Augmented Reality Trains Surgeons Well, but Does the Skill Survive the Operating Room?

Augmented reality has become one of the most seductive promises in surgical education. Unlike virtual reality, which replaces a trainee’s entire field of view with a synthetic environment, AR overlays digital cues, annotations, and three-dimensional guidance directly onto the real world, preserving the tactile feedback that surgeons depend on. Head-mounted displays and projector-based telestration systems can trace critical anatomy onto a laparoscopic video feed or project a patient-specific aneurysm onto a physical model, and a growing body of evidence shows that trainees who learn with these tools perform measurably better during training sessions. But a new systematic review published in Global Surgical Education, the journal of the Association for Surgical Education, asks a harder and far more consequential question: does any of that skill actually last, and does it survive the transition to a real operating theatre?

The review, led by Ee Hng Ian Lim, Aaleyah Adam, and Athanasios Hassoulas of Cardiff University School of Medicine, is the first to formally appraise skill retention and operative transfer as outcomes in their own right. Previous syntheses, including recent meta-analyses by El Ashry and colleagues, Xiong and colleagues, and Farooq and colleagues, had established with reasonable consistency that AR improves technical performance during training. None of them, however, measured whether that performance persists without continued practice, or whether it carries over into live surgery. The Cardiff team searched PubMed/MEDLINE, EMBASE, the Cochrane Library, and CINAHL from January 2000 to March 2026, following PRISMA 2020 guidance and prospectively registering the protocol on PROSPERO. Of 304 records screened after de-duplication, 18 underwent full-text review and just nine studies, encompassing 324 participants, met the inclusion criteria.

The included studies spanned four specialties: five in laparoscopic or visceral surgery, two in urology, and one each in orthopaedics and neurosurgery. All were prospective comparative designs, seven of them randomised controlled trials, and all used validated assessment instruments such as OSATS, GOALS, and GEARS rather than subjective measures of confidence or satisfaction. Eligibility was deliberately strict: studies had to report skill retention beyond the immediate training session, or transfer to live intraoperative or high-fidelity ex-vivo performance. Purely virtual systems without a physical component were excluded, as were studies reporting only satisfaction outcomes. Methodological quality was assessed with the Medical Education Research Study Quality Instrument, or MERSQI, and risk of bias in randomised trials with the Cochrane RoB2 tool, with particular attention to whether outcome assessors were blinded.

On retention, the picture is encouraging but bounded. Four studies assessed performance after a delay, with intervals ranging from one to eight weeks. The strongest evidence came from Shepard and colleagues, who found that checklist scores for ultrasound-guided percutaneous nephrolithotomy access remained significantly improved from baseline at eight weeks at both lower pole and upper pole sites, with only modest, non-significant decline from post-test to retention assessment. Brinkman and colleagues found that laparoscopic training gains on the LAP Mentor II simulator were maintained without significant decline at one week. Dodier and colleagues reported that aneurysm occlusion rates in an AR-guided neurosurgical simulator improved significantly at fourteen days, though between-group comparisons did not reach significance and baseline imbalances complicated interpretation. Notably, Sugand and colleagues used a one-week washout design rather than a true retention paradigm, so the reviewers caution it cannot be read as confirmatory retention evidence.

Transfer results split along a striking line. Five studies assessed transfer to high-fidelity ex-vivo models, and four of the five favoured AR on validated instruments. All three studies assessing transfer to porcine box-trainer cholecystectomy reported significant AR advantages: Felinska and colleagues found higher global and task-specific OSATS scores; Wild and colleagues reported higher GOALS and OSATS scores along with lower complication rates of 60 percent versus 86.7 percent; and Cizmic and colleagues, across a ten-session training programme, found the AR group superior on GOALS and OSATS scores, critical view of safety achievement, completion rates, and complication rates. Chowriappa and colleagues found significantly higher GEARS scores and lower NASA-TLX time pressure and effort in AR-trained urology residents performing robot-assisted urethrovesical anastomosis, with a crossover arm replicating the findings.

Then comes the outlier that defines the field’s largest evidence deficit. Huettl and colleagues conducted the only study of any AR platform assessing performance in a live operating theatre, using real-time AR annotation during laparoscopic surgery. The result: no significant difference between AR and control groups in GOALS total scores, critical view of safety achievement, or operative time. Intriguingly, the AR group received 59.4 percent fewer gestural and 36.1 percent fewer verbal corrections from trainers, yet were rated as receiving lower-quality assistance, suggesting AR changed the structure of the teaching interaction rather than the participants’ underlying technical performance. The reviewers note several possible explanations for the divergence: ex-vivo settings strip away the unpredictable anatomy, time pressure, and consequences of error that characterise real operations, and Huettl’s real-time annotation modality differs fundamentally from the pre-operative simulation training used in every ex-vivo study.

One consistent null finding deserves attention. Operative time showed no significant difference in any of the eight studies that measured it, regardless of specialty, platform, or comparator. The authors argue this is probably a ceiling effect rather than a true null: in novice populations, duration is driven more by hesitation and error recovery than by movement quality, so AR may sharpen precision without shortening procedures. This fits the Fitts-Posner model of motor learning, in which qualitative gains precede speed and automaticity. Chowriappa’s cognitive load data, showing significantly lower perceived time pressure and effort in AR-trained participants, is at least consistent with AR functioning as a cognitive scaffold that offloads spatial information from working memory.

The methodological caveats are substantial. All nine studies scored as high quality on MERSQI, ranging from 13.5 to 16.5 out of 18, yet no study achieved an overall low-risk judgement on RoB2: seven raised some concerns and two were rated high risk, driven predominantly by unblinded outcome assessment. In Felinska’s study, gaze data annotation was performed by an unblinded assessor; in Wild’s, the training tutor assessed outcomes in real time with the AR overlay still visible, precluding blinding altogether. Randomisation sequence generation was inadequately reported in eight of nine studies. No study reported effect sizes, meaning statistical significance cannot be read as educational or clinical meaningfulness, and no study measured patient-level outcomes, complication rates in patients, or safety data. The maximum follow-up across the entire evidence base was eight weeks, leaving consolidation beyond that window entirely uncharted.

The reviewers’ conclusions are deliberately circumscribed. AR simulation produces measurable technical benefits in controlled ex-vivo and simulator-based settings, predominantly in novice learners, which justifies cautious integration as a pre-operative or early-curriculum adjunct, most defensibly for laparoscopic cholecystectomy and robotic surgery where the evidence is most internally consistent. What is not supported matters just as much: nothing demonstrates benefit beyond an eight-week horizon, procedural efficiency gains, or AR as a substitute for conventional intraoperative supervision. The highest research priority is adequately powered, pre-registered, blinded intraoperative transfer trials reaching Kirkpatrick Levels 3 and 4, alongside retention studies beyond three months, enrolment of senior trainees, mandatory effect size reporting, and standardised platform characterisation. Until then, the authors argue, AR should be treated as complementary rather than standalone, a genuinely promising but evidentially immature technology that must meet the same methodological bar as any other training intervention before system-level adoption.

Subject of Research: Augmented reality simulation for surgical skill retention and transfer to operative performance

Article Title: Augmented reality simulation for surgical skill retention and operative transfer: a systematic review

Article References: Lim, E. H. I., Adam, A., & Hassoulas, A. (2026). Augmented reality simulation for surgical skill retention and operative transfer: a systematic review. Global Surgical Education – Journal of the Association for Surgical Education, 5(1), Article 182. https://doi.org/10.1007/s44186-026-00585-5

Image Credits: AI Generated

DOI: 10.1007/s44186-026-00585-5

Keywords: augmented reality, surgical education, simulation training, skill retention, skill transfer, systematic review, laparoscopic surgery, surgical training, operating room, clinical competence, medical education, randomised controlled trials