Generative AI-Assisted Personalized Feedback for Laparoscopic Skill Learning in Medical Undergraduates

Laparoscopic skills training requires learners to develop spatial orientation, coordinated use of both hands, stable instrument control, and accurate manipulation of materials. Formative feedback helps learners recognize technical problems and organize subsequent practice. Providing frequent individualized feedback can require substantial instructor time, while differences in how performance information is interpreted may affect the consistency of training recommendations. Generative artificial intelligence (AI) may support this process by organizing multiple sources of structured performance information into focused and actionable guidance.

This study examines the additional educational value of AI-assisted personalized feedback in a setting where both study groups receive structured feedback. The active comparator uses predefined rules and templates to provide training recommendations. The scientific question is whether generative AI can improve the integration and explanation of performance evidence, identify technical bottlenecks, and adapt practice recommendations within the study's predefined priority and target rules. The study is conducted in a laparoscopic simulation environment and evaluates learning among medical undergraduates.

The training program separates practice tasks from standardized assessment tasks. Transfer training uses a 14-ball pegboard task that requires controlled grasping, movement, placement, and coordinated instrument use. Cutting training uses study-developed shape-cutting tasks that require traction, exposure, controlled cutting, directional adjustment, and protection of the training material. Fundamentals of Laparoscopic Surgery (FLS) tasks are reserved for scheduled assessments. This separation is intended to evaluate whether skills developed through the training exercises transfer to standardized tasks that are not used for routine repeated practice.

Both groups follow the same training sequence and planned training dose. Training is generally scheduled once per week, with 60 minutes of net practice per session organized into 6-minute rounds. Equipment, materials, general task instructions, and feedback opportunities are standardized. Instructors provide common explanations, safety corrections, and assistance with equipment problems. Additional group-specific individual technical coaching outside the assigned feedback approach is not part of the planned intervention. Actual attendance, training completion, and relevant deviations are recorded.

Feedback is based on structured information confirmed by the research team. Inputs include task duration, completion status, objective error events, observer-rated technical performance, and participant self-assessments. Technical ratings are derived from independent assessments of anonymized videos, with discrepancies handled through the predefined review procedure. The feedback systems use these confirmed records and do not independently score raw videos. Missing, non-evaluable, or inconsistent information is handled according to the study's data review rules and is not replaced by invented performance data.

In the AI group, predefined prompts guide the system to integrate the available evidence, identify one or two priority technical problems, and translate these problems into specific practice recommendations. The feedback is intended to explain the relevant technical bottleneck, organize a short sequence of practice steps, and provide observable goals for the next training session. Recommendations remain subject to predefined task rules and target criteria. The system is not permitted to create unsupported numerical thresholds, infer unobserved behavior, or attribute performance to unmeasured psychological or personal characteristics.

The template group receives feedback generated by a local deterministic rule-based system. Predefined triggering conditions, priority rules, target criteria, and fixed statements determine the feedback content. Participant-specific performance values are inserted into the applicable statements. This system can identify current training priorities, recommend practice activities, and specify next-session goals. It therefore provides an active educational comparison with structured guidance. Both groups receive feedback with a similar basic organization, presentation format, and opportunity for reading before practice.

Four formative feedback events are scheduled, with two events in the transfer phase and two in the cutting phase. In each phase, the first feedback event uses the relevant phase-entry performance data and is delivered before the first training session. The second uses the same participant's previous frozen raw input together with current corresponding performance data and is delivered before the second training session. The longitudinal comparison is intended to identify improvement, persistent problems, worsening performance, or new errors and to adjust the current recommendations accordingly. Previously generated feedback text is not required as input for the second feedback event.

The AI workflow and template rules will be finalized and frozen before formal study implementation. Feedback intended for delivery undergoes a predefined release procedure. For AI outputs, review focuses on substantive issues such as incorrect use of input data, unsupported factual statements, errors in longitudinal comparison, unsafe or non-verifiable recommendations, and missing essential content. Outputs with blocking errors are regenerated or withheld according to the study procedures. Reviewers do not freely rewrite feedback solely to improve style or express a personal preference for another reasonable training recommendation. Original outputs, review decisions, and delivered versions are retained.

Participants are not proactively informed whether their feedback was produced by generative AI or the template system. Feedback uses a common visual format and delivery approach without source labels. This procedure is intended to reduce expectations associated with the feedback source, although writing style may allow some participants to form their own impressions. Technical assessors independently review videos identified by anonymized codes without access to study group or actual assessment time point. Staff responsible for generating, reviewing, and delivering feedback are not blinded and do not perform the blinded technical outcome assessments.

AI inputs contain only the de-identified structured information needed for the assigned feedback event. Raw training videos, direct personal identifiers, and patient information are not submitted to the generative AI workflow. Feedback generation is restricted to the relevant participant and the permitted current and previous training data. Future standardized assessment results and retention data are excluded from formative feedback generation. The intervention is intended to support simulation-based learning and does not establish clinical competence or authorization to perform surgery.

The randomized comparison evaluates whether the AI-assisted approach adds learning benefits beyond structured rule-based feedback under otherwise standardized training conditions. The exploratory design supports estimation of the direction and magnitude of any additional benefit, together with its uncertainty. Interpretation will remain specific to the study population, simulation tasks, and feedback procedures used in this trial.