Abstract

Background:

Artificial intelligence (AI)-assisted education has become an increasingly important instructional model in medical education and raises questions about how it shapes students’ self-directed learning processes and technology adoption.

Objectives and methods:

This study examined factors associated with self-directed learning and AI-related behavioral intention among medical students using a cross-sectional survey design. A model integrating self-directed learning dimensions with TAM- and UTAUT2-related constructs was tested with 600 valid questionnaires, and data were analyzed using partial least squares structural equation modeling (PLS-SEM) in SmartPLS 4.

Results:

The results supported 21 of the 24 hypothesized paths. Motivation emerged as the strongest predictor in the model and was significantly associated with attitude and all three self-directed learning dimensions. Attitude was also significantly associated with self-planning, self-management, and self-monitoring. Self-planning was positively associated with self-management, and self-management was positively associated with self-monitoring. In the technology acceptance pathway, perceived ease of use and perceived usefulness were associated with behavioral intention, and behavioral intention was associated with actual behavior. Facilitating conditions and social influence were also associated with behavioral intention. The model explained substantial variance across key constructs, ranging from 47.6% in actual behavior to 69.9% in self-management.

Conclusion:

These findings suggest that motivational support, structured self-directed planning activities, and adequate digital infrastructure may be relevant considerations for AI integration in health sciences education. The study provides preliminary evidence that a model integrating self-directed learning dimensions with TAM and UTAUT2 related constructs may help explain AI-assisted learning behavior in this population and highlights the need for longitudinal research to clarify the directionality of these associations.

Introduction

Artificial intelligence, a key 21st-century technology, is transforming various fields, including business, science, and healthcare (1), and plays a crucial role in education (2). AI applications, such as ChatGPT and AI-assisted diagnostic tools, are increasingly used in medical education and clinical training across specialties like ophthalmology (3), radiology (4), and residency programs (5). AI offers solutions to educational challenges (6, 7); for example, Fang et al. (5) demonstrated successful use of an AI-based system for detecting pathological myopia in ophthalmology training. By identifying knowledge gaps, providing personalized tutoring, and enhancing feedback, AI-assisted teaching methods may improve both teaching quality and learning outcomes (8–10).

With AI increasingly integrated into medical education, self-directed learning is crucial for medical students (11–13). Defined as an active process where learners set goals and manage their learning (14), self-directed learning is associated with academic performance (15, 16). Students with high self-directed learning skills achieve better results by actively engaging in learning, improving their goal-setting, resource access, and analysis skills (17). In medical education, where knowledge is rapidly updated and lifelong learning is essential, self-directed learning is particularly important for the development of independent professional competence (11, 18).

To examine these factors, this study integrates the Technology Acceptance Model (TAM) and the Unified Theory of Acceptance and Use of Technology 2 (UTAUT2). TAM provides a well-established foundation for examining how perceived usefulness and perceived ease of use shape learners’ behavioral intentions toward AI-assisted learning tools (19). However, as TAM was originally developed to explain individual-level technology acceptance (20), it does not readily account for the broader contextual and social factors that are likely to shape technology adoption in educational settings, such as social influence from peers and instructors and the availability of digital infrastructure. UTAUT2 addresses this limitation by incorporating these contextual constructs explicitly (21). In the context of self-directed learning, which involves both individual-level processes such as goal setting and self-monitoring and contextual conditions that shape how learners engage with learning tools (22), neither framework alone is sufficient; their integration therefore provides a more comprehensive theoretical basis for the present study. In the present study, TAM was used to represent technology-related beliefs, whereas UTAUT2 was used to represent broader contextual influences associated with AI-related behavioral intention.

In addition to technology-related and contextual perspectives, learner-related psychological factors such as motivation and attitude may also be relevant to self-directed learning in AI-assisted educational settings. Accordingly, this study aims to examine the learner-related, technology-related, and contextual factors associated with self-directed learning and AI-related behavioral intention among medical students in a medical university context. Understanding these associations may provide useful evidence for educators and institutions seeking to design AI-assisted learning environments that more effectively support students’ motivational engagement, self-directed learning processes, and technology adoption. Given the increasing integration of AI tools in health sciences education and the growing importance of self-directed learning competencies for future healthcare professionals, such evidence is particularly timely and relevant for informing curriculum design and instructional practice in this field.

Supporting theorySelf-directed learning dimensions

Self-directed learning is both an effective pedagogy and a crucial competency, vital for improved educational outcomes and lifelong learning (23). Knowles defined it as a process where individuals, independently or with support, diagnose learning needs, set objectives, identify resources, implement strategies, and assess outcomes. Building on this, Ginzburg et al. (24) emphasized that self-directed learners must also identify, analyze, and synthesize relevant information, continuously self-regulate their learning, and embrace lifelong learning. Medical students must continually update their knowledge and engage in lifelong learning due to the life-critical nature of medicine (25). Self-directed learning is crucial for this lifelong engagement (26, 27), and improving these skills can boost motivation, reduce burnout, and ease teaching burdens. Self-directed learning is therefore a critical competency for medical students, and examining the learner-related and contextual factors associated with it is of clear importance in medical education.

Technology acceptance perspectives

The Technology Acceptance Model (TAM) explains how users come to accept and use new technologies. According to TAM, perceived ease of use and perceived usefulness shape behavioral intention, which in turn influences actual use behavior (19). External variables may also be associated with perceived ease of use and perceived usefulness, allowing researchers to incorporate additional factors relevant to technology use (28). TAM has been widely applied to examine learners’ acceptance of emerging technologies, including artificial intelligence in educational contexts (29, 30).

The Unified Theory of Acceptance and Use of Technology 2 (UTAUT2) also provides a useful perspective for understanding technology adoption (31). Its UTAUT2 has been shown to offer strong explanatory value in technology adoption research and highlights the importance of behavioral intention and contextual conditions in understanding technology use (21, 32). In the present study, TAM was used to represent technology-related beliefs, whereas UTAUT2 was used to represent broader contextual influences relevant to AI-related behavioral intention. Together, these perspectives provide a methodological basis for examining learner-related, technology-related, and contextual factors within the proposed conceptual framework.

Motivation and attitude

Recognizing the role of psychological factors, especially motivation and attitude, in shaping behavior, the model incorporates these constructs. Drawing on Bandhu’s motivation theory (33), in learning contexts, these factors are important for goal-directed behavior and may influence how learners regulate their learning. Therefore, motivation and attitude were included in the present study as key learner-related constructs.

MethodConceptual framework, research questions, and hypotheses

This study developed a conceptual framework to examine self-directed learning among medical students in AI-assisted learning contexts. The framework draws on selected constructs from the Technology Acceptance Model (TAM), the Unified Theory of Acceptance and Use of Technology 2 (UTAUT2), motivational theory, and the literature on self-directed learning. Specifically, it incorporates learner-related, technology-related, and contextual variables to explain the hypothesized relationships shown in Figure 1. The study adopted a cross-sectional survey design, and all hypotheses were derived from established theoretical frameworks and tested in a confirmatory manner.

Motivation

Self-determination theory, developed by Deci and Ryan, proposes that human behavior is voluntary and self-regulated (34). It also suggests that learning is more effective when driven by intrinsic motivation and supported by autonomy (35). Ntrinsic motivation includes internal factors such as curiosity, interest, and emotion, whereas extrinsic motivation is related to external rewards or reliance on specific techniques or tools (35). Prior literature has further emphasized the importance of motivation in self-directed learning (36). These theoretical perspectives support the inclusion of motivation as an important learner-related factor in self-directed learning. In this study, motivation refers to the psychological drive that encourages students to engage in self-directed learning. Accordingly, motivation was treated as a key learner-related construct in the conceptual framework.

H1a: Motivation positively influences attitude.

H1b: Motivation positively influences self-planning.

H1c: Motivation positively influences self-management.

H1d: motivation positively influences self-monitoring.

Attitude

Attitude refers to learners’ evaluative orientation toward AI-assisted self-directed learning. Prior literature suggests that attitude is associated with how learners approach learning tasks and regulate their learning behaviors (35). In technology-enhanced learning contexts, students’attitudes may also be shaped by their perceptions of the learning environment (37). Accordingly, attitude was included in the present study as an important learner-related construct, and the following hypotheses were proposed:

H2a: Attitude positively influences self-planning.

H2b: Attitude positively influences self-management.

H2c: Attitude positively influences self-monitoring.

Self-directed learning dimensions

Self-directed learning involves learners’ active regulation of their learning processes. Prior literature has described self-directed learning as involving planning, management, and monitoring of one’s own learning activities (38). In this study, self-planning refers to the organization of learning goals and activities, self-management refers to the regulation and adjustment of learning tasks, and self-monitoring refers to the ongoing review and evaluation of learning progress. These dimensions are conceptually related: effective self-planning may support self-management, and self-management may further facilitate self-monitoring. Accordingly, the following hypotheses were propose.

H3a: Self-planning has a positive influence on self-management.

H3b: Self-management positively influences self-monitoring.

TAM model and UTAUT2 model

In the present study, perceived usefulness and perceived ease of use were included as core technology-related constructs derived from TAM. Perceived usefulness refers to the extent to which learners believe that using AI-related learning tools may enhance their learning, whereas perceived ease of use refers to the extent to which such tools are perceived as easy to use (39). Prior studies have suggested that learner-related factors may be associated with students’evaluations of learning technologies (40). Consistent with TAM, perceived ease of use was hypothesized to positively influence perceived usefulness, and both constructs were expected to positively influence behavioral intention, which in turn was expected to positively influence actual behavior (39). This study aligns with Davis’s framework by proposing the following hypotheses:

H4a: Motivation positively influences perceived usefulness.

H4b: Motivation positively influences perceived ease of use.

H4c: Self-planning positively influences perceived usefulness.

H4d: Self-planning positively influences perceived ease of use.

H4e: Self-management positively influences perceived usefulness.

H4f: Self-management positively influences perceived ease of use.

H4g: Self-monitoring positively influences perceived usefulness.

H4h: Self-monitoring positively influences perceived ease of use.

H4i: Perceived ease of use positively influences perceived usefulness.

H4j: Perceived usefulness positively influences behavioral intention.

H4k: Perceived ease of use positively influences behavioral intentions.

H4l: Behavioral intentions positively influences actual behavior.

External environmental factors

In addition to learner-related and technology-related factors, contextual conditions may also shape students’ intention to use AI-related learning tools. In this study, social influence, teacher characteristics, and facilitating conditions were included as contextual variables. Social influence reflects perceived expectations or encouragement from important others, teacher characteristics reflect instructor-related factors relevant to AI-assisted learning, and facilitating conditions refer to the perceived availability of resources and support for technology use (21), Accordingly, the following hypotheses were proposed:

H5a: Social Influences Positively Influence Behavioral Intentions.

H5b: Teacher characteristics positively influence behavioral intentions.

H5c: Facilitating conditions positively influence behavioral intentions.

Integration of structural equation modeling. AT, attitude; MT, motivation; SP, self-planning; SMA, self-management; SMO, self-monitoring; PEU, perceived ease of use; PU, perceived usefulness; BI, behavior intention; AB, actual behavior, SI, social influence; TC, teacher characteristic; FC, facilitating condition.

Participant process

Before the formal survey, a pilot survey involving 151 students was conducted using an online questionnaire, and minor wording revisions were made to improve questionnaire clarity and quality. The formal survey adopted a cross-sectional design and a convenience sampling strategy. An online questionnaire was distributed via QR code in classrooms, small-group learning settings, and other offline contexts at a medical university. Participation was voluntary and anonymous. Eligible participants were students currently enrolled in undergraduate or postgraduate programs at the institution who provided informed consent prior to completion.

Data collection was conducted in three rounds. In the first round, 300 questionnaires were allocated for data quality assessment, and the recovery rate (91.3%) was satisfactory. In the second and third rounds, 300 and 228 questionnaires were collected respectively, bringing the total number of collected questionnaires to 828. For cases with one or two missing item responses within a construct, missing values were imputed using the mean value of the remaining items within the same construct prior to analysis. Questionnaires were excluded from analysis if they contained unanswered items (98 cases), were completed in an unreasonably short time suggesting insufficient engagement (72 cases), or contained logically inconsistent responses identified through attention check items (58 cases). A total of 600 valid responses were finally retained for analysis, yielding an effective response rate of 72.5%, which meets the ‘10-fold rule’ for the minimum sample size for structural equation modeling (41). Ethical approval was obtained from the Ethics Committee of Dalian Medical University (DMU-Ethics-2024-026). All procedures complied with the Declaration of Helsinki, and informed consent was obtained from all participants before questionnaire completion.

Based on the final sample of 600 valid questionnaires, the majority of participants were female (n = 434, 72.3%) and of Han ethnicity (n = 527, 87.8%). Most respondents were aged 18–20 years (n = 405, 67.5%), followed by those aged 21–23 years (n = 152, 25.3%). In terms of academic year, most participants were in grades 1–2 (n = 403, 67.2%), while 143 (23.8%) were in grades 3–5 and 54 (9.0%) were master’s students. Regarding majors, the largest proportion of participants were from Health Services and Management (n = 114, 19.0%), followed by Labor and Social Security (n = 104, 17.3%) and Public Utilities Management (n = 100, 16.7%). The characteristics of the participants are shown in Figure 2.

Demographic characteristics.

Measurement

The research questionnaire comprised 47 items derived from previously validated instruments and adapted to align with the present study’s model and objectives (42–47), including the Self-Directed Learning Scale, the TAM questionnaire, and UTAUT-related items. The questionnaire was organized into two sections: the first section collected basic demographic information, and the second section contained the measurement items. Items were originally developed in English and translated into Chinese by the research team, with wording refined based on pilot survey feedback. All items were assessed using a five-point Likert scale ranging from 1 (“strongly disagree”) to 5 (“strongly agree”), with reverse-scored items recoded prior to analysis. The full item wording, construct–item mapping, and original sources are presented in Appendix 1.

Data processing and analysis

Data were analyzed using a two-stage partial least squares structural equation modeling (PLS-SEM) procedure. Descriptive statistics and correlation analyses were conducted in SPSS 27, structural equation modeling was performed in SmartPLS 4, and all figures were generated in Origin 2024.

In the first stage, the reflective measurement model was evaluated. Internal consistency reliability was assessed using Cronbach’s alpha and composite reliability (CR), with values of 0.70 or higher considered acceptable. Convergent validity was assessed using indicator outer loadings and average variance extracted (AVE), with outer loadings generally expected to exceed 0.70 and AVE values to exceed 0.50 (48). Discriminant validity was assessed using both the Fornell–Larcker criterion and the heterotrait–monotrait ratio (HTMT) (49). In addition to HTMT point estimates, HTMT inference was conducted using bootstrapped confidence intervals, and discriminant validity was considered acceptable when the confidence interval did not include 1 (50).

In the second stage, the structural model was evaluated. Multicollinearity among predictor constructs was assessed using variance inflation factors (VIF), with values below 5 considered acceptable (48). Explanatory power was assessed using the coefficient of determination (R2), while effect sizes (f2) and predictive relevance (Q2) were also examined. Hypotheses were tested using bootstrapping with 5,000 resamples to estimate t-values and bias-corrected 95% confidence intervals for direct effects. Model fit was assessed using the standardized root mean square residual (SRMR), with values below 0.08 indicating acceptable fit. The normed fit index (NFI) was also reported as a supplementary fit index and interpreted cautiously (51).

ResultsMeasurement model assessment

We assessed the reliability and validity of the reflective measurement model. As shown in Appendix 2; Table 1, the indicator outer loadings ranged from 0.771 to 0.929, and all retained indicators were above the recommended threshold. Cronbach’s α values for the multi-item constructs ranged from 0.749 to 0.941, while composite reliability (CR) values ranged from 0.752 to 0.941. These results indicate satisfactory internal consistency reliability of the measurement model. In addition, the average variance extracted (AVE) values ranged from 0.666 to 0.855, all exceeding the recommended threshold of 0.50, thereby supporting convergent validity (Table 1).

Endogenous constructsCronbach’s alphaCRAVESMO0.7490.7520.666SMA0.7760.7770.690TC0.8060.8440.717SP0.8150.8180.730SI0.8250.8460.74PEU0.8320.8510.7501MT0.9170.9190.752FC0.8440.8440.763PU0.8450.8630.764BI0.8750.8760.800AT0.9410.9410.809AB0.9150.920.855

Questionnaire reliability and validity.

Discriminant validity was subsequently assessed using both the Fornell–Larcker criterion and the heterotrait–monotrait ratio (HTMT). As presented in Figure 3, the square roots of the AVE values for all constructs were greater than their correlations with other constructs, indicating acceptable discriminant validity according to the Fornell–Larcker criterion. In addition, HTMT inference was conducted using bootstrap confidence intervals. All HTMT confidence intervals were below 1, further supporting discriminant validity. The detailed HTMT results are provided in Appendix 2; Table 2.

Correlation analysis.

EndogenousQ2 predictR-squareAB0.5490.476AT0.6930.695BI0.4780.551PEU0.5150.580PU0.6060.685SMA0.6260.700SMO0.4930.588SP0.6110.645

Structural model quality (R2, Q2).

Structural model assessment

Following the guidelines established by Hair et al. (41, 48), we conducted a comprehensive assessment of the structural model, focusing on collinearity diagnostics, model fit, explanatory power (R2), predictive relevance (Q2), and effect sizes (f2).

Initially, collinearity within the structural model was examined. The variance inflation factor (VIF) values for all predictor constructs ranged from 1.0 to 4.02. Although some predictors showed relatively elevated VIF values, all values remained below the threshold of 5, suggesting that multicollinearity was within an acceptable range.

Model fit was further assessed using the standardized root mean square residual (SRMR) and the normed fit index (NFI). The SRMR values were 0.053 for the saturated model and 0.110 for the estimated model, while the corresponding NFI values were 0.810 and 0.791, respectively. These results suggest that the saturated model showed acceptable fit, whereas the estimated model fit remained less than ideal and should therefore be interpreted with caution.

Regarding explanatory power, the model explained 47.6% of the variance in actual behavior, 69.5% in attitude, 55.1% in behavioral intention, 58.0% in perceived ease of use, 68.5% in perceived usefulness, 70.0% in self-management, 58.8% in self-monitoring, and 64.5% in self-planning. These R2 values indicate moderate to substantial explanatory power overall. In addition, all Q2 values were greater than zero, supporting the predictive relevance of the model (see Table 2).

The effect size estimates (f2) varied considerably across the structural paths, and were interpreted according to Cohen’s (65) benchmarks: f2 ≥ 0.35 indicates a large effect, f2 ≥ 0.15 a medium effect, f2 ≥ 0.02 a small effect, and f2 < 0.02 a negligible effect. A large effect was observed for the relationship between behavioral intention and actual behavior (f2 = 0.908). Motivation also showed a relatively strong effect on attitude (f2 = 2.281), although this result should be interpreted with caution given the magnitude of the estimate. A moderate effect was found for motivation on self-planning (f2 = 0.227). Most of the remaining paths showed only small or negligible effect sizes, as detailed in Table 3.

Pathf-squareEffect sizeMT → AT2.281LargeBI → AB0.908LargeMT → SP0.226MediumMT → PU0.137SmallAT → SMA0.094SmallAT → SP0.090SmallMT → SMA0.089SmallPEU → PU0.084SmallAT → SMO0.080SmallMT → PEU0.078SmallSP → SMA0.076SmallSP → PU0.070SmallSMA → SMO0.061SmallPU → BI0.060SmallSP → PEU0.043SmallPEU → BI0.038SmallFC → BI0.030SmallSI → BI0.027SmallMT → SMO0.022SmallSMO → PEU0.015NegligibleSMA → PEU0.012NegligibleSMA → PU0.003NegligibleTC → BI0.003NegligibleSMO → PU0.001Negligible

Effect sizes (f2) for path.

Path factor

Regarding the direct structural paths, the results provided support for the proposed model structure. Specific path coefficients, T-statistics, and significance levels are summarized in Table 4.

For H1 and H2, motivation emerged as the most dominant antecedent, showing a large effect on attitude (β = 0.834, p < 0.001, f2 = 2.281) and significant positive associations with all three self-directed learning dimensions (H1b–H1d: all p < 0.05). Attitude further demonstrated significant positive associations with all three self-directed learning dimensions (H2a–H2c: all p < 0.01), supporting H2a through H2c.

Regarding H3, self-planning significantly predicted self-management (β = 0.253, p < 0.001), and self-management in turn significantly influenced self-monitoring (β = 0.280, p < 0.001), confirming a sequential pattern among these components and supporting H3a and H3b.

For the TAM and UTAUT related paths (H4), motivation and self-planning significantly predicted both perceived usefulness and perceived ease of use (H4a–H4d: all p < 0.001). Self-management showed a non-significant negative association with perceived usefulness (H4e: β = −0.052, p = 0.204), and self-monitoring showed a non-significant association with perceived usefulness (H4g: β = 0.029, p = 0.310), leading to the rejection of H4e and H4g. Self-management and self-monitoring both showed significant positive associations with perceived ease of use (H4f: p < 0.05; H4h: p < 0.05). Perceived ease of use significantly predicted perceived usefulness (H4i: β = 0.251, p < 0.001), and both perceived usefulness and perceived ease of use significantly predicted behavioral intention (H4j–H4k: all p < 0.001). Behavioral intention showed a large positive association with actual behavior (H4l: β = 0.690, p < 0.001, f2 = 0.908).

Among the external environmental factors, social influence (H5a: β = 0.160, p < 0.01) and facilitating conditions (H5c: β = 0.196, p < 0.001) both significantly predicted behavioral intention, whereas teacher characteristics did not reach significance (H5b: β = 0.064, p = 0.139), leading to the rejection of H5b.

HypothesesPathβT statisticsp valuesH1a