This study used an RCT design to assess the impact of AI-assisted patient health education using voice cloning technology and ChatGPT on patient education outcomes. Participants were randomly assigned to 1 of 3 groups: The 3 groups received different educational methods but the same educational content and duration of education. Following the educational intervention, the effectiveness of the education was comprehensively evaluated at fixed time points using ChatGPT. To ensure the validity and accuracy of the evaluation tool, a pre-test of the ChatGPT evaluation tool was conducted prior to the formal data collection. This study not only explored the potential of AI voice cloning technology in personalized patient education but also used ChatGPT as an assessment tool to further validate its reliability and practicality in evaluating health education interventions. Ultimately, the study aimed to provide theoretical insights and practical recommendations for improving medical health education models, enhancing patient education outcomes, and promoting treatment adherence and quality of life.

An independent tool validation study was conducted from January 2024 to June 2024 to verify the ChatGPT-based compliance scoring tool (no participant enrollment). Administrative preparation, staff training, and software and hardware debugging were completed in December 2024, and this period did not involve participant enrollment. Trial registration (ChiCTR2500101882) was initiated on January 15, 2025, and finalized on April 30, 2025, before enrollment began in May 2025. Participant enrollment and intervention delivery were conducted from May 2025 to June 2025, with 1-month follow-up assessments completed by July 2025.

The trial was conducted and reported in accordance with the CONSORT (Consolidated Standards of Reporting Trials) 2025 statement and the CONSORT-EHEALTH checklist for digital and AI-based health interventions (Checklists 1 and 2) [28,29].

This single-center, 3-arm, parallel-group, superiority RCT (allocation ratio 1:1:1) was conducted at The First Hospital of Lanzhou University. System development, intervention standardization, and staff training were completed in December 2024 and January 2025. Participant recruitment and data collection were conducted from May 2025 to July 2025.

The clinical trial protocol was reviewed and approved by the Ethics Committee of The First Hospital of Lanzhou University (approval number: LDYYLL-2025‐805) and conducted in accordance with the principles of the Declaration of Helsinki.

All participants received detailed verbal and written information regarding the study purpose, procedures, potential risks, and expected benefits and provided written informed consent prior to enrollment. The consent process also covered permission for the use and publication of anonymized data. Participants were informed that their involvement was voluntary and that they could withdraw from the study at any time without affecting their standard medical care.

All collected data were deidentified before analysis and stored in encrypted, password-protected databases accessible only to the research team. Personal identifiers were removed to ensure participant confidentiality and data security.

The use of voice cloning technology was explicitly covered in the consent form. Participants and physicians provided separate written authorization for the use of their voice recordings solely for research purposes. All voice samples and generated models were permanently deleted after study completion.

Participants did not receive monetary compensation but were offered free access to the AI-based health education program and follow-up consultations as part of their clinical care. No identifiable photographs nor audiovisual materials were used in this publication.

The study participants were patients hospitalized in a tertiary hospital from May 2025 to July 2025 who needed to receive medical education.

The inclusion criteria were (1) age ≥18 years of any gender; (2) clear diagnosis, patients who needed to receive medical education; (3) clear consciousness, able to understand and cooperate with the study; and (4) voluntary participation in this study and signed informed consent. The exclusion criteria were (1) severe hearing impairment, unable to normally receive voice education; (2) severe cognitive impairment, unable to understand education content; (3) expected hospitalization time <3 days; and (4) previously participated in similar studies.

Based on the pre-experimental results, with the education content compliance rate as the main observation indicator, α set at 0.05 (2-sided), β set at 0.10, a test power of 90%, and an expected effect size of 0.3 among the 3 groups, we calculated that each group needed a sample size of 54 participants using G*Power 3.1 software. Considering a possible dropout rate of approximately 10%, we finally determined that we would enroll 60 participants in each group, for a total of 180 participants. Prior to participation, all patients were informed that the educational intervention would be delivered using AI-generated voice recordings. Written informed consent was obtained from all participants.

This independent tool validation study was conducted separately from the randomized trial and did not involve participant enrollment. To validate the AI-based scoring used for the primary outcome, we conducted a separate pilot validation study (January 2024 through June 2024) among 30 volunteers who were not part of the randomized trial. Phase 1 (verbatim-matching workflow) used iFlytek automatic speech recognition (Mandarin medical context; no manual correction) to transcribe each participant’s recitation; the transcribed text was pasted into the ChatGPT web interface together with the standard education text for rubric-based scoring. Phase 2 (semantic audio-based workflow) re-evaluated the same recordings by uploading the original audio directly to the ChatGPT web interface (WAV/MP3; 16-kHz sampling rate; 60‐120 s per recording); each recording was scored once. Both workflows were benchmarked against consensus ratings from 3 senior nursing experts using the same rubric. Agreement with expert ratings was substantial in Phase 1 (weighted κ=.72) and higher in Phase 2 (weighted κ=.87), with strong internal consistency of the evaluation items (Cronbach α=0.89). Therefore, Phase 2 was selected for the main RCT to minimize transcription-related errors and enhance reproducibility. Detailed prompts, rubric, and example inputs and outputs are provided in Multimedia Appendix 1.

Research assistants who had received standardized training were responsible for data collection, using uniform data collection forms to record patient information and evaluation results. An electronic database was established, with all data independently entered by 2 personnel and cross-checked to ensure accuracy. To protect patient privacy, all data were stored in coded form, and regular data quality checks were conducted to ensure completeness and reliability. Missing data were promptly identified and imputed.

All voice recordings collected in this study were considered sensitive personal data. They were stored on secure, password-protected servers at The First Hospital of Lanzhou University, accessible only to authorized research personnel. For analysis, the recordings were anonymized. Following the completion of the study, all original recordings will be permanently deleted in accordance with institutional data retention and privacy policies.

Statistical analyses were performed using SPSS version 26.0 (IBM Corp). All analyses followed the intention-to-treat (ITT) principle. Missing outcome data (6/182, 3.3%) were evaluated using the missing completely at random test from Little, which indicated that missingness was completely at random (χ²₁₂=14.5, P=.27). Consequently, multiple imputation by chained equations was used to generate 5 imputed datasets (m=5), and pooled estimates were obtained using the rules by Rubin. Sensitivity analyses using per-protocol (complete-case) data yielded consistent findings (absolute difference in mean compliance<1%). Continuous variables are presented as mean (SD), and categorical variables are presented as n (%). Between-group comparisons were conducted using 1-way ANOVA with post hoc tests (least significant difference or Bonferroni, as appropriate). Categorical variables were compared using χ² tests or Fisher exact tests. All tests were 2-sided with a significance level of P<.05.

The primary outcome of this study was the education content compliance rate, which reflected how well patients understood and followed the health education provided. Secondary outcomes included knowledge retention, patient satisfaction, and treatment adherence.

Education level and knowledge retention were measured using a structured questionnaire developed by the research team based on the hospital’s standardized education materials.

The questionnaire consisted of multiple-choice and short-answer questions assessing patients’ understanding and recall of the key educational points immediately after education and before discharge.

Each correct response was scored as 1 point, with higher total scores indicating better knowledge retention.

The disease-related knowledge questionnaire was developed based on the hospital’s standardized education materials and underwent content review by 3 senior nursing experts.

Figure 1 illustrates the participant flow. ITT analysis was performed for all participants (n=180) using multiple imputation by chained equations for missing data. The analysis section in the figure shows the per-protocol analyses (n=174).

A total of 180 patients were recruited to participate in this study. During the follow-up period, 6 participants were lost to follow-up due to transfer to another hospital (n=1) and withdrawal of consent (n=1) in the control group; early discharge (n=1) and withdrawal of consent (n=2) in intervention group 1; and failure to complete follow-up assessments (n=1) in intervention group 2. Therefore, 174 patients completed the study (58 in the control group, 57 in intervention group 1, 59 in intervention group 2), for a total dropout rate of 3.3% (6/180). After multiple imputation, the analysis results remained consistent with those of the complete-case dataset, indicating the robustness of the findings. Comparison of general information such as age, gender, education level, occupation, and disease type among the 3 groups of patients showed no statistically significant differences (P>.05), indicating comparability (Table 1).

Immediately after education, the mean education compliance score in the control group was 73.2 (SD 8.5; 95% CI 71.0‐75.4), compared with mean scores of 86.7 (SD 7.3; 95% CI 84.8‐88.6) in intervention group 1 and 92.5 (SD 6.8; 95% CI 90.8‐94.2) in intervention group 2. The 1-way ANOVA showed a statistically significant difference among the 3 groups (F_2,171=103.427, P<.001). Before discharge, the mean score in the control group was 68.5 (SD 9.1; 95% CI 66.2‐70.8), while the mean scores in intervention group 1 and intervention group 2 were 82.3 (SD 8.1; 95% CI 80.2‐84.4) and 88.6 (SD 7.4; 95% CI 86.7‐90.5), respectively. The between-group difference remained statistically significant (F_2,171=95.682, P<.001; Table 2).

The corresponding ITT results based on the multiply imputed datasets are provided in Supplementary Table S2.

Immediately after education, the mean knowledge mastery score in the control group was 76.3 (SD 9.2; 95% CI 73.9‐78.7), which was significantly lower than the mean scores in intervention group 1 (88.5, SD 8.4; 95% CI 86.3‐90.7) and intervention group 2 (90.2, SD 8.1; 95% CI 88.1‐92.3). A statistically significant difference was observed among the 3 groups (F_2,171=48.362, P<.001). Before discharge, the control group had a mean knowledge mastery score of 71.8 (SD 9.8; 95% CI 69.2‐74.4), whereas intervention group 1 and intervention group 2 achieved higher mean scores of 84.7 (SD 8.9; 95% CI 82.3‐87.1) and 87.3 (SD 8.5; 95% CI 85.1‐89.5), respectively. The between-group difference remained statistically significant (F_2,171=46.175, P<.001). At 1 month after discharge, the mean knowledge mastery score further declined in the control group (65.2, SD 10.5; 95% CI 62.4‐68.0). In contrast, both intervention groups maintained relatively higher levels of knowledge mastery, with mean scores of 78.6 (SD 9.3; 95% CI 76.1‐81.1) in intervention group 1 and 83.5 (SD 8.9; 95% CI 81.2‐85.8) in intervention group 2. The overall difference among the 3 groups remained statistically significant (F_2,171=54.793, P<.001; Table 3).

Significant differences were observed among the 3 groups across all dimensions of education satisfaction, including education content, education method, education time, education effect, and overall satisfaction (all P<.001). Across all items, the 95% CIs of the 2 intervention groups were consistently higher than those of the control group, with minimal overlap between groups. This indicates that the observed differences in education satisfaction were not only statistically significant but also stable and precise, reflecting a reliable improvement associated with the intervention (Table 4).

Among the 3 groups, significant differences in treatment adherence were observed across all dimensions, including drug treatment adherence, lifestyle adjustment, follow-up review, and total adherence score (all P<.001). For drug treatment adherence, the control group scored a mean of 7.6 (SD 1.5; 95% CI 7.2‐8.0), which was lower than the mean scores for intervention group 1 (8.5, SD 1.2; 95% CI 8.2‐8.8) and intervention group 2 (9.1, SD 0.9; 95% CI 8.9‐9.3), with a statistically significant overall difference (F_2,171=25.364, P<.001). Similar patterns were observed for lifestyle adjustment and follow-up review adherence. The 95% CIs of both intervention groups were consistently higher than those of the control group, with limited overlap, indicating that the observed improvements in treatment adherence were stable and precise. Regarding the total adherence score, the control group achieved a mean of 21.6 (SD 4.2; 95% CI 20.5‐22.7), whereas intervention group 1 and intervention group 2 achieved higher mean scores of 24.7 (SD 3.5; 95% CI 23.8‐25.6) and 26.6 (SD 2.8; 95% CI 25.9‐27.3), respectively. The between-group difference remained statistically significant (F_2,171=32.593, P<.001; Table 5).

One month after discharge, significant differences were observed among the 3 groups in all SF-36 dimensions, including physical functioning, role-physical, bodily pain, general health, vitality, social functioning, role-emotional, and mental health (all P<.001). For each dimension, the 95% CIs of both intervention groups were consistently higher than those of the control group, indicating that the observed improvements in health-related quality of life were robust and precise.

Notably, intervention group 2 generally showed narrower 95% CIs, suggesting less variability and greater consistency in patient-reported outcomes compared with the control group (Table 6).

Before education, there were no significant differences in anxiety nor depression scores among the 3 groups (anxiety: F_2,171=0.068, P=.93; depression: F_2,171=0.185, P=.83), indicating baseline comparability. One month after discharge, significant differences were observed among the groups for both anxiety and depression scores (P<.001). Specifically, for anxiety, the control group scored a mean of 6.5 (SD 2.9; 95% CI 5.8‐7.2), intervention group 1 scored a mean of 5.2 (SD 2.5; 95% CI 4.6‐5.8), and intervention group 2 scored a mean of 4.3 (SD 2.2; 95% CI 3.8‐4.8). For depression, the mean scores were 5.8 (SD 2.6; 95% CI 5.2‐6.4), 4.6 (SD 2.3; 95% CI 4.1‐5.1), and 3.7 (SD 2.0; 95% CI 3.3‐4.1), respectively.

The 95% CIs of the intervention groups did not overlap with those of the control group, indicating that the reductions in anxiety and depression after intervention were stable and precise (Table 7).

This RCT investigated the effectiveness of an AI-assisted patient education system integrating voice cloning technology and compared the educational outcomes of physician voice cloning versus patient self-voice cloning. In addition, the study explored the feasibility of using ChatGPT as a supportive tool to assist with evaluating education effectiveness. Consistent with our 3 prespecified hypotheses, the results demonstrated that AI-assisted voice cloning–based education significantly improved education content compliance, knowledge mastery, satisfaction, treatment adherence, and short-term psychological and quality-of-life outcomes compared with traditional education. These findings are consistent with prior research indicating that technology-assisted and personalized education can enhance patient engagement and learning outcomes [10,22]. Notably, education delivered using patients’ own cloned voices yielded superior effects compared with physician voice cloning across multiple outcome domains, supporting the hypothesis that self-referential personalization enhances educational effectiveness.

Voice cloning technology represents an emerging application of AI in patient education, offering distinct advantages over conventional education models. By enabling repeated exposure to standardized educational content delivered in a familiar and emotionally resonant voice, voice cloning may enhance attention, comprehension, and memory consolidation [30,31]. Consistent with previous studies demonstrating that familiar auditory cues improve trust and acceptance in health communication, both AI-assisted intervention groups in this study outperformed the traditional education group in education compliance, knowledge mastery, and satisfaction [30,32,33]. Beyond confirming the general effectiveness of voice-based AI education, this study provides novel evidence regarding differences between voice sources. Patients who received education via their own cloned voice achieved higher compliance, satisfaction, knowledge retention, treatment adherence, and better short-term psychological outcomes than those educated using physician voice cloning. Previous studies have largely focused on clinician-narrated or professionally recorded content [34], whereas empirical comparisons involving patient self-voice have been scarce. The observed superiority of self-voice education may be explained by the self-reference effect, whereby information related to the self is processed more deeply and remembered more effectively [17,35,36]. Research in cognitive psychology suggests that self-related stimuli—particularly self-generated or self-similar auditory information—enhance attention, emotional engagement, and memory encoding [16,35]. In this study, hearing one’s own voice narrate medical information may have strengthened self-identification and personal relevance, thereby reinforcing learning and adherence. Furthermore, the self-voice group demonstrated superior outcomes in multiple quality-of-life domains and had greater reductions in anxiety and depression. These findings align with evidence that personalized and self-relevant health communication can enhance perceived control, self-efficacy, and emotional regulation among patients [14,36]. Importantly, the reported effect sizes indicated moderate-to-large intervention effects, suggesting that the observed improvements were not only statistically significant but also clinically meaningful. Reporting both effect sizes and confidence intervals enhances transparency and interpretability and is consistent with CONSORT 2025 recommendations for RCTs [28].

This study also explored the use of ChatGPT as an auxiliary tool to assist with evaluating patient education outcomes based on standardized educational content. Pre-experimental testing demonstrated good agreement between ChatGPT-assisted scoring and expert evaluation, suggesting that LLMs may support structured and reproducible assessment under controlled conditions. These findings are consistent with emerging literature indicating that LLMs can perform semantic analysis and text evaluation tasks with acceptable reliability in medical and educational contexts [6,37]. Compared with traditional manual evaluation, AI-assisted assessment may offer advantages in efficiency, scalability, and consistency, particularly in settings with limited human resources. However, ChatGPT should not be regarded as a fully objective evaluator. Assessment outcomes may still be influenced by prompt design, model architecture, and algorithmic biases. Therefore, AI-assisted evaluation should be considered a complementary tool rather than a replacement for expert assessment, and its use should be accompanied by appropriate human oversight [38,39].

Several limitations should be acknowledged. First, the follow-up period was limited to 1 month, restricting conclusions regarding long-term adherence, psychological outcomes, and sustained quality of life benefits. Longer follow-up periods are needed to assess the durability of intervention effects, as recommended in previous digital health education studies [40]. Second, this was a single-center study involving hospitalized patients with relatively homogeneous educational backgrounds, which may limit generalizability. Prior research suggests that educational level and health literacy can influence the effectiveness of digital health interventions [41]. Multicenter studies with more diverse populations are therefore warranted. Third, although voice cloning technology offers a promising approach, current speech synthesis systems still have limitations in emotional nuance and naturalness, which may affect user engagement [13]. Continued technological advancements may further enhance the effectiveness of AI-assisted education. Finally, although AI-assisted education may reduce clinical workload and improve efficiency, this study did not include a formal cost-effectiveness analysis. Future research should systematically evaluate economic outcomes to support large-scale implementation decisions [40]. Because ChatGPT is a continuously updated product, we standardized the scoring workflow (fixed rubric and prompts, prespecified audio format, and single scoring per recording) and provide the complete scoring materials to facilitate reproducibility. Additionally, voice-based delivery systems may misinterpret medication names or specialized terminology, underscoring the need for clinician review of generated content [42].

Implementation should address unit- and hospital-level factors separately. Targeted strategies should improve staff training, digital access, and policy support. Addressing demographic disparities is key to promoting equity and care quality.

This study introduces an innovative patient education model integrating AI voice cloning and ChatGPT, representing a novel approach distinct from previous studies that primarily relied on standard text-to-speech or professionally recorded content. The key innovation lies in using patients’ own cloned voices for health education delivery, leveraging the self-reference effect to enhance learning outcomes. Compared with prior research focusing on clinician-narrated content, this study provides the first empirical evidence that self-voice education produces superior outcomes across multiple domains including compliance, satisfaction, and psychological well-being. These findings contribute to the field by establishing a theoretical and practical framework for personalized AI-driven patient education. In real-world clinical settings, this approach offers a scalable, cost-effective solution to enhance patient engagement, particularly valuable in resource-limited environments where individualized education is challenging to deliver. Future research should focus on multicenter validation, longer follow-up periods, and exploration of optimal voice cloning parameters to maximize educational effectiveness.