Introduction

Artificial intelligence has entered surgery less as an autonomous operator and more as an observer, assessor, and potential source of feedback. Contemporary systems can analyze operative video, instrument motion, simulator telemetry, procedural phases, error patterns, and other workflow characteristics. Some platforms are being evaluated as “surgical coaches” that may provide scalable feedback, objective assessment, and more standardized training.

The key clinical question isn’t just if AI can spot technical errors or give feedback. It is whether AI-assisted coaching improves outcomes that matter to patients and health systems. Better simulator scores are not equivalent to fewer bile duct injuries, anastomotic leaks, reoperations, readmissions, or deaths.

The early evidence is encouraging but incomplete. AI coaching can improve selected technical-performance measures under controlled simulation or structured training conditions. Real-world procedural data are beginning to emerge, particularly in laparoscopic cholecystectomy. Nevertheless, the field remains closer to surgical education and quality improvement than to proven patient-outcome improvement.

Why This Topic Matters Now

Surgical coaching has a strong conceptual foundation. Technical skill varies among surgeons, and observational studies have found associations between higher video-assessed technical-skill ratings and lower complication rates after selected operations.

In bariatric surgery, Birkmeyer et al. found that lower peer-rated technical skill was associated with higher risk-adjusted rates of complications, reoperation, readmission, and emergency department visits. In a separate study of laparoscopic right hemicolectomy, Stulberg et al. found that higher technical-skill ratings were associated with lower rates of complications, unplanned reoperation, and a composite of death or serious morbidity.

These studies support an association between intraoperative technical performance and selected patient outcomes. They do not establish that increasing a performance score through AI coaching will necessarily improve patient outcomes.

Traditional video-based coaching can improve technical performance, but it requires expert time, structured review, repeated feedback, and a culture in which case review is treated as professional development rather than surveillance. AI offers a potential way to scale parts of this process by screening video, identifying procedural phases, detecting selected error patterns, quantifying performance features, and presenting standardized feedback.

The risk is that descriptions of AI capability may advance faster than the evidence. A system can’t be called effective for patient outcomes unless the study shows it improves the composite performance score. In surgery, careful evidence language is part of patient-safety practice.

What Counts as an AI Surgical Coach?

For this review, an AI surgical coach is a system that uses machine learning, computer vision, automated video analysis, sensor data, simulator data, or related computational methods to assess surgical performance and provide feedback, guidance, or educational support.

These systems may be used in several settings:

• In simulation-based training, learners perform standardized tasks outside the operating room.
• In post-case video review, AI analyzes recorded operative footage and identifies potential coaching opportunities.
• In AI-augmented human coaching, a surgeon educator uses AI-derived performance information to individualize feedback.
• In more advanced applications, real-time systems may identify anatomy, procedural phases, instruments, or risk-relevant events during an operation.

This article focuses on coaching, performance assessment, and performance improvement. It does not address autonomous robotic surgery.

Current Evidence: Encouraging, but Mostly Indirect

Technical Skill Is Associated With Clinical Outcomes

The rationale for surgical coaching is strengthened by observational evidence linking technical-skill ratings with postoperative outcomes in selected procedures.

Birkmeyer et al. found that peer-rated technical skill during laparoscopic gastric bypass was associated with risk-adjusted complication rates. Stulberg et al. subsequently found that technical-skill scores based on laparoscopic right hemicolectomy videos were associated with complication and reoperation rates following colectomy.

These studies establish neither causation nor a validated surrogate endpoint. Surgeon experience, judgment, case selection, team performance, institutional resources, and other factors may contribute to the observed relationships. The studies suggest that improving technical skills can lead to safer surgery. However, they do not prove that AI coaching will provide clinical benefits.

Video-Based Coaching Provides the Educational Foundation

Video-based coaching predates contemporary AI systems. A systematic review and meta-analysis by Augestad et al. found that video-based coaching improved technical performance among medical students and surgical residents. However, the included studies varied substantially in procedure, learner level, intervention design, coaching method, assessment instrument, and outcome measure.

A later meta-analysis focused on surgical residents found greater pre-to-post improvement among residents receiving video-based coaching, but also found substantial heterogeneity and no clearly significant difference in post-coaching scores between groups.

AI coaching should therefore be understood as an extension of established coaching methods rather than a replacement for the educational principles that make coaching effective. These principles include trust, repetition, specificity, expert interpretation, reflective practice, and psychologically safe feedback.

AI Simulation Trials Show Measure-Specific Benefits

Randomized simulation trials have found that AI tutoring can improve selected performance measures, but the results depend heavily on the assessment method and AI system.

In a randomized clinical trial involving 70 medical students, Fazlollahi et al. compared an AI tutor with remote expert instruction and no feedback during simulated neurosurgical training. Participants receiving AI tutoring achieved higher machine-derived performance scores during practice and a realistic transfer task. However, global Objective Structured Assessment of Technical Skills, or OSATS, ratings did not differ significantly among groups. This discrepancy is important because it shows that apparent benefit can depend on whether performance is evaluated by the AI system or by blinded human raters.

Yilmaz et al. compared real-time AI feedback, in-person expert instruction, and no real-time feedback in 97 medical trainees. AI feedback produced higher composite performance scores than the comparison groups during training. Expert-rated OSATS scores were similar between AI and in-person instruction, and participants receiving AI feedback reported greater extraneous cognitive load.

A 2026 systematic review and meta-analysis identified 40 studies for narrative synthesis and included four studies with 268 participants in the quantitative analysis. AI tutoring was associated with a small improvement in expert-rated OSATS scores, with a mean difference of 0.20 points. The certainty of evidence was low, the magnitude was of uncertain educational significance, and sensitivity analyses showed that the result was not robust to removal of individual studies. No significant difference was identified in machine-derived ICEMS scores, and AI tutoring was associated with greater extraneous cognitive load.

Taken together, the simulation evidence supports educational promise, but not uniform superiority. It also illustrates the hazards of treating an AI-generated score as an objective clinical truth. AI assessment systems operationalize performance standards derived from human-labeled data, model design choices, and selected training metrics.

AI-Augmented Human Coaching May Offer Complementary Benefits

One of the most clinically relevant findings is that AI may be most useful when it supports rather than replaces expert educators.

In a randomized clinical trial of 88 medical students, Giglio et al. compared intelligent tutoring alone, expert instruction using the same scripted feedback as the AI tutor, and personalized expert instruction informed by AI-derived error data. The AI-informed personalized expert group achieved higher performance scores and better skill transfer than intelligent tutoring alone.

A smaller randomized crossover study by Yilmaz et al. examined the sequence of AI and expert instruction. The results suggested that the timing and sequence of the two instructional approaches may influence learning. However, the trial included only 25 students, used simulation-based outcomes, and produced findings that should be interpreted as exploratory rather than definitive.

The current evidence does not establish a universally superior coaching model. It does suggest that AI and expert instruction have different strengths. AI can continuously quantify selected behaviors and deliver standardized feedback. Human educators can integrate anatomy, procedural strategy, case complexity, learner readiness, communication, judgment, and the wider clinical context.

The more appropriate comparison is therefore not simply “AI versus surgeon educator.” It is how AI-derived information can be incorporated into expert-led education without displacing clinical judgment or narrowing competence to what the model can measure.

Early Real-World Procedural Data Are Promising but Limited

The most clinically relevant early procedural evidence comes from laparoscopic cholecystectomy.

Wu et al. conducted a multicenter randomized controlled trial evaluating an AI-assisted coaching program for novice surgeons. Twenty-two surgeons from 10 hospitals were enrolled, and 18 completed the study. Surgeons assigned to the coaching program improved their Laparoscopic Cholecystectomy Rating Form scores over time and had higher scores than the self-learning group at the end of the study. Achievement of the critical view of safety increased within the coaching group.

This study is important because it goes beyond isolated simulation exercises. It evaluates performance during actual procedures, providing valuable insights. Nevertheless, it did not establish that AI coaching reduced bile duct injury, conversion to open surgery, readmission, reoperation, or other patient-centered outcomes.

The critical view of safety is a guideline-supported anatomic-identification strategy intended to reduce misidentification during laparoscopic cholecystectomy. The multisociety safe-cholecystectomy guideline conditionally recommends use of the critical view of safety for identification of the cystic duct and cystic artery. Much of the supporting evidence is low certainty, and improved critical-view achievement should not be treated as proof that an AI coaching system reduces bile duct injury.

Bile duct injury is uncommon. A systematic review of more than 500,000 laparoscopic cholecystectomies reported pooled bile duct injury rates of approximately 0.32% to 0.52%, with lower reported rates in more recent periods. Demonstrating a reliable reduction in such an uncommon outcome would require substantially larger studies, robust outcome ascertainment, and adjustment for surgeon experience, case complexity, institutional practices, and other confounders.

Table 1. What the Evidence Shows and What Remains Unproven

Evidence domain	What current evidence supports	What remains unproven
Observational technical-skill studies	Higher skill ratings are associated with fewer complications after selected procedures	That raising a skill score causes better patient outcomes
Traditional video-based coaching	Improvement in selected technical-performance measures	Uniform effectiveness across procedures, learners, and settings
AI simulation trials	Improvement in some machine-derived or expert-rated training measures	Consistent superiority to expert instruction or transfer to clinical outcomes
AI-augmented human instruction	Potential benefit from combining AI-derived data with personalized expert feedback	The optimal human-AI model across specialties and learner levels
Real-world cholecystectomy coaching	Improved procedural ratings and a safety-process measure in an early trial	Fewer bile duct injuries, complications, readmissions, or deaths
Surgical scene understanding	Feasibility of identifying phases, instruments, anatomy, and visual cues	Reliable clinical benefit during routine operative care

Clinical Relevance Beyond Surgery

Surgeons and surgical educators are the primary users of AI coaching systems. Still, the issue has broader relevance for internists, subspecialists, pharmacists, advanced practice clinicians, quality leaders, informaticians, and hospital administrators.

Many clinicians participate in preoperative optimization or postoperative complication management. They may be asked to interpret whether an AI-supported intervention has improved actual outcomes or only an educational or process measure.

Hospital committees are also increasingly asked to evaluate AI tools. These decisions require the same disciplined questions applied to drugs, devices, diagnostics, and procedural technologies:

• What is the intended use?
• What population, procedure, and environment were studied?
• Was the comparator appropriate?
• Was the endpoint technical, process-based, or clinical?
• Was the observed difference clinically meaningful?
• Was performance externally validated?
• How will errors, drift, privacy risks, workflow disruption, and automation bias be monitored?

AI-derived performance information may eventually influence credentialing, privileging, maintenance of competence, morbidity and mortality review, or quality dashboards. If poorly governed, these data may become punitive, biased, misleading, or detached from clinical context. If appropriately validated and implemented, they may support more consistent feedback and identify opportunities for improvement.

Safety, Regulatory, and Implementation Issues

AI surgical coaching tools operate across a varied regulatory and safety landscape. Some systems may function primarily as educational software. Others may provide patient-specific decision support, anatomy identification, risk alerts, or intraoperative guidance.

Whether a system is regulated as a medical device depends on its intended use, claims, functionality, and role in diagnosis or treatment. Institutions should distinguish a product’s general marketing language from its specific FDA marketing authorization, if any. The FDA’s AI-Enabled Medical Device List is useful but is not comprehensive, and inclusion does not imply authorization for every possible clinical use.

For products that influence patient care, institutions should verify:

• The authorized intended use and applicable labeling
• The patient population, procedure, and hardware environment in which the system was evaluated
• Whether local use falls within or outside the authorized use
• How software changes and model updates will be managed
• Whether the manufacturer has an appropriate validation and monitoring plan
• How clinicians will identify, report, and respond to incorrect outputs

The FDA’s final guidance on predetermined change control plans describes a framework through which manufacturers may prospectively specify certain modifications to AI-enabled device software functions and the methods used to develop, validate, and implement them. This does not eliminate the need for institutional monitoring, local governance, or review of whether the current software version remains appropriate for local use.

Important safety concerns include automation bias, excessive reliance on AI output, false reassurance, inaccurate anatomy recognition, poor performance in atypical anatomy, data drift, inadequate video quality, cybersecurity vulnerabilities, privacy loss, inequitable performance, and increased cognitive demand.

An erroneous anatomy label during a complex laparoscopic cholecystectomy could contribute to harm if it is accepted uncritically. The risk arises not only from the algorithm, but also from the clinician’s use of the output and the institution’s deployment, training, monitoring, and incident-response processes.

For most current applications, simulation-based feedback or post-case coaching with expert oversight represents a more conservative deployment model than patient-specific real-time guidance. This is a risk-management judgment, not a conclusion from comparative safety trials. Real-time intraoperative guidance requires a higher evidentiary and safety threshold when it can influence operative decisions.

Postmarket surveillance should include local incident reporting, performance audits, version tracking, and review of relevant manufacturer and FDA information. Reports in the FDA Manufacturer and User Facility Device Experience database can help identify potential safety signals, but passive adverse-event reports cannot establish event incidence or causality.

Table 2. Practical Questions Before Deploying an AI Surgical Coach

Deployment question	Minimum expectation	Risk addressed
Intended use	Education, quality improvement, or clinical guidance is explicitly defined	Prevents use beyond the supporting evidence or authorization
Validation	Procedure, learner level, patient population, equipment, and setting match local use	Reduces generalizability error
Outcome selection	Technical, process, and clinical endpoints are clearly distinguished	Prevents vague or inflated benefit claims
Human oversight	Qualified experts retain responsibility for interpretation and clinical decisions	Limits automation bias
Version control	Model and software versions are documented and changes are assessed	Detects unreviewed performance changes
Safety monitoring	Incident reporting, performance audits, and escalation pathways are established	Identifies harm, drift, and recurrent failure
Data governance	Consent, privacy, storage, access, retention, and secondary use are addressed	Protects patients and clinicians
Equity review	Performance is assessed across relevant case-mix and subgroup characteristics	Reduces biased implementation
Professional use	Coaching data are separated from punitive personnel action unless validity is established	Protects fairness and psychological safety

A Practical Approach for Clinicians and Health Systems

Initial implementation should generally focus on education and quality improvement rather than direct patient-care decision-making. The first question is whether the tool improves a meaningful and measurable behavior in the local environment.

A stepwise implementation pathway may include:

1. Expert review of the system’s intended use, output, and clinical face validity.
2. Retrospective validation using representative local operative videos or simulation data.
3. Silent prospective deployment without presenting outputs to clinicians.
4. Comparison with standard coaching, expert review, or existing training methods.
5. Limited educational use with clear human oversight.
6. Prospective monitoring of technical, process, usability, workload, equity, and patient-safety outcomes.
7. Defined criteria for pausing or withdrawing the system if performance deteriorates.

Institutions should also decide how feedback will be presented. A numerical score without explanation may be less useful than a specific coaching point linked to a video segment. Feedback should be timely, interpretable, technically valid, and educationally safe.

Public ranking or high-stakes credentialing based on immature AI metrics should be avoided. Before an AI measure is used for employment, privileging, disciplinary action, or public reporting, the institution should establish that the measure is valid, reliable, appropriately risk-adjusted, resistant to gaming, and relevant to the competency being assessed.

Patient and Procedure Selection

AI coaching is most plausible in procedures with high volume, reproducible steps, routinely captured video, and reasonably well-defined technical or safety behaviors.

Laparoscopic cholecystectomy is a logical early application because the critical view of safety is visible, teachable, and guideline-supported. Nevertheless, achieving critical view remains a process measure rather than a validated guarantee against bile duct injury.

Other laparoscopic and robotic procedures may be suitable because operative video and instrument data are frequently available. Suitability must be evaluated separately for each procedure, platform, population, and intended use.

Less appropriate early targets include rare procedures, operations with highly variable anatomy, emergent cases with unusual presentations, and settings in which the system has not been externally validated. Caution is also warranted when video quality is inconsistent, local workflow cannot support review and follow-up, or clinical staff cannot reliably recognize situations in which the model may fail.

What Would Count as Patient-Outcome Evidence?

AI surgical coaching may be considered an educationally useful intervention when it improves valid measures of learning, feedback quality, skill retention, or competency. Stronger evidence is required before it can be described as a patient-outcome or patient-safety intervention.

Relevant clinical endpoints will vary by procedure and may include:

• Postoperative complications
• Procedure-specific injury
• Conversion to open surgery
• Bleeding or transfusion
• Anastomotic leak
• Readmission
• Reoperation
• Length of stay
• Mortality
• Functional recovery
• Patient-reported outcomes
• Cost-effectiveness

Because many serious surgical complications are uncommon, large pragmatic trials, registry-linked studies, cluster-randomized trials, or stepped-wedge designs may be necessary. Studies should report case mix, surgeon experience, procedure complexity, institutional resources, learning-curve effects, implementation fidelity, and prespecified subgroup analyses.

Clinical AI studies should also use applicable reporting standards. CONSORT-AI provides AI-specific reporting items for randomized trials, SPIRIT-AI addresses clinical-trial protocols, and DECIDE-AI addresses early-stage clinical evaluation of AI decision-support systems.

Adherence to these standards can improve transparency, reporting completeness, replicability, and critical appraisal. It does not, by itself, ensure sound study design, unbiased execution, appropriate analysis, or clinically meaningful results.

Limitations of the Evidence

The current evidence base has several limitations.

Many studies are simulation-based, enroll medical students or early trainees, and use composite performance scores. These settings are valuable for evaluating educational efficacy but do not reproduce the full cognitive, team-based, ethical, and clinical demands of independent operative care.

The AI systems, procedures, datasets, feedback mechanisms, and outcome measures differ substantially. Findings from one system cannot be assumed to apply to another. Recent systematic reviews of surgical training and surgical scene understanding have identified small datasets, limited external validation, inconsistent reporting, algorithmic opacity, narrow procedural representation, and limited clinical integration.

Some studies use outcomes generated by the same AI framework that delivers or informs the training intervention. This creates a risk that the intervention is optimized to improve the metric by which it is judged. Independent human ratings, external validation, and patient-centered outcomes are therefore essential.

Technical-performance measures may be useful for formative feedback or competency assessment. They should not be presented as evidence of safer surgery or improved patient outcomes unless that relationship has been adequately validated.

The current literature also provides limited evidence on practicing surgeons, long-term skill retention, rare procedures, unusual anatomy, diverse patient populations, workflow effects, professional behavior, liability, and unintended consequences.

Future Directions

The next phase of research should move from proof-of-concept performance studies toward pragmatic clinical evaluation.

Future trials should include appropriate comparators, prespecified outcomes, external validation, version reporting, independent outcome assessment, and postdeployment surveillance. They should also evaluate workload, cognitive burden, surgeon acceptance, privacy, equity, liability, deskilling, overreliance, and the possibility that learners optimize their behavior for the algorithm rather than for the patient.

Research should identify which aspects of coaching are best delivered by AI, which require expert interpretation, and how feedback should change as learners progress from novice to independent practice.

The most plausible future model is not an autonomous AI coach replacing expert educators. It is a governed hybrid system in which AI identifies selected patterns and coaching opportunities while qualified surgeons preserve context, judgment, mentorship, and responsibility for patient care.

References

Augestad, K. M., Butt, K., Ignjatovic, D., Keller, D. S., & Kiran, R. (2020). Video-based coaching in surgical education: A systematic review and meta-analysis. Surgical Endoscopy, 34(2), 521-535. https://doi.org/10.1007/s00464-019-07265-0. PMID: 31748927.

Birkmeyer, J. D., Finks, J. F., O’Reilly, A., Oerline, M., Carlin, A. M., Nunn, A. R., Dimick, J., Banerjee, M., & Birkmeyer, N. J. O. (2013). Surgical skill and complication rates after bariatric surgery. New England Journal of Medicine, 369(15), 1434-1442. https://doi.org/10.1056/NEJMsa1300625. PMID: 24106936.

Brunt, L. M., Deziel, D. J., Telem, D. A., Strasberg, S. M., Aggarwal, R., Asbun, H., et al. (2020). Safe cholecystectomy multi-society practice guideline and state-of-the-art consensus conference on prevention of bile duct injury during cholecystectomy. Annals of Surgery, 272(1), 3-23. https://doi.org/10.1097/SLA.0000000000003791. PMID: 32404658.

Carstens, M., Vasisht, S., Zhang, Z., Barbur, I., Reinke, A., Maier-Hein, L., Hashimoto, D. A., & Kolbinger, F. R. (2025). Artificial intelligence for surgical scene understanding: A systematic review and reporting quality meta-analysis. npj Digital Medicine, 9(1), Article 59. https://doi.org/10.1038/s41746-025-02227-4. PMID: 41407878.

Escobar-Castillejos, D., Barrera-Animas, A. Y., Noguez, J., Magana, A. J., & Benes, B. (2025). Transforming surgical training with AI techniques for training, assessment, and evaluation: Scoping review. Journal of Medical Internet Research, 27, e58966. https://doi.org/10.2196/58966. PMID: 41252719.

Fazlollahi, A. M., Bakhaidar, M., Alsayegh, A., Yilmaz, R., Winkler-Schwartz, A., Mirchi, N., et al. (2022). Effect of artificial intelligence tutoring vs expert instruction on learning simulated surgical skills among medical students: A randomized clinical trial. JAMA Network Open, 5(2), e2149008. https://doi.org/10.1001/jamanetworkopen.2021.49008. PMID: 35191972.

Giglio, B., Albeloushi, A., Alhaj, A. K., Alhantoobi, M., Saeedi, R., Davidovic, V., et al. (2025). Artificial intelligence-augmented human instruction and surgical simulation performance: A randomized clinical trial. JAMA Surgery, 160(9), 993-1003. https://doi.org/10.1001/jamasurg.2025.2564. PMID: 40768205.

Hanna, F., Gad, M. M., Helmy, M. B. F., Eisa, M. M., Omran, M. W. Z., Youssef, M., et al. (2026). Artificial intelligence-augmented tutoring vs expert instruction on learning simulated general surgical skills: A systematic review and meta-analysis. BMC Medical Education, 26, Article 952. https://doi.org/10.1186/s12909-026-09606-9. PMID: 42271404.

Hashimoto, D. A., Marwaha, J. S., Lee, S. A., Schwaitzberg, S., & Duffourc, M. (2026). Risk and liability in the deployment of AI systems for surgery: A SAGES white paper. Surgical Endoscopy, 40(7), 5468-5479. https://doi.org/10.1007/s00464-026-12881-8. PMID: 42260175.

Liu, X., Rivera, S. C., Moher, D., Calvert, M. J., Denniston, A. K., & the SPIRIT-AI and CONSORT-AI Working Group. (2020). Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: The CONSORT-AI extension. BMJ, 370, m3164. https://doi.org/10.1136/bmj.m3164. PMID: 32909959.

Pucher, P. H., Brunt, L. M., Davies, N., Linsk, A., Munshi, A., Rodriguez, H. A., et al. (2018). Outcome trends and safety measures after 30 years of laparoscopic cholecystectomy: A systematic review and pooled data analysis. Surgical Endoscopy, 32(5), 2175-2183. https://doi.org/10.1007/s00464-017-5974-2. PMID: 29556977.

Rivera, S. C., Liu, X., Chan, A.-W., Denniston, A. K., Calvert, M. J., & the SPIRIT-AI and CONSORT-AI Working Group. (2020). Guidelines for clinical trial protocols for interventions involving artificial intelligence: The SPIRIT-AI extension. BMJ, 370, m3210. https://doi.org/10.1136/bmj.m3210. PMID: 32907797.

Stulberg, J. J., Huang, R., Kreutzer, L., Ban, K., Champagne, B. J., Steele, S. R., Johnson, J. K., Holl, J. L., Greenberg, C. C., & Bilimoria, K. Y. (2020). Association between surgeon technical skills and patient outcomes. JAMA Surgery, 155(10), 960-968. https://doi.org/10.1001/jamasurg.2020.3007. PMID: 32838425.

U.S. Food and Drug Administration. (2025). Artificial intelligence in software as a medical device. Retrieved July 23, 2026, from https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-software-medical-device.

U.S. Food and Drug Administration. (2025). Marketing submission recommendations for a predetermined change control plan for artificial intelligence-enabled device software functions: Guidance for industry and Food and Drug Administration staff. Retrieved July 23, 2026, from https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketing-submission-recommendations-predetermined-change-control-plan-artificial-intelligence.

U.S. Food and Drug Administration. (2026). Artificial intelligence-enabled medical devices. Retrieved July 23, 2026, from https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-enabled-medical-devices.

U.S. Food and Drug Administration. (2026). About Manufacturer and User Facility Device Experience (MAUDE) Database. Retrieved July 23, 2026, from https://www.fda.gov/medical-devices/mandatory-reporting-requirements-manufacturers-importers-and-device-user-facilities/about-manufacturer-and-user-facility-device-experience-maude-database.

Vasey, B., Nagendran, M., Campbell, B., Clifton, D. A., Collins, G. S., Denaxas, S., et al. (2022). Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. BMJ, 377, e070904. https://doi.org/10.1136/bmj-2022-070904. PMID: 35584845.

Wu, S., Tang, M., Liu, J., Qin, D., Wang, Y., Zhai, S., et al. (2024). Impact of an AI-based laparoscopic cholecystectomy coaching program on surgical performance: A randomized controlled trial. International Journal of Surgery, 110(12), 7816-7823. https://doi.org/10.1097/JS9.0000000000001798. PMID: 38896869.

Yilmaz, R., Bakhaidar, M., Alsayegh, A., Hamdan, N. A., Fazlollahi, A. M., Tee, T., et al. (2024). Real-time multifaceted artificial intelligence vs in-person instruction in teaching surgical technical skills: A randomized controlled trial. Scientific Reports, 14, Article 15130. https://doi.org/10.1038/s41598-024-65716-8. PMID: 38956112.

Yilmaz, R., Alsayegh, A., Bakhaidar, M., Fazlollahi, A. M., Hamdan, N. A., Tee, T., et al. (2025). Combining real-time AI and in-person expert instruction in simulated surgical skills training: Randomized crossover trial. npj Artificial Intelligence, 1, Article 36. https://doi.org/10.1038/s44387-025-00032-8.