Self-Reported AI Competency and Educational Needs for Integrated AI and Digital Rehabilitation Education Among Occupational Therapy Students: A Cross-Sectional Survey

Abstract

Background and Aim

The growing use of generative artificial intelligence (AI) in healthcare calls for occupational therapy education that combines critical appraisal, clinical reasoning, and digital rehabilitation. This study aimed to examine students’ interest in AI, perceived AI competency, and needs for integrated AI and digital rehabilitation education, and identified factors associated with educational content needs.

Participants and Methods

An exploratory cross-sectional survey using convenience sampling and five researcher-developed scales without full psychometric validation was conducted among occupational therapy students from three universities. A total of 250 questionnaires were submitted, and 230 complete responses were analyzed. Descriptive statistics, reliability tests, Welch group comparisons, Pearson correlations, and hierarchical multiple regression with sensitivity analyses were used.

Results

Mean scores were 3.64±0.77 for AI interest, 3.42±0.67 for self-reported AI competency, 3.93±0.68 for educational content needs, 3.95±0.63 for educational priority, and 3.82±0.64 for instructional method preference. All five domains were positively correlated. Educational priority and instructional method preference showed the strongest correlation (r=0.701). AI interest (β=0.490, p<0.001) and self-reported AI competency (β=0.154, p=0.015) were associated with educational content needs in the conventional regression; the final model explained 34.3% of the variance. The competency association was sensitive to robust inference (HC3 p=0.0504).

Conclusion

Students expressed preferences for integrated, practice-oriented education addressing AI use, occupational therapy clinical reasoning, client-appropriate tool selection, assessment, and intervention planning.

Introduction

Artificial intelligence (AI) and digital technologies have rapidly expanded across healthcare and rehabilitation and are being used in diverse forms, including telerehabilitation, rehabilitation robotics, virtual and augmented reality, wearable sensors, AI-based assessment and motion analysis, and education and counseling support using generative AI. The World Health Organization (WHO) emphasizes that digital technologies can improve the accessibility, efficiency, and sustainability of healthcare services and states that safety, transparency, accountability, privacy protection, and human-centered decision-making should also be considered in the use of AI.^1,2 Accordingly, healthcare professionals are increasingly expected not only to use digital technologies but also to evaluate and apply them clinically and responsibly. AI refers broadly to computational systems that perform tasks such as prediction, classification, or generation; generative AI produces new content such as text or images. Digital rehabilitation uses digital technologies to support rehabilitation and may or may not involve AI. For example, AI-based motion analysis incorporates AI, whereas telerehabilitation, computerized cognitive rehabilitation, VR/AR, robotics, and wearables do so only when AI functions are specifically implemented.

Occupational therapy is a healthcare profession that supports people who experience difficulties in occupational performance due to illness or disability in re-engaging in meaningful and necessary occupations. Occupational therapists plan assessment and intervention by considering not only physical, cognitive, and psychosocial functions but also clients’ habits and roles, living environments, and social participation.³ These characteristics increase the potential for integrating AI and digital technologies that can analyze diverse functional and environmental information and support individualized intervention.⁴ In particular, as occupational therapy services expand from hospitals to homes and communities in response to population aging, increases in chronic diseases and dementia, growing mental health concerns, and strengthened community-based services,^5,6 the ability of occupational therapists to understand and clinically apply digital rehabilitation technologies is becoming increasingly important.

The recent proliferation of generative AI tools such as ChatGPT and Gemini has also brought substantial changes to higher education. Generative AI can be used to search for learning materials, summarize content, support writing, generate ideas, and solve problems. In healthcare education, its applications are expanding to the development of clinical cases, organization of assessment data, exploration of intervention ideas, preparation of educational materials for clients and caregivers, and development of counseling scenarios.^7,8 In occupational therapy education, the potential educational use of generative AI for constructing clinical cases and exploring intervention ideas has also been suggested.^9,10

However, increased experience with generative AI does not necessarily mean that students will naturally acquire the competence to use it appropriately in clinical contexts. AI-generated outputs may include factually incorrect information, poorly supported claims, hallucinations, and bias, and may not be appropriate for a client’s functional level, condition, environment, or treatment goals.⁹ In healthcare, uncritical use of AI-generated content may lead to inappropriate clinical judgments or the provision of incorrect information. Therefore, occupational therapy students need not only prompt-writing and information-search skills but also the ability to evaluate the accuracy and reliability of AI outputs, identify errors and bias, and judge the appropriateness of outputs based on professional knowledge and clinical evidence.¹¹ They also need the competence to use AI responsibly while considering ethical issues such as privacy, copyright, plagiarism, and client safety.^2,7

Current occupational therapy education focuses on professional knowledge and clinical performance skills, which are essential for developing the core competencies required of occupational therapists.¹² However, to respond to the rapidly changing AI and digital healthcare environment, education should extend beyond basic generative AI use to connect diverse digital rehabilitation technologies—including AI-based assessment and motion analysis, computerized cognitive rehabilitation, rehabilitation robotics, VR/AR, wearable sensors, and telerehabilitation—with actual occupational therapy cases.^13,14 In particular, students need systematic education in reviewing and revising AI-generated or AI-analyzed outputs according to occupational therapy knowledge and clinical reasoning and in selecting AI and digital tools appropriate to client characteristics and treatment goals.

Although studies examining perceptions, readiness, intention to use AI, and educational needs among healthcare students have increased,^8,15,16 research that comprehensively investigates occupational therapy students’ interest in AI-based occupational therapy, current self-reported AI competency, specific educational content needs, priorities for AI and digital rehabilitation tools they wish to experience directly, and preferred instructional methods remains limited.^17,18 It is therefore necessary to identify the AI competencies in which occupational therapy students perceive relative difficulty, the educational content and digital tools they consider important, and the relationships among these domains. Specifically, the relationship between OT students’ self-reported AI competency and their perceived needs and preferences for education that also addresses digital rehabilitation remains unclear. Examining these topics together reflects their potential use within the same clinical workflow, rather than assuming a single competency. As OT contributes to multidisciplinary rehabilitation, these findings may generate questions for other healthcare programs; their applicability requires separate investigation.

Accordingly, this study aimed to assess occupational therapy students’ interest in AI-based occupational therapy, self-reported AI competency, educational content needs, educational priorities, and instructional method preferences, and to analyze differences according to participant characteristics and relationships among the major domains. In addition, by identifying factors associated with educational needs for AI-integrated occupational therapy education, this study sought to provide foundational data for curriculum development that goes beyond basic generative AI use and incorporates occupational therapy knowledge and clinical reasoning, critical appraisal of AI outputs, ethics and safety, and practical application of digital rehabilitation tools.

Materials and Methods

Study Design

This study was an exploratory cross-sectional survey designed to investigate self-reported generative AI competency and educational needs for integrated AI and digital rehabilitation education among occupational therapy students. The survey assessed experience with generative AI, interest in AI-based occupational therapy, self-reported AI competency, educational content needs, priorities for digital rehabilitation tool education, and preferences for instructional methods. A structured self-report questionnaire was developed based on relevant previous studies and international guidelines to collect data appropriate to the study objectives.

Participants and Data Collection

This study was conducted from 1 July to 10 August 2026 using convenience sampling among first- to fourth-year occupational therapy students enrolled at three universities in Cheongju, Busan, and Ulsan, South Korea. The minimum required sample size was calculated using G*Power 3.1. For hierarchical multiple regression, assuming a medium effect size (f²=.15), a significance level of α=0.05, power (1−β)=.95, and seven predictors, the minimum required sample was 153. Considering incomplete responses and possible attrition, 250 students were recruited. The universities were selected based on accessibility to the research team and the feasibility of recruitment.

Before participation, students were informed about the purpose and procedures of the study, the research and statistical use of their responses, privacy and confidentiality, voluntary participation, and the right to withdraw at any time. Electronic informed consent was obtained from students who voluntarily agreed to participate before they proceeded to the online questionnaire. Students without prior experience using AI or digital rehabilitation tools were instructed to respond based on their current perceptions. A total of 250 questionnaires were submitted. Of these, data from 230 students who provided complete responses for the five main domains—AI interest, self-reported AI competency, educational content needs, educational priority, and instructional method preference—as well as sex were included in the final analysis. Thus, 20 responses (8.0%) were excluded, and 92.0% of received responses were analyzed; this proportion is not an institutional response rate. All 230 cases had complete year, sex, AI-use frequency, and main-domain item data. Departmental administrative assistants distributed the survey link. The recruitment message explained that the findings could inform future curriculum development and encouraged participation. No incentives were offered, and nonparticipation carried no disadvantage. Twenty submitted questionnaires were excluded because of incomplete responses.

Measures

The questionnaire was developed by the researchers based on relevant previous studies and international guidelines to assess occupational therapy students’ self-reported AI competency and educational needs for integrated AI and digital rehabilitation education. Item development was informed by the Medical Artificial Intelligence Readiness Scale for Medical Students (MAIRS-MS),¹⁹ which conceptualizes AI readiness among medical students in terms of cognition, ability, vision, and ethics, and by the concept of AI literacy,²⁰ which describes the core competencies needed to understand, critically evaluate, and effectively use AI technologies. Items addressing generative AI experience, educational use, and instructional needs were informed by a previous study of rehabilitation sciences students.²¹ Items related to privacy, accuracy, bias, accountability, and safe use were based on UNESCO’s AI competency framework for students²² and WHO guidance on ethics and governance of AI for health.¹

After the initial questionnaire was developed, two occupational therapy experts with more than 10 years of experience as occupational therapy faculty reviewed the content validity, clarity of wording, relevance to occupational therapy education, and appropriateness of item composition within each domain. Based on their feedback, ambiguous or redundant wording was revised, and items concerning the integration of occupational therapy knowledge and clinical reasoning, selection of AI and digital tools appropriate to client characteristics and treatment goals, priorities for digital rehabilitation tool education, and instructional methods were modified and refined. The final questionnaire comprised participant characteristics and AI-use experience, AI interest, self-reported AI competency, educational content needs, digital tool educational priority, and instructional method preference. These framework-informed, researcher-developed domains are used for exploratory description; expert review and Cronbach’s α do not establish unidimensionality or construct validity. The cited frameworks informed item development, but formal item-by-item framework mapping, pilot testing, cognitive interviews, and exploratory factor analysis were not conducted. The complete questionnaire, including instructions, response options, and scale anchors, is provided in Supplementary File S2.

AI interest consisted of four items assessing interest in AI technology itself, interest in using generative AI, interest in applying AI and digital technologies to occupational therapy assessment and intervention, and intention to enroll in related courses. Each item was rated on a 5-point Likert scale from 1 (strongly disagree) to 5 (strongly agree), with higher scores indicating greater interest in AI-based occupational therapy. Cronbach’s α in this study was 0.901.

Self-reported AI competency consisted of seven items: writing prompts appropriate to a given purpose; judging the accuracy and reliability of AI outputs; identifying errors, exaggeration, bias, or unclear evidence; linking AI outputs to occupational therapy knowledge and clinical reasoning; selecting AI and digital tools appropriate for clients; understanding ethical standards related to privacy, copyright, and plagiarism; and using AI safely and responsibly. Items were rated on a 5-point Likert scale, with higher scores indicating greater perceived AI competency. Cronbach’s α in this study was 0.902. The tool-selection item was included to reflect the clinical context in which AI and other digital tools may be considered together. Because selecting a non-AI digital tool is not itself an AI competency, this item broadens the domain and limits its interpretation as a measure specific to AI.

Educational content needs consisted of 10 items covering basic understanding of generative AI; prompt writing; review of output accuracy and reliability; responses to errors such as hallucinations and bias; integration with occupational therapy knowledge and clinical reasoning; digital rehabilitation technologies; online counseling and telerehabilitation; client-tailored tool selection; practice in analyzing assessment data and developing intervention plans; and ethics and safety, including privacy protection and prevention of copyright infringement and plagiarism. Each item was rated from 1 (not at all necessary) to 5 (very necessary), with higher scores indicating greater educational needs. Cronbach’s α in this study was 0.938.

Educational priority consisted of seven items: computerized cognitive rehabilitation programs, upper- and lower-extremity rehabilitation robots, virtual- and augmented-reality-based rehabilitation tools, wearable sensor and motion-analysis systems, online counseling and telerehabilitation platforms, tools for creating digital patient education content, and generative AI tools for creating documents and materials. The need to use each tool directly within a course was rated from 1 (not at all necessary) to 5 (very necessary), with higher scores indicating a higher educational priority for the corresponding digital tool. Cronbach’s α in this study was 0.906.

Instructional method preference consisted of eight items: instructor-led lectures on core concepts, step-by-step guided practice, analysis of actual clinical cases, activities to identify errors in AI outputs and review supporting evidence, hands-on experience with digital rehabilitation tools, small-group discussion and problem-solving activities, team projects, and guest lectures and demonstrations by field experts. Each item was rated on a 5-point Likert scale from 1 (do not prefer at all) to 5 (strongly prefer), with higher scores indicating greater preference for the corresponding instructional method. Cronbach’s α in this study was 0.866.

Questionnaire scoring and missing responses. Domain scores were arithmetic means of the 4 interest, 7 competency, 10 educational-content, 7 tool-priority, and 8 instructional-preference items, respectively. No additional reverse coding was applied when calculating the domain means. All main-domain items were complete for the 230 analyzed respondents; no missing-item values were imputed. Ancillary characteristic comparisons used available cases. Domain scores ranged from 1 to 5, with higher scores indicating higher levels of the respective construct; response options and scale anchors are provided in Supplementary File S2.

Ethical Considerations

The survey was conducted entirely online and anonymously and involved no clinical intervention. No personally identifiable information, including names, student identification numbers, telephone numbers, or Email addresses, was collected or recorded, and no sensitive health information was obtained. Before participation, all students were informed of the purpose and procedures of the study, the voluntary nature of participation, confidentiality of their responses, the intended research use of the data, and their right to decline or discontinue participation without disadvantage. Participants were required to indicate their agreement electronically before proceeding to the questionnaire, and the consent decision was electronically recorded through the online survey system and retained as part of the research records.

Under Article 13 of the Enforcement Rule of the Bioethics and Safety Act of the Republic of Korea, certain human-subject research that does not collect or record personally identifiable information may be eligible for exemption from Institutional Review Board (IRB) review when the specified low-risk conditions are met.²³ The present study was designed as an anonymous, minimal-risk, non-interventional survey without collection or recording of personally identifiable information. However, formal IRB approval or a prospective IRB exemption determination was not obtained before commencement of the study. Article 16(1) of the Bioethics and Safety Act of the Republic of Korea defines written informed consent as including consent provided through electronic documents.²⁴ Accordingly, the electronic informed consent obtained in this study constituted written informed consent under applicable Korean legislation.

Data Analysis

Data were analyzed using IBM SPSS Statistics version 30.0 for the original analyses; the present revision was independently reproduced using Python with NumPy and SciPy 1.17.0. Participant characteristics were summarized as frequencies and percentages; domain scores were summarized as means and standard deviations, and internal consistency was assessed using Cronbach’s α. Group differences were examined using Welch t-tests or Welch ANOVA, with Games–Howell post hoc comparisons for nominally significant Welch ANOVAs. For the competency comparison by AI-use frequency, Welch ANOVA was undefined because all three never-users had identical scores, so the Kruskal–Wallis test was used. Shapiro–Wilk and median-centered Levene tests assessed distributional and variance characteristics; Kruskal–Wallis tests provided sensitivity checks for other group comparisons. Hedges g or descriptive eta squared was reported with 95% stratified percentile bootstrap intervals (4,000 replicates). Holm correction was applied across the 25 characteristic-by-domain tests; nominal and adjusted p-values are reported.

Pearson correlations were supplemented by Fisher-z confidence intervals, Spearman correlations, and Holm correction across the 10 correlations. Hierarchical multiple regression retained sex and year of study in Model 1, ordinal AI-use frequency in Model 2, and AI interest and self-reported competency in Model 3. Regression diagnostics included residual Shapiro–Wilk, Breusch–Pagan, Ramsey RESET, VIF, leverage, and Cook’s distance. Sensitivity analyses used HC3 robust standard errors, indicator-coded frequency categories with 1–3 uses/week as the reference, and a six-item competency score excluding the digital-tool-selection item. All 230 cases were retained in the primary models. Two-sided nominal significance was set at p<0.05; the additional analyses are exploratory.

Results

Participant Characteristics

A total of 230 students were included in the final analysis. First-year students comprised the largest group (n=79, 34.3%), followed by third-year students (n=64, 27.8%), fourth-year students (n=49, 21.3%), and second-year students (n=38, 16.5%). There were 141 women (61.3%) and 89 men (38.7%). The most common frequency of generative AI use was 1–3 times per week (n=92, 40.0%), followed by four or more times per week or almost daily (n=65, 28.3%) and 1–3 times per month (n=59, 25.7%). Regarding experience with digital rehabilitation, the largest group had learned only the concepts in class (n=82, 35.7%), whereas 52 students (22.6%) reported no related experience. Most students rated their generative AI competency as average (n=137, 59.6%). AI tools previously used were assessed using multiple responses; ChatGPT was the most commonly used (n=197, 85.7%), followed by Google Gemini (n=186, 80.9%) and Claude (n=69, 30.0%) (Table 1).

Table 1.

Participant Characteristics and Generative AI and Digital Rehabilitation Use Characteristics (N=230)

Characteristic	Category	N	%
Year of study	First year	79	34.3
	Second year	38	16.5
	Third year	64	27.8
	Fourth year	49	21.3
Sex	Male	89	38.7
	Female	141	61.3
Frequency of generative AI use	Never used	3	1.3
	Less than once a month	11	4.8
	1–3 times/month	59	25.7
	1–3 times/week	92	40.0
	≥4 times/week or almost daily	65	28.3
Experience with digital rehabilitation	Learned concepts only in class	82	35.7
	Exposure through demonstrations or site visits	20	8.7
	Direct use in departmental practical training	32	13.9
	Observed or used during clinical fieldwork	44	19.1
	No related experience	52	22.6
Self-rated generative AI competency	Very poor	7	3.0
	Poor	46	20.0
	Average	137	59.6
	Proficient	36	15.7
	Very proficient	4	1.7
AI tools used*	ChatGPT	197	85.7
	Google Gemini	186	80.9
	Microsoft Copilot	15	6.5
	Claude	69	30.0
	Perplexity	31	13.5
	Image/video generation AI	50	21.7
	No experience using AI tools	5	2.2

Notes: *Multiple responses were allowed.

AI Interest, Self-Reported AI Competency, Educational Content Needs, Educational Priority, and Instructional Method Preference

The mean AI interest score was 3.64±0.77, with intention to enroll in related courses receiving the highest item score (3.75±0.88). The mean self-reported AI competency score was 3.42±0.67; safe and responsible AI use received the highest score (3.73±0.83), whereas selecting AI and digital tools appropriate for clients received the lowest score (3.22±0.88). The mean educational content needs score was 3.93±0.68, and education on AI use based on occupational therapy knowledge and clinical reasoning received the highest score (4.04±0.83).

The mean digital tool educational priority score was 3.95±0.63, with computerized cognitive rehabilitation programs (4.08±0.80) and wearable sensor and motion-analysis systems (4.07±0.76) receiving high ratings. The mean instructional method preference score was 3.82±0.64. Hands-on experience with digital rehabilitation tools received the highest score (4.23±0.75), whereas team projects received the lowest score (3.15±1.12) (Table 2).

Table 2.

Self-Reported AI Competency and Educational Needs for Integrated AI and Digital Rehabilitation Education

Domain	Item	M±SD
AI interest	Interest in AI technology itself	3.51±0.90
	Interest in using generative AI such as ChatGPT	3.66±0.88
	Interest in applying AI and digital technologies to occupational therapy assessment and intervention	3.64±0.86
	Intention to enroll in a course on AI-based occupational therapy if offered	3.75±0.88
	Overall domain	3.64±0.77
Self-reported AI competency	Writing questions or prompts appropriate to the purpose	3.40±0.84
	Judging the accuracy and reliability of AI search/response outputs	3.43±0.78
	Identifying errors, exaggeration, bias, and unclear evidence in AI outputs	3.44±0.82
	Linking AI outputs with occupational therapy knowledge and clinical reasoning	3.26±0.83
	Selecting AI and digital tools appropriate to client characteristics and treatment goals	3.22±0.88
	Understanding ethical standards including privacy, copyright, and plagiarism	3.48±0.95
	Using AI safely and responsibly in assignments or clinical situations	3.73±0.83
	Overall domain	3.42±0.67
Educational content needs	Understanding the principles, functions, and limitations of generative AI	3.67±0.89
	Prompt writing and searching for discipline-related materials	3.89±0.86
	Reviewing the accuracy and reliability of AI responses and verifying evidence	4.00±0.87
	Responding to incorrect AI information such as hallucinations and bias	3.97±0.89
	Using AI based on occupational therapy knowledge and clinical reasoning	4.04±0.83
	Understanding and using computerized cognitive rehabilitation, rehabilitation robots, VR/AR, and wearable technologies	4.00±0.84
	Online counseling, telerehabilitation, and digital education platforms	3.80±0.85
	Selecting and comparing tools appropriate to client function, environment, and goals	3.94±0.80
	Practice in analyzing assessment data and developing intervention plans using AI and digital tools	4.00±0.75
	Ethics and safety, including privacy protection and prevention of copyright infringement and plagiarism	4.00±0.82
	Overall domain	3.93±0.68
Digital tool educational priority	Computerized cognitive rehabilitation programs	4.08±0.80
	Upper- and lower-extremity rehabilitation robots	4.05±0.83
	Virtual- and augmented-reality-based rehabilitation tools	3.99±0.84
	Wearable sensor and motion-analysis systems	4.07±0.76
	Online counseling and telerehabilitation platforms	3.75±0.81
	Tools for creating digital patient education content	3.91±0.76
	Generative AI tools for creating documents and materials	3.83±0.75
	Overall domain	3.95±0.63
Instructional method preference	Instructor-led lectures on core concepts	4.03±0.79
	Step-by-step guided practice	4.05±0.84
	Analysis of actual clinical cases	4.05±0.81
	Activities to identify errors in AI outputs and review supporting evidence	3.77±0.87
	Hands-on experience with digital rehabilitation tools	4.23±0.75
	Small-group discussion and problem-solving activities	3.40±0.99
	Team projects	3.15±1.12
	Guest lectures and demonstrations by field experts	3.89±0.90
	Overall domain	3.82±0.64

Notes: All item and domain scores in Table 2 were calculated using complete-response data from the same 230 participants.

Abbreviations: M, mean; SD, standard deviation.

Differences in Self-Reported AI Competency and Educational Needs According to Participant Characteristics

AI interest differed by frequency of generative AI use (Welch F(4,13.66)=4.59, p=0.015). Across single-item self-rated competency groups, differences were observed in AI interest (Welch F(4,15.67)=5.76, p=0.005), self-reported AI competency (Welch F(4,15.90)=18.57, p<0.001), educational content needs (Welch F(4,16.41)=4.18, p=0.016), and educational priority (Welch F(4,19.12)=23.65, p<0.001). These p-values were unadjusted. No significant differences were found by sex or year of study.

After Holm correction across 25 comparisons, only the differences in self-reported AI competency and educational priority across self-rated competency groups remained significant (adjusted p=0.00018 and.000009, respectively). In the rank-based sensitivity analyses, the AI-use-frequency difference in interest was significant after correction (adjusted p=0.003), whereas the self-rated-competency difference in educational priority was not (adjusted p=0.603). Detailed subgroup statistics and post hoc comparisons are presented in Supplementary Table S1.

Correlations Among the Five Major Domains

Pearson correlation analysis showed significant positive correlations among all five major domains (Table 3). Interest in AI-based occupational therapy was significantly and positively correlated with self-reported AI competency (r=0.474, p<0.001), educational content needs (r=0.553, p<0.001), educational priority (r=0.506, p<0.001), and instructional method preference (r=0.424, p<0.001). Educational content needs showed strong positive correlations with educational priority (r=0.680, p<0.001) and instructional method preference (r=0.634, p<0.001), and the strongest correlation was observed between educational priority and instructional method preference (r=0.701, p<0.001).

Table 3.

Correlations Among the Five Major Domains

Variable	1	2	3	4	5
1. Interest in AI-based occupational therapy	1
2. Self-reported AI competency	0.474*** [0.367, 0.568]	1
3. Educational content needs	0.553*** [0.457, 0.637]	0.389***[0.274, 0.494]	1
4. Educational priority	0.506*** [0.403, 0.596]	0.344*** [0.224, 0.453]	0.680*** [0.604, 0.744]	1
5. Instructional method preference	0.424*** [0.312, 0.524]	0.309*** [0.187, 0.421]	0.634*** [0.550, 0.706]	0.701*** [0.628, 0.761]	1

Notes: Entries are Pearson r [95% CI]; confidence intervals use the Fisher z transformation (N=230). ***p<0.001. All ten Pearson correlations remain significant after Holm correction (all adjusted p<0.001). Spearman coefficients range from.309 to.674 (all Holm-adjusted p<0.001); neither analysis rules out common-method bias.

Shared response formats and a general tendency to endorse AI/digital education may contribute to these associations. All ten correlations remained significant after Holm correction; Spearman correlations were also positive (ρ=0.309–0.674; all adjusted p<0.001). Neither the rank-based sensitivity analysis nor multiplicity correction rules out common-method bias.

Factors Associated with Educational Needs for AI-Integrated Occupational Therapy Education

Hierarchical multiple regression analysis examined factors associated with educational content needs (Table 4). Model 1, including sex and year of study, explained 2.1% of the variance and was not significant (R²=.021, F(4,225)=1.19, p=0.316). Adding generative AI-use frequency in Model 2 increased the explained variance to 2.9%, but the model remained nonsignificant (R²=.029, ΔR²=.008, F(5,224)=1.35, p=0.245). Adding AI interest and self-reported AI competency in Model 3 increased the explained variance to 34.3% (adjusted R²=.322, ΔR²=.314). The final model was significant (F(7,222)=16.55, p<0.001). AI interest showed the strongest positive association with educational content needs (β=0.490, p<0.001), followed by self-reported AI competency (β=0.154, p=0.015). Sex, year of study, and AI-use frequency were not significant.

Table 4.

Hierarchical Multiple Regression Analysis of Factors Associated with Educational Content Needs

Model	Predictor	B [95% CI]	SE	β	t	p	HC3 p (Model 3)
Model 1	Sex (male)	−0.149 [−0.330, 0.032]	0.092	−0.108	−1.62	0.107	—
	Second year	0.126 [−0.137, 0.388]	0.133	0.069	0.94	0.347	—
	Third year	0.094 [−0.130, 0.318]	0.114	0.062	0.83	0.409	—
	Fourth year	0.136 [−0.107, 0.379]	0.123	0.083	1.10	0.271	—
	Model summary	R2=.021; Adj. R2=.003; ΔR2=.021; F(4,225)=1.19				0.316
Model 2	Sex (male)	−0.123 [−0.307, 0.061]	0.094	−0.089	−1.32	0.190	—
	Second year	0.139 [−0.124, 0.402]	0.133	0.077	1.04	0.298	—
	Third year	0.098 [−0.126, 0.321]	0.113	0.065	0.86	0.390	—
	Fourth year	0.132 [−0.111, 0.374]	0.123	0.080	1.07	0.286	—
	Frequency of generative AI use	0.070 [−0.028, 0.168]	0.050	0.094	1.40	0.163	—
	Model summary	R2=.029; Adj. R2=.008; ΔR2=.008; F(5,224)=1.35				0.245
Model 3	Sex (male)	−0.127 [−0.280, 0.026]	0.078	−0.092	−1.63	0.105	0.141
	Second year	0.142 [−0.076, 0.359]	0.110	0.078	1.29	0.200	0.198
	Third year	−0.013 [−0.199, 0.174]	0.095	−0.009	−0.14	0.892	0.886
	Fourth year	0.067 [−0.134, 0.269]	0.102	0.041	0.66	0.511	0.578
	Frequency of generative AI use	−0.022 [−0.107, 0.062]	0.043	−0.030	−0.52	0.605	0.628
	Interest in AI-based occupational therapy	0.429 [0.316, 0.542]	0.057	0.490	7.51	<0.001	<0.001
	Self-reported AI competency	0.154 [0.030, 0.278]	0.063	0.154	2.45	0.015	0.0504
	Model summary	R2=.343; Adj. R2=.322; ΔR2=.314; F(7,222)=16.55				<0.001

Notes: References: female and first year. Frequency was scored 1–5, assuming a linear category trend rather than equal intervals of actual use. CIs use unrounded estimates and residual df=225, 224, and 222. Main columns use conventional OLS; HC3 p-values are shown for Model 3 (dashes: not displayed). Model 3 diagnostics: residual Shapiro–Wilk W=0.9859, p=0.023; Breusch–Pagan χ²(7)=7.516, p=0.377; RESET F(2,220)=.209, p=0.812; VIF=1.07–1.44. Competency HC3 95% CI: [−.0003,0.3087]. Categorical-frequency model (reference: 1–3 uses/week; N=230): R²=.344, F(10,219)=11.506, p<0.001; nested comparison F(3,219)=.163, p=0.921. Interest B=0.434, conventional and HC3 p<0.001; competency B=0.146, conventional p=0.024, HC3 p=0.0704. The never-use group (n=3) limits precision; nonsignificant diagnostic tests do not establish model adequacy.

Abbreviations: B, unstandardized coefficient; SE, standard error; β, standardized coefficient; Adj. R², adjusted R²; ΔR², change in R²; CI, confidence interval.

VIF values indicated no serious multicollinearity (1.07–1.44). Residuals departed from normality (Shapiro–Wilk p=0.023), while the Breusch–Pagan and Ramsey RESET tests were nonsignificant. With HC3 robust standard errors, AI interest remained significant (p<0.001), whereas self-reported AI competency did not reach the.05 threshold (p=0.0504; 95% CI [−.0003,0.3087]). Treating AI-use frequency as categorical produced similar explanatory power (R²=.344); competency was significant under conventional inference (p=0.024) but not under HC3 (p=0.070). Excluding the tool-selection item yielded a similar pattern (B=0.157, conventional p=0.014; HC3 p=0.0508). Detailed diagnostics and categorical-frequency results are reported in the Table 4 note.

Discussion

This study examined self-reported generative AI competency and needs for integrated AI and digital rehabilitation education among occupational therapy students and identified factors associated with educational content needs. Overall, students showed moderate-to-high levels of interest in AI-based occupational therapy and self-reported AI competency, and their educational content needs and priorities for digital tool education were relatively high. Conventional regression revealed that interest in AI-based occupational therapy and self-reported AI competency were also associated with educational content needs. Students considered it important for education to connect AI and digital technologies with occupational therapy knowledge and clinical reasoning and to apply them to actual clinical situations.

Interest in AI-based occupational therapy and self-reported AI competency were moderate to high, and educational content needs and priorities for digital tool education were relatively high. Among the educational content items, the need for AI use based on occupational therapy knowledge and clinical reasoning was rated highest. This indicates that students perceived a need for education not only in the general use of generative AI but also in how to apply and evaluate it within occupational therapy and actual clinical situations. Recent studies of medical and health sciences students have similarly reported overall moderate levels of AI readiness and emphasized the need for systematic education covering practical competence and ethical use.^8,15 Taken together, these findings suggest that one potential direction for occupational therapy education is to move beyond understanding and basic use of AI and to emphasize its application through professional knowledge and clinical reasoning.

Among the detailed self-reported AI competency items, students reported relatively lower competence in selecting AI and digital tools appropriate for client characteristics and treatment goals and in linking AI outputs to occupational therapy knowledge and clinical reasoning. At the same time, educational needs were high for AI use based on occupational therapy knowledge and clinical reasoning, analysis of assessment data and development of intervention plans, and digital rehabilitation technologies. These findings suggest that the profession-specific competencies students perceive as insufficient are closely aligned with their perceived educational needs. One potential educational implication is to go beyond prompt writing and information retrieval by strengthening tool selection based on client function, environment, and treatment goals, critical appraisal of AI outputs, and integration of AI outputs with clinical reasoning. Previous studies have similarly reported that generative AI can help explore intervention ideas, but final clinical judgment still requires a review process that considers client needs and evidence, supported by systematic educational guidance.^9,25 Educational needs were also high for evaluating the accuracy and reliability of AI outputs, responding to hallucinations and bias, and addressing ethical and safety issues such as privacy protection, copyright, and plagiarism prevention. These preferences suggest that AI education may address not only technical use but also output verification and responsible use, which are consistent with WHO principles of safety, transparency, and accountability in healthcare AI.^1,2 The relatively high rating for safe and responsible use may reflect confidence rather than verified ethical knowledge or safe behavior and may coexist with lower confidence in clinically applying AI.

In digital tool education, computerized cognitive rehabilitation programs and wearable sensor and motion-analysis systems received high priority ratings. Regarding instructional methods, students expressed strong preferences for hands-on experience with digital rehabilitation tools, step-by-step practice, and analysis of actual clinical cases. These findings indicate a preference for practice-oriented education in which students apply AI and digital technologies to actual occupational therapy situations rather than learning about them only theoretically. One possible curricular approach for future testing is a staged practice structure centered on actual clinical cases, progressing from assessment-data analysis and tool selection to intervention planning and evidence verification. This proposal distinguishes critical appraisal of generative AI outputs from the selection and use of digital rehabilitation tools, which need not incorporate AI. Its effectiveness was not tested in this study.

For the analyses based on participant characteristics, nominal differences in interest in AI-based occupational therapy related to the frequency of generative AI use were observed. In addition, AI interest, self-reported AI competency, educational content needs, and educational priority differed nominally with self-rated AI competency. In contrast, no differences related to sex or year of study were observed. These exploratory findings suggest that perceived educational needs are associated with the AI-use experience and perceived competency of students; they do not establish that these factors are more important than year of study or sex. A previous study of healthcare students in Korea similarly found no significant differences in AI literacy related to sex or year of study and reported that interest in AI was more strongly associated with AI literacy than with previous AI education experience.¹⁶ Future studies should test whether AI-use experience, self-reported and objectively assessed competency, and year of study can help inform differentiated instruction. The difference in the multi-item score across self-rated competency groups involves closely related self-report measures and adds limited independent evidence. After Holm correction, only the self-rated-competency differences in the multi-item competency score and educational priority remained significant; the latter did not survive the rank-based sensitivity analysis after correction. These exploratory differences should, therefore, not determine curricular placement.

The scores for all five major domains were significantly and positively correlated. Educational content needs were strongly associated with educational priority and instructional method preference, and the correlation between educational priority and instructional method preference was the strongest (r=0.701). This suggests that students who perceive a greater need for education in AI and digital rehabilitation technologies also tend to have stronger preferences for related educational content and instructional approaches. A previous study of occupational therapy students also found positive attitudes toward the educational use of AI and demand for specific AI learning functions and modes of use tailored to course characteristics.²⁶ These findings support the possibility that the need for AI education is linked to preferences for specific educational content and learning methods.

The hierarchical multiple regression analysis further supported this pattern. Neither Model 1, which included sex and year of study, nor Model 2, which additionally included frequency of generative AI use, was statistically significant. However, the explanatory power of the final model increased to 34.3% after interest in AI-based occupational therapy and self-reported AI competency were added. Interest in AI-based occupational therapy showed the strongest positive association with educational content needs (β=0.490, p<0.001), and self-reported AI competency was also positively associated under conventional inference (β=0.154, p=0.015). In contrast, frequency of generative AI use was not significant in the final model. These findings suggest that interest in applying AI to one’s professional field and perceived AI competency may be more closely related to educational needs than frequent use of generative AI alone. A previous study of healthcare students in Korea found that interest in AI was closely associated with AI literacy and that higher AI literacy was associated with more positive attitudes toward AI and greater intention to use AI in clinical contexts.¹⁶ A study of medical students likewise identified a significant relationship between interest in AI and AI literacy and a positive relationship between AI literacy and attitudes toward AI.²⁷ Together, these findings highlight the importance of considering both interest and competency rather than AI-use experience alone. Future studies could examine the educational value of providing students with opportunities to apply AI to occupational therapy clinical problem-solving and to critically evaluate its outputs. The interest association persisted with HC3 robust inference and categorical frequency coding. In contrast, the competency association was borderline with HC3 (p=0.0504) and was not significant in the categorical-frequency HC3 analysis (p=0.070); its interpretation should remain tentative. Some observations had elevated Cook’s distance, and the original 230-case sample was retained.

Although the participants in this study were occupational therapy students, the findings raise questions for future multidisciplinary research. The educational topics examined here—critical appraisal of AI-generated information, ethical and safe use, evidence verification, client-tailored technology selection, and practice-oriented use of digital rehabilitation tools—may also warrant investigation in other rehabilitation and healthcare disciplines, including physical therapy and nursing. Occupational therapists, physical therapists, and nurses increasingly share digital health information, contribute to coordinated rehabilitation planning, and use technology-supported assessment, education, and intervention. Therefore, a common educational foundation in responsible AI use may be worth investigating as a potential support for multidisciplinary healthcare practice. However, physical therapy and nursing students were not sampled, and neither cross-profession applicability nor educational effectiveness was established.

Limitations and Future Directions

This study has limitations. First, convenience sampling was used to recruit occupational therapy students from three universities in Cheongju, Busan, and Ulsan, limiting the generalizability of the findings to all occupational therapy students in South Korea. Future studies should include universities from a wider range of regions, including the Seoul metropolitan area, to determine whether the findings generalize. Second, self-reported AI competency was measured with a questionnaire and may differ from the actual ability of the students to use AI. Future studies should incorporate performance-based assessments such as prompt writing, detection of errors in AI outputs, evidence verification, and application to clinical cases. Third, the questionnaire was developed by the researchers for the specific aims of this study and underwent expert content review, but its construct validity and criterion validity were not independently established. Although internal consistency was high across the five domains, future studies should further evaluate the psychometric properties of the instrument in larger and more diverse samples. Fourth, because this was a cross-sectional survey, associations among AI interest, self-reported AI competency, and educational content needs were identified, but causal relationships could not be established. Future longitudinal or intervention studies should apply AI and digital rehabilitation education within actual courses and examine changes in self-reported AI competency, clinical reasoning, and digital tool-selection ability before and after education. Additional limitations include possible faculty–student power dynamics, social-desirability bias, common-method variance, and conceptual overlap among the domains. Further instrument development and validation, including cognitive interviewing and factor analysis, are needed to clarify domain structure. The very small subgroup sizes and method-dependent findings, even after supplementary checks and multiplicity correction, limit subgroup interpretation. Detailed item-level missingness patterns and differences between included and excluded respondents could not be evaluated from the available data; potential selection bias associated with complete-case analysis cannot be ruled out. Categorical-frequency and HC3 sensitivity analyses supported the interest association but showed that the competency association was less stable.

Conclusion

In this study, occupational therapy students reported moderate-to-high levels of interest and perceived competency in AI-based occupational therapy, while educational content needs and educational priorities for integrated AI and digital rehabilitation education were relatively high. In particular, interest in AI-based occupational therapy was consistently associated with educational content needs, whereas the self-reported competency association was sensitive to robust inference. Students expressed high educational needs for AI use based on occupational therapy knowledge and clinical reasoning, selection of tools appropriate to client characteristics and treatment goals, analysis of assessment data and development of intervention plans, and hands-on experience with digital rehabilitation tools.

These reported preferences suggest that future AI education in occupational therapy could extend beyond basic generative AI use or prompt writing and could address evidence verification of AI outputs, client-tailored tool selection, ethics and safety, and practical application of digital rehabilitation technologies around occupational therapy knowledge and clinical reasoning. The findings of this study may provide preliminary information for developing integrated AI and digital rehabilitation curricula and practice-oriented courses in occupational therapy education. The survey does not establish objective competency deficits, a single integrated AI–digital rehabilitation competency, or curricular effectiveness. Applicability to other healthcare disciplines requires future study.

Funding Statement

This research was supported by Choonhae Health Sciences University (2026).

Ethics Approval and Informed Consent

This study was a non-interventional, anonymous online questionnaire survey involving occupational therapy students. No personally identifiable information, including names, student identification numbers, telephone numbers, or Email addresses, was collected or recorded, and no sensitive health information was obtained. Under Article 13 of the Enforcement Rule of the Bioethics and Safety Act of the Republic of Korea, certain human-subject research that does not collect or record personally identifiable information may be eligible for exemption from IRB review when the specified low-risk conditions are met.²³ The study was designed as an anonymous, minimal-risk, non-interventional survey without collection or recording of personally identifiable information. However, formal IRB approval or a prospective IRB exemption determination was not obtained before commencement of the study.

Prior to accessing the questionnaire, all participants were provided with information regarding the purpose and procedures of the study, voluntary participation, confidentiality, use of the data for research purposes, and their right to decline or discontinue participation at any time without disadvantage. Participants were required to provide informed consent electronically before proceeding to the questionnaire. Article 16(1) of the Bioethics and Safety Act of the Republic of Korea recognizes written informed consent as including consent provided through electronic documents.²⁴ Accordingly, the electronic informed consent obtained in this study constituted written informed consent under applicable Korean legislation. The consent decision was electronically recorded through the online survey system before participants were permitted to proceed to the questionnaire and was retained as part of the research records.

Consent for Publication

Not applicable. The manuscript contains no information, images, recordings, or other material that could identify individual participants.

Disclosure

The authors report no conflicts of interest in this work.