The discriminant validity results reported in Table 4 (Fornell–Larcker matrix) and Table 5 (HTMT matrix) show that, in the Fornell–Larcker matrix, the square roots of AVE on the diagonal for each latent construct (WB = 1.520, UR = 1.896, FC = 1.507, SI = 1.718, PEOU = 1.400, PU = 1.675) are all greater than the off-diagonal correlation coefficients between that construct and other latent variables (ranging from 0.253 to 0.816), thereby satisfying the criterion that “the square root of AVE should exceed the inter-construct correlations.” In the HTMT matrix, all HTMT values between latent variables fall within the range of 0.142–0.562, which are below the critical threshold of 0.85, indicating clear discriminant boundaries among constructs and the absence of serious multicollinearity problems. Taken together, these findings suggest that the measurement model exhibits satisfactory discriminant validity; the conceptual independence and distinctiveness of each latent construct are well supported, meeting the requirements for subsequent empirical analyses.

Table 4.

Discriminant validity (Fornell–Larcker Matrix)

	WB	UR	FC	SI	PEOU	PU
WB	1.520
UR	0.761	1.896
FC	0.669	0.345	1.507
SI	0.571	0.256	0.344	1.718
PEOU	0.547	0.313	0.816	0.253	1.400
PU	0.786	0.465	0.333	0.467	0.365	1.675

PEOU Perceived Ease of Use, PU Perceived Usefulness, SI Social Influence, FC Facilitating Conditions, UR Usage Requirements, WB Well-Being

The on-diagonal items in bold represent the square roots of the AVE; off-diagonal elements are the correlation

Table 5.

Discriminant validity (HTMT Matrix)

	WB	UR	FC	SI	PEOU	PU
WB
UR	0.448
FC	0.442	0.204
SI	0.353	0.142	0.214
PEOU	0.375	0.192	0.562	0.163
PU	0.492	0.261	0.209	0.275	0.239

PEOU Perceived Ease of Use, PU Perceived Usefulness, SI Social Influence, FC Facilitating Conditions, UR Usage Requirements, WB Well-Being

Goodness of fit indices of the structure model

The goodness-of-fit results for the confirmatory factor analysis model in Table 6 indicate that all fit indices meet the corresponding evaluation criteria. Among the absolute fit indices, CMIN/DF = 1.846 (< 3), GFI = 0.915 (> 0.90), AGFI = 0.885 (meeting the more lenient criterion of ≥ 0.85), and RMSEA = 0.051 (within the acceptable threshold of ≤ 0.08). For the incremental fit indices, NFI = 0.944, TLI = 0.967, IFI = 0.974, CFI = 0.973, and RFI = 0.931, all exceeding the recommended cutoff value of 0.90. Regarding the parsimony fit indices, PCFI = 0.794, PNFI = 0.770, and PGFI = 0.676, all surpassing the acceptable benchmark of 0.50. Taken together, these results demonstrate that the confirmatory factor analysis model exhibits good overall fit, and that the data align well with the theoretical model, thereby providing a sound basis for subsequent empirical analyses.

Table 6.

Model fit for confirmatory factor analysis

Category	Fit Index	Fit Criterion	Model Result	Model Fit Status
Absolute fit indices	CMIN/DF	< 3	1.846	Yes
	GFI	> 0.9	0.915	Yes
	AGFI	≥ 0.85 (lenient criterion) / > 0.90 (ideal criterion)	0.885	Yes (lenient criterion)
	RMSEA	≤ 0.08 (acceptable) / ≤ 0.05 (ideal)	0.051	Yes (acceptable)
Incremental fit indices	NFI	> 0.9	0.944	Yes
	TLI	> 0.9	0.967	Yes
	IFI	> 0.9	0.974	Yes
	CFI	> 0.9	0.973	Yes
	RFI	> 0.9	0.931	Yes
Parsimonious fit indices	PCFI	≥ 0.5	0.794	Yes
	PNFI	≥ 0.5	0.770	Yes
	PGFI	≥ 0.5	0.676	Yes

Hypothesis testing

The model estimation results presented in Table 7; Fig. 4 indicate that the path coefficients for all hypothesized relationships are positive, and the key statistical indicators meet the criteria for significance. The unstandardized path estimates range from 0.139 to 0.306, with all standardized regression coefficients being significantly positive. The critical ratios (C.R.) all exceed the conventional threshold for statistical significance. In terms of p-values, those for H1, H3, H4, and H5c are 0.014, 0.007, 0.001, and 0.005, respectively (all < 0.05), while the p-values for H2, H5, H5a–H5b, and H5d are all below 0.001 (denoted by ***). These results indicate that perceived ease of use (PEOU), perceived usefulness (PU), social influence (SI), and facilitating conditions (FC) not only exert significant positive direct effects on well-being (WB), but also have significant positive indirect effects on WB via usage demand (UR). Overall, all research hypotheses are fully supported, and the structural paths in the model are statistically meaningful.

Table 7.

Model measurement results

Hypothesis				Estimates	S.E.	C.R.	P	Testing the hypothesis
H1	WB	<---	PEOU	0.170	0.069	2.455	0.014	Established
H2	WB	<---	PU	0.222	0.058	3.836	***	Established
H3	WB	<---	SI	0.139	0.051	2.702	0.007	Established
H4	WB	<---	FC	0.209	0.066	3.188	0.001	Established
H5	WB	<---	UR	0.222	0.058	3.862	***	Established
H5a	UR	<---	PEOU	0.288	0.077	3.718	***	Established
H5b	UR	<---	PU	0.278	0.063	4.398	***	Established
H5c	UR	<---	SI	0.163	0.058	2.819	0.005	Established
H5d	UR	<---	FC	0.306	0.072	4.239	***	Established

PEOU Perceived Ease of Use, PU Perceived Usefulness, SI Social Influence, FC Facilitating Conditions, UR Usage Requirements, WB Well-Being

*** indicates p < 0.001

Fig. 4.

Structural equation model diagram

Mediation analysis of usage demand

The mediation analysis results in Table 8 show that all four paths with usage demand (UR) as the mediating variable pass the significance test (all p-values < 0.05), and the 95% confidence intervals (Lower–Upper) for the indirect effects do not include zero. This indicates that UR exerts a significant partial mediating effect on the relationships between perceived ease of use (PEOU), perceived usefulness (PU), social influence (SI), facilitating conditions (FC), and well-being (WB).

Table 8.

Mediation effect analysis

Mediation Path	Effect	Estimate	Lower	Upper	P
PEOU—UR—WB	Indirect Effect	0.059	0.020	0.122	0.002
	Direct Effect	0.156	0.001	0.301	0.047
	Total Effect	0.214	0.053	0.377	0.005
PU—UR—WB	Indirect Effect	0.060	0.017	0.137	0.001
	Direct Effect	0.216	0.098	0.340	0.001
	Total Effect	0.277	0.146	0.405	0.001
SI—UR—WB	Indirect Effect	0.038	0.010	0.090	0.009
	Direct Effect	0.146	0.010	0.282	0.036
	Total Effect	0.183	0.049	0.307	0.007
FC—UR—WB	Indirect Effect	0.066	0.023	0.143	0.001
	Direct Effect	0.204	0.032	0.381	0.016
	Total Effect	0.271	0.113	0.437	0.001

PEOU Perceived Ease of Use, PU Perceived Usefulness, SI Social Influence, FC Facilitating Conditions, UR Usage Requirements, WB Well-Being

*** indicates p < 0.001

Specifically, the estimated indirect effects along the four paths are 0.059, 0.060, 0.038, and 0.066, respectively, while the corresponding direct effects are 0.156, 0.216, 0.146, and 0.204. The total effects range from 0.183 to 0.277. These results suggest that PEOU, PU, SI, and FC not only have significant positive direct effects on WB, but also promote WB indirectly by enhancing UR. Together, the direct and indirect effects constitute the overall mechanisms through which these antecedent variables influence well-being.

Construction of the ANN neural network model

As shown in Fig. 5, to further optimize the model structure and enhance predictive accuracy, this section integrates the path analysis results of Structural Equation Modeling (SEM) with the nonlinear fitting capabilities of Artificial Neural Networks (ANN), thereby constructing a SEM–ANN–based integrated model of older adults’ well-being. Drawing on the analytical framework proposed by Liébana et al. [55], and in accordance with the variable relationships validated by SEM and the hierarchical mapping principles of ANN, a two-stage neural network architecture is designed. Model A uses perceived ease of use (PEOU), perceived usefulness (PU), social influence (SI), and facilitating conditions (FC) as input-layer variables, with usage demand (UR) as the output layer. Model B then takes the aforementioned four antecedent variables together with the mediator UR as input-layer variables and well-being (WB) as the output layer, thereby forming a hierarchical prediction mechanism spanning “antecedent variables–mediator–outcome variable.”

Fig. 5.

Artificial neural network model construction

To avoid model overfitting and assess generalization capability, this study adopts 10-fold cross-validation to train and test the ANN models. The dataset is first split into a training set (90%) and a test set (10%) at a 9:1 ratio. The training set is then randomly partitioned into 10 subsets; in each iteration, 9 subsets are used for training and 1 subset for validation, and this process is repeated 10 times until all subsets have been used for validation. An early stopping mechanism is also implemented: the mean squared error (MSE) on the validation set is monitored, and training is automatically terminated if the validation MSE does not decrease for 10 consecutive epochs, with the current best model parameters retained to prevent degradation in generalization performance due to overtraining. The hyperparameters of the model are configured as follows: Sigmoid activation function, Adam optimizer, learning rate of 0.001, 10 nodes in the hidden layer, and a maximum of 100 training epochs.

Root mean square error test

This study evaluates model predictive performance using two complementary indicators: Mean Squared Error (MSE) and Root Mean Squared Error (RMSE). MSE amplifies the impact of prediction deviations through squared errors, making it more sensitive to observations with large errors; RMSE shares the same unit of measurement as the dependent variable and intuitively reflects the average deviation between predicted and actual values. Taken together, these two metrics provide a comprehensive assessment of model fit. In the subsequent results analysis, both the magnitude and fluctuation range of these indicators will be considered to make an integrated judgement regarding model stability and predictive accuracy. The formulas are as follows:

1

2

In both formulas, denotes the true value, the predicted value, and the sample size.

The root mean squared error (RMSE) results of the artificial neural network (ANN) models reported in Table 9 indicate that the RMSE values for both the training and test sets are at relatively low levels for Model A (inputs: PEOU, PU, SI, FC; output: UR) and Model B (inputs: PEOU, PU, SI, FC, UR; output: WB). Specifically, for Model A, the average RMSE is 0.281 on the training set (standard deviation, SD = 0.155) and 0.269 on the test set (SD = 0.225); for Model B, the average RMSE is 0.229 on the training set (SD = 0.172) and 0.212 on the test set (SD = 0.237). A comparison shows that the average RMSE values for both the training and test sets of Model B are lower than those of Model A, and the test-set RMSE of Model B (0.212) is notably lower than that of Model A (0.269). This suggests that, after incorporating the mediating variable UR, Model B achieves higher predictive accuracy for older adults’ well-being (WB). Meanwhile, although some fluctuations are observed in the RMSE standard deviations of both models, they remain within a reasonable range overall, indicating that the models exhibit good stability and generalization capability and can effectively capture the nonlinear relationships among variables.

Table 9.

Root mean square error test for artificial neural network models

	Model A		Model B
	Input: PEOU, PU, SI, FC, UR		Input: PEOU, PU, SI, FC, UR
	Output: WB		Output: WB
Neural network	Training	Testing	Training	Testing
ANN1	0.303	0.252	0.240	0.212
ANN2	0.321	0.147	0.295	0.268
ANN3	0.252	0.320	0.232	0.145
ANN4	0.300	0.307	0.210	0.218
ANN5	0.277	0.265	0.225	0.117
ANN6	0.244	0.252	0.203	0.262
ANN7	0.274	0.280	0.227	0.285
ANN8	0.276	0.323	0.208	0.176
ANN9	0.294	0.256	0.255	0.180
ANN10	0.264	0.283	0.194	0.254
Mean	0.281	0.269	0.229	0.212
SD	0.155	0.225	0.172	0.237

PEOU Perceived Ease of Use, PU Perceived Usefulness, SI Social Influence, FC Facilitating Conditions, UR Usage Requirements, WB Well-Being

Sensitivity analysis

This study adopts Garson’s algorithm to compute the relative importance of each input variable. The core principle is to decompose the connection weights of the neural network in order to quantify the contribution of each input variable to the output variable. The specific steps are as follows: (1) extract the weight matrices between the input and hidden layers, and between the hidden and output layers of the ANN model; (2) calculate the absolute contribution value of each input variable by summing the absolute products of its corresponding input–hidden and hidden–output weights; and (3) normalize the contribution values of all input variables (i.e., contribution of a given variable / total contribution of all variables × 100%) to obtain the normalized relative importance (%). This method effectively reflects the predictive priority of each variable within a nonlinear model, and the results are reproducible.

The normalized importance analysis of the ANN models presented in Table 10 shows that, in Model A (with UR as the output), the ranking of input variables in terms of their normalized relative importance for predicting usage demand (UR) is: perceived ease of use (PEOU, 100.000%) > facilitating conditions (FC, 97.193%) > perceived usefulness (PU, 87.368%) > social influence (SI, 65.965%). Among these, PEOU and FC are the core determinants of UR, with importance values close to or reaching 100%. In Model B (with WB as the output), the ranking of input variables in terms of their normalized relative importance for predicting well-being (WB) is: perceived usefulness (PU, 100.000%) > facilitating conditions (FC, 85.551%) > usage demand (UR, 80.228%) > social influence (SI, 61.977%) > perceived ease of use (PEOU, 52.852%). PU emerges as the most critical predictor of WB, followed closely by FC and UR. Overall, social influence (SI) exhibits the lowest normalized importance in both models, whereas PEOU, PU, and FC play central roles in influencing UR and WB, respectively. The mediating variable UR also shows a relatively high predictive importance in Model B, further confirming its key mediating role in the “antecedent variables → UR → WB” pathway. These findings provide quantitative evidence to inform the subsequent development of targeted strategies for enhancing older adults’ well-being.

Table 10.

Analysis of the importance of normalisation in artificial neural network models

	Model A (Output: UR)				Model B (Output: WB)
Neural network	PEOU	PU	SI	FC	PEOU	PU	SI	FC	UR
ANN1	0.197	0.241	0.115	0.447	0.094	0.401	0.035	0.170	0.300
ANN2	0.184	0.257	0.225	0.334	0.198	0.288	0.093	0.326	0.096
ANN3	0.343	0.254	0.212	0.191	0.157	0.237	0.161	0.236	0.209
ANN4	0.358	0.264	0.124	0.254	0.096	0.321	0.247	0.157	0.179
ANN5	0.232	0.262	0.195	0.310	0.168	0.232	0.184	0.210	0.206
ANN6	0.301	0.239	0.186	0.273	0.136	0.207	0.163	0.268	0.226
ANN7	0.398	0.209	0.217	0.175	0.118	0.297	0.215	0.183	0.187
ANN8	0.256	0.230	0.178	0.336	0.146	0.209	0.172	0.241	0.232
ANN9	0.233	0.285	0.276	0.206	0.142	0.196	0.157	0.241	0.264
ANN10	0.350	0.253	0.155	0.242	0.133	0.242	0.200	0.214	0.212
Average relative imporance	0.285	0.249	0.188	0.277	0.139	0.263	0.163	0.225	0.211
Normanlized relative importance (%)	100.000	87.368	65.965	97.193	52.852	100.000	61.977	85.551	80.228

PEOU Perceived Ease of Use, PU Perceived Usefulness, SI Social Influence, FC Facilitating Conditions, UR Usage Requirements, WB Well-Being

Results and discussion

Drawing on a hybrid analysis using SEM and ANN, this study validates and refines a causal chain linking “technological characteristics → fulfillment of usage needs → enhancement of psychological well-being.” Specifically, perceived ease of use (PEOU), perceived usefulness (PU), and facilitating conditions (FC) not only exert direct effects on older adults’ psychological well-being (WB), but also generate significant indirect effects by enhancing usage needs (UR). Among these, facilitating conditions serve as a key antecedent that triggers initial use, whereas perceived usefulness functions as the core variable sustaining continued use and converting it into well-being gains. The nonlinear results from the ANN further indicate that the relative importance of different factors varies across the technology use lifecycle: in the introductory phase, priority should be given to lowering technical barriers and strengthening external support; in the deep-use phase, the emphasis should shift to the technology’s actual value in addressing core issues such as health problems and loneliness. These findings provide evidence-based support for reorienting relevant policies from merely “promoting technology adoption” to “advancing well-being realization.” In other words, policy instruments should balance the facilitation of initial uptake with the pursuit of long-term benefits, rather than relying solely on metrics such as device distribution or platform activation rates as primary performance indicators.

Focusing on the dual orientation of “needs–value”: from “functional stacking” to “precise value alignment”

Drawing on the path coefficients from the SEM analysis and the variable importance rankings from the ANN, this study proposes technology optimization pathways around two key nodes: stimulating usage needs (UR) and enhancing well-being (WB). The findings indicate that the “advancement” of the technology itself is not the primary concern for older adults; rather, the crucial issue is whether it can accurately match their core needs (UR). Accordingly, the development of intelligent companions should abandon the strategy of “maximal and comprehensive” functional accumulation and instead focus on a dual-core orientation of medical value and emotional value. First, policy should be leveraged to promote the deep integration of medical, eldercare, and community services into the operational processes of such platforms, so that the platform functions not merely as an information display tool but as a “service hub” connecting family doctors, community health centers, and eldercare institutions. In this way, perceived usefulness can be translated into tangible improvements in daily life. Second, in response to the widespread experience of loneliness among older adults, a large body of empirical research has shown that when technological functions are highly aligned with older adults’ personally meaningful life goals, technology adoption becomes more stable and can substantially improve subjective well-being and depressive symptoms [40, 56]. Social robots and intelligent companions are particularly effective in reducing loneliness and enhancing positive affect when they address emotional and social needs [57]. Intelligent companions should therefore reinforce their role in facilitating social connection, not only by providing high-definition video communication channels with family members, but also by incorporating virtual companionship functions with affective computing capabilities. When the system can detect emotional fluctuations in older adults and respond with human-like verbal comfort or conversational prompts, the technology transcends its purely instrumental nature and becomes a “digital companion” with emotional soothing functions. Such systems have a markedly stronger effect on emotional stability and life satisfaction among older adults than static, rule-based systems [58].

Enhancing perceived ease of use: building a “zero cognitive load” natural interaction ecology

The empirical results confirm that perceived ease of use (PEOU) is the “threshold factor” for older adults to develop usage intention (UR). Given age-related declines in physiological functioning and fluid intelligence, any operation with a high cognitive load (such as multi-level menus or complex touch gestures) is likely to trigger technological anxiety and obstruct the pathway through which well-being is enhanced. A substantial body of research has demonstrated that operational complexity and learning costs are major barriers to older adults’ adoption of digital health technologies and social robots [59, 60]. Therefore, it is first necessary—through standards and regulatory mechanisms—to incorporate age-friendly interaction principles such as large fonts, high contrast, voice-first interaction, multimodal feedback, and minimal procedural steps into industry norms and procurement thresholds for smart eldercare products. Second, a “proactive service” logic should be introduced. To further lower the operational threshold, system design should leverage machine learning to anticipate user habits—for example, automatically broadcasting weather and medication reminders in the morning, or automatically adjusting lighting and security modes in the evening. This “invisible” mode of technological presence enables older adults to enjoy the benefits of technology without a strong sense of interacting with it, thereby alleviating “technophobia” and fostering a sense of control and satisfaction through a relaxed user experience. Randomized controlled trials on social robots designed to enhance older adults’ well-being have shown that when robotic interventions produce clear, short-term improvements in daily life (such as better sleep, emotional relief, or increased social interactions), older adults significantly increase their subjective evaluation of the robot’s value and their willingness to use it over the long term [61].

Optimizing facilitating conditions: enhancing service accessibility and convenience

“Facilitating conditions” go far beyond merely providing devices or internet access; their core lies in systematically removing all potential barriers that older adults may encounter throughout the entire process of technology use, thereby creating a convenient and trustworthy usage ecosystem. First, the government should work together with enterprises to establish a city-wide, unified smart eldercare hotline and a network of offline service centers. Whether it is device malfunction, questions about software operation, or requests for emergency assistance, problems should be solvable through a single phone call or one on-site visit. Empirical studies have shown that small-scale interventions combining telephone hotlines, in-home digital guidance, and peer tutoring can significantly enhance older adults’ digital self-efficacy, reduce the frustration associated with discontinuing use, and indirectly alleviate depression and loneliness [62, 63]. Second, universal “age-friendly” design standards for smart eldercare products (such as large fonts, simplified interfaces, and voice interaction as the default mode) should be promoted, alongside stringent regulations on the protection of older adults’ personal information and data privacy. When older adults perceive that systems are easy to operate and that their privacy is fully respected and protected, their sense of trust and willingness to use these technologies can be fundamentally strengthened. In addition, efforts should be made to achieve interoperability among different services—such as healthcare, domestic services, transportation, and e-government—through unified identity authentication and standardized interfaces. This would enable older adults to access multiple services via a single terminal, reducing the burden of switching between multiple accounts and complex interfaces. Existing empirical evidence indicates that platform-based eldercare technology solutions that integrate services across domains not only increase usage frequency but also significantly improve subjective quality of life and social participation [64].

Conclusions

Limitations and future directions

Despite the advances achieved in this study in terms of theoretical model construction and methodological innovation, several limitations remain and should be addressed in future research.

First, there are substantial constraints regarding sample recruitment and representativeness. The data in this study were drawn primarily from community-dwelling older adults in urban areas and selected county-level regions, using a non-probability sampling strategy that combined convenience sampling and snowball sampling, with self-administered questionnaires as the main mode of data collection. While this design is feasible and operationally advantageous at an exploratory stage, it also entails a pronounced risk of selection bias: the sample may over-represent older adults who already possess relatively high levels of digital literacy, enjoy stronger social support networks, hold more positive attitudes toward new technologies, and have some prior usage experience; conversely, those with limited digital access, weaker social ties, and more vulnerable economic or health conditions are likely to be under-represented. Consequently, the findings of this study are primarily applicable to older adults whose characteristics are similar to those of the present sample—namely, older individuals living in urban or town communities with at least some opportunities for technology exposure—and should not be directly generalized to the broader older population, particularly not to rural residents, institutionalized older adults, individuals with severe functional impairments, or more isolated and disadvantaged groups. Future research, where conditions permit, should strive to adopt stratified random sampling or other probability-based sampling strategies, and intentionally include digitally marginalized and high-risk older populations (e.g., very old adults living alone, those with low education, low income, or partial/total disability). Weighted analyses or stratified analyses could be employed to enhance external validity and the generalizability of the results. At the same time, multi-source data (such as platform backend behavioral data and community registration records) may be used to correct for sample bias and conduct sensitivity analyses.

Second, this study relies on cross-sectional data, which makes it difficult to capture the dynamic processes underlying the relationship between technology adoption and well-being among older adults. Future studies are encouraged to employ longitudinal designs or experimental/intervention studies to dynamically examine the causal relationships and long-term effects among technology acceptance, changing needs, and improvements in well-being, thereby enabling a more comprehensive understanding of temporal ordering and the underlying mechanism chains among these variables. Third, although this study combines SEM and ANN, thereby balancing theory testing with data-driven prediction, there remains room for improvement in terms of variable selection and model complexity. Future research could incorporate additional dimensions, such as mental health, social support networks, and digital literacy, or employ more sophisticated analytical tools—such as multilevel structural equation modeling (MSEM) or mixed-methods approaches—to explore the interactions among different influencing factors and their heterogeneous effects across subgroups.

Furthermore, as artificial intelligence and smart eldercare technologies continue to evolve, older adults’ attitudes toward, and needs for, emerging technologies are also undergoing constant change. Future research should pay closer attention to the dynamic psychological and behavioral adaptation processes of older adults in the context of ongoing technological iteration, and explore how intelligent and personalized service design can better meet their diverse and differentiated well-being needs. Despite the aforementioned limitations, this study nonetheless provides an initial theoretical foundation and empirical reference for understanding the associative mechanisms between the adoption of smart eldercare services and the enhancement of well-being among community-dwelling older adults in urban settings. Looking ahead, with continual refinement of research methods and diversification of data sources, both theory and practice in this field are expected to become richer and more nuanced, thereby offering stronger support for achieving healthy ageing and the high-quality development of smart eldercare.

Conclusions

By innovatively constructing a hybrid analytical framework that integrates Structural Equation Modeling (SEM) and Artificial Neural Networks (ANN), and by combining the Technology Acceptance Model (TAM) with the Unified Theory of Acceptance and Use of Technology (UTAUT), this study systematically unpacks the complex mechanisms through which smart eldercare services enhance the well-being of urban older adults. At the level of underlying mechanisms, the study confirms a logical chain of “technology perceptions → usage demand → psychological well-being.” SEM results indicate that perceived ease of use, perceived usefulness, social influence, and facilitating conditions are not only antecedents of technology adoption, but also exert significant direct and indirect effects on older adults’ well-being by stimulating their intrinsic “usage demand.” Among these factors, perceived usefulness (PU) plays a dual driving role: it directly increases adoption intention, while indirectly enhancing well-being by meeting health and daily living needs. At the same time, social influence and facilitating conditions are identified as key environmental variables in building a positive technological ecosystem, providing essential social support and material guarantees for older adults’ adaptation to technology.

Methodologically, the triangulation of SEM and ANN reveals the nonlinear nature of variable importance. The in-depth mining capability of the ANN model complements the limitations of SEM’s linear analysis by employing sensitivity analysis to generate a precise ranking of the antecedent variables. The results show that perceived usefulness (PU) and usage demand (UR) are the most critical indicators for predicting older adults’ well-being, with normalized importance values substantially higher than that of perceived ease of use (PEOU). This finding suggests that once older adults have crossed the initial “ease-of-use threshold,” the decisive factors for their well-being become whether the technology can effectively address concrete pain points (usefulness) and satisfy deeper needs (demand–technology fit. Although social support and environmental factors rank lower in relative importance, their foundational role in alleviating technological anxiety and lowering entry barriers remains indispensable.

In terms of theoretical and practical contributions, this study not only validates an integrated model through dual methodological testing, but also theoretically clarifies the internal mechanism by which technology is transformed into psychological well-being via “need satisfaction.” This offers a new perspective for overcoming the traditional separation between technological attributes and psychological outcomes in prior research. Practically, the findings provide a scientific basis for realizing “high-quality digital living” in an ageing society: the optimization of smart eldercare services should not be confined to interface simplification (ease of use), but should instead focus on the creation of core value (usefulness) and the precise alignment with individualized needs.

Authors’ contributions

B.L.: conceptualization, methodology, supervision, writing—original draft, writing—review and editing, software, investigation, formal analysis. Y.L.: validation, software, investigation. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by 2025 Zhejiang Provincial Philosophy and Social Sciences Planning “Provincial and Municipal Cooperation” Project: “A Study on the Construction and Implementation Pathways of the Wenzhou Smart Elderly Care Service System, Driven by Design and Empowered by Digital Intelligence” (25SSHZ071YB); Research on Aged-friendly Interaction Design Strategies and Practical Paths of Smart Products under the Perspective of Active Aging (RC202443).

Data availability

The data used and analyzed during the current study are not publicly available due to ethical and privacy reasons, but are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the ethical principles outline in the Declaration of Helsinki. This study was approved by the Human Research Ethics Committee of Donghua University(RLSSZYJ202501100057), and all participants provided written informed consent.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.