AI-Assisted Feedback for Learning Standardized Tuina Skills

Overview

The goal of this educational study is to determine whether personalized feedback generated using sensor data and generative artificial intelligence (AI) can improve the learning of standardized Tuina skills among rehabilitation trainees. Tuina is a form of manual therapy that requires learners to control the location, force, rhythm, and consistency of their hand movements. A total of 81 rehabilitation trainees will be randomly assigned to one of three training groups: AI-supported personalized feedback, sensor-based data feedback without AI-generated recommendations, or traditional instructor feedback. All groups will receive the same standardized demonstration, training tasks, practice duration, and number of practice sessions. The main question is whether trainees receiving AI-supported personalized feedback achieve better retention of standardized Tuina skills four weeks after training. The researchers will also compare immediate skill performance, force and rhythm control, transfer of skills to a related task, learning efficiency, self-efficacy, cognitive load, satisfaction, and the safety and acceptability of AI-generated feedback. Skill performance will be assessed using a blinded Objective Structured Clinical Examination (OSCE) and objective sensor-based measurements. The study activities will be conducted using a mechanical simulation model and a pressure-sensing system, rather than on patients.

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.