AI-Assisted Feedback for Learning Standardized Tuina Skills

Manual rehabilitation skills are complex sensorimotor skills that require learners to integrate anatomical localization, body mechanics, force control, rhythm, consistency, safety, and professional communication. Conventional instruction generally relies on teacher demonstration, observation, and verbal feedback. This approach may be limited by instructor availability and may not provide continuous and objective information about force and rhythm during each practice attempt.

Sensor-based training systems can quantify performance characteristics such as mean force, deviation from the target force, force variability, operating frequency, rhythm variability, and the proportion of time spent within the target range. However, novice learners may have difficulty converting numerical data and performance curves into specific strategies for improvement. Generative AI may help translate objective performance data into immediate, structured, and actionable feedback.

This is a single-center, prospective, assessor-masked, three-arm, parallel-group randomized educational study. Eighty-one rehabilitation trainees will be allocated in a 1:1:1 ratio to an AI-supported personalized feedback group, a sensor-based data feedback group, or a traditional instructor feedback group. All participants will receive the same instructional materials, standardized demonstration, safety instructions, practice tasks, number of practice opportunities, and total training time.

The AI-supported personalized feedback will follow a structured task-gap-action format. The feedback will identify the target skill, describe the difference between the trainee's measured performance and the predefined target, and provide specific recommendations for the next practice attempt. The AI system will receive only coded sensor measurements and predefined task standards. It will not receive participant names, student identification numbers, facial images, voices, patient information, or clinical records. AI-generated feedback will be restricted to educational use and will be subject to instructor oversight and predefined safety rules.

Assessments will be performed at baseline, immediately after the training intervention, and four weeks after training. The primary outcome is the total OSCE score four weeks after training, representing retention of the standardized Tuina skill. OSCE performances will be coded and evaluated independently by assessors who are unaware of group allocation. Objective sensor-based measures will be used to evaluate force accuracy, force variability, rhythm accuracy, rhythm stability, and the proportion of performance within the target range.

Secondary evaluations will include immediate OSCE performance, performance on a related transfer task, the number of practice attempts required to reach a predefined competency standard, learning self-efficacy, cognitive load, learning engagement, satisfaction, trust in AI feedback, instructor feedback time, and the frequency of AI-generated feedback requiring instructor modification or correction.

The study is an educational intervention and will not involve patient treatment, invasive procedures, biological specimen collection, or clinical decision-making. All research-related Tuina practice will be performed on a mechanical simulation model equipped with a pressure-sensing system.

• AI-Supported Personalized Feedback (Behavioral): Participants will complete four standardized training sessions over two weeks, with each session lasting approximately 45 to 60 minutes. A pressure-sensing system will measure force accuracy, force variability, operating frequency, rhythm stability, and time within the target range. After each practice attempt, a generative AI system will provide feedback using a structured task-gap-action format. The feedback will describe the target task, identify differences between measured performance and the predefined standard, and recommend specific actions for the next practice attempt. AI output will be restricted to educational use and overseen by instructors.
• Sensor-Based Data Feedback (Behavioral): Participants will complete four standardized training sessions over two weeks, with each session lasting approximately 45 to 60 minutes. After each practice attempt, participants will view numerical sensor measurements, force-time curves, rhythm information, and predefined target ranges. No generative AI interpretation, personalized action plan, or AI-generated recommendation will be provided.
• Traditional Instructor Feedback (Behavioral): Participants will complete four standardized training sessions over two weeks, with each session lasting approximately 45 to 60 minutes. Instructors will observe performance and provide conventional verbal feedback based on the standardized teaching protocol. Participants will not receive sensor-derived performance displays or AI-generated feedback.

Primary outcomes

• Standardized Tuina Skill Performance Measured by Total OSCE Score [Four weeks after completion of the training intervention]: Participants will complete a standardized Objective Structured Clinical Examination (OSCE) assessing anatomical localization, body mechanics, force control, rhythm, procedural consistency, safety, communication, and overall performance. Coded performance recordings will be independently scored by two assessors who are unaware of group allocation. The mean of the two assessors' total scores will be used for analysis. Higher scores indicate better standardized Tuina skill performance.

Secondary outcomes

• Immediate Standardized Tuina Skill Performance Measured by Total OSCE Score [Immediately after completion of the training intervention]: Standardized Tuina skill performance will be independently evaluated by two masked assessors using the same OSCE scoring rubric as the primary outcome. The mean of the two assessors' total scores will be analyzed. Higher scores indicate better performance.
• Absolute Error From the Target Force [Baseline, immediately after the intervention, and four weeks after the intervention]: The pressure-sensing system will calculate the absolute difference, in newtons, between the participant's mean applied force and the predefined target force. A lower absolute error indicates greater force accuracy.
• Coefficient of Variation of Applied Force [Baseline, immediately after the intervention, and four weeks after the intervention]: Force stability will be calculated as the standard deviation of applied force divided by the mean applied force and expressed as a percentage. A lower coefficient of variation indicates more stable force control.
• Absolute Error From the Target Operating Frequency [Baseline, immediately after the intervention, and four weeks after the intervention]: The absolute difference, in hertz, between the participant's measured operating frequency and the predefined target frequency will be calculated. A lower value indicates greater rhythm accuracy.
• Proportion of Practice Time Within the Target Performance Range [Baseline, immediately after the intervention, and four weeks after the intervention]: The pressure-sensing system will calculate the percentage of recorded practice time during which both applied force and operating frequency remain within the predefined acceptable target ranges. Values range from 0% to 100%, with higher percentages indicating better performance.
• Performance on a Related Tuina Skill Transfer Task [Immediately after the intervention and four weeks after the intervention]: Participants will perform a related standardized Tuina task that was not practiced in the same form during training. Performance will be evaluated using a predefined transfer-task rubric by assessors unaware of group allocation. Higher scores indicate better transfer of the learned skill.
• Number of Practice Attempts Required to Reach the Competency Standard [During the four training sessions over two weeks]: The number of practice attempts required for a participant to first meet the predefined competency threshold will be recorded. A lower number of attempts indicates greater learning efficiency. Participants who do not reach the threshold will be recorded as not achieving competency during the intervention period.
• Proportion of AI-Generated Feedback Requiring Instructor Modification [During the four training sessions over two weeks]: For participants in the AI-supported feedback arm, instructors will record whether each AI-generated feedback message is accepted without modification, modified for accuracy or clarity, or blocked because it is potentially inappropriate or unsafe. The outcome will be calculated as the percentage of AI-generated feedback messages requiring modification or blocking.

Sex: All, ages: 18 Years to any, accepts healthy volunteers

Inclusion Criteria:

* Aged 18 years or older.
* Currently enrolled in or receiving training in rehabilitation medicine, physical therapy, rehabilitation therapy, or a related health profession.
* Has not previously achieved the predefined competency standard for the standardized Tuina skill evaluated in this study.
* Able to understand the study instructions and independently complete the training tasks and study questionnaires.
* Able and willing to attend all scheduled training sessions and the four-week follow-up assessment.
* Willing to participate voluntarily and provide written informed consent.

Exclusion Criteria:

* Current acute injury, clinically significant pain, or functional limitation involving the hand, wrist, upper limb, shoulder, neck, or lower back that may interfere with repeated manual skill practice.
* Previous systematic training in the same standardized Tuina technique with performance at or above the predefined competency standard.
* Any medical, physical, cognitive, or psychological condition that, in the investigator's judgment, may make participation unsafe or prevent valid completion of the study procedures.
* Unable to understand the study procedures or provide informed consent.
* Direct involvement in the design, randomization, intervention delivery, outcome assessment, data management, or statistical analysis of this study.
* Concurrent participation in another training study that may substantially affect performance of the standardized Tuina skill.

• Xing-Chen Zhou, Principal investigator, The First Affiliated Hospital, Zhejiang University School of Medicine, 79 Qingchun Rd., Shangcheng District, Hangzhou, Zhejiang