Artificial intelligence in urological education: a systematic review

Abstract

Background

Artificial intelligence (AI) is currently widely applied across various fields, especially in the field of education where it holds significant implications. This systematic review aims to explore the specific role of AI in urological education.

Methods

Following the PRISMA 2020 guidelines, we searched the PubMed, Scopus, and Web of Science databases for articles related to AI in urological education published up to February 10, 2026. The quality of included studies was assessed using the Medical Education Research Study Quality Instrument (MERSQI), and educational outcomes were classified based on the modified Kirkpatrick framework.

Results

A total of 15 eligible study records were finally included. The use of AI in urology education can be categorized into two groups: interventions that are used directly in the educational process, and enabling technologies that provide support for education. It was showed that combining AI in urology education had a positive impact on both learners and teachers. For learners such as medical students, urology residents, or interns, the AI-based hybrid teaching model increased learning efficiency and improved students’ theoretical exam scores. Meanwhile, the combination of high-fidelity virtual reality simulation training and intelligent AI feedback showed potential in facilitating the acquisition of clinical skills, in which simulation scores were positively correlated with surgical potential. On the other hand, as the supporting or enabling technology, AI assisted teachers in designing teaching materials and developing assessment programs. In terms of evaluating learners’ performance, AI provided fair and objective evaluation results. In addition, machine learning and computer vision algorithms incorporating AI could automatically recognize surgical steps with high precision.

Conclusion

The application of AI possesses the capacity to drive transformative changes in urological education. AI can serve as an effective tool to improve students’ learning efficiency, theoretical knowledge level and clinical skills through assisted teaching. Moreover, as the enabling technology, the application of AI may be potential in providing support during the teaching process. But AI should be utilized in a responsible and ethical manner.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12909-026-09687-6.

Introduction

Artificial intelligence (AI) is currently receiving an increasing amount of attention and is widely applied across numerous fields [1]. Simply put, AI refers to the technology that enables computer systems to simulate human intelligence processes. It encompasses four core branches that are interconnected yet each with its own distinct focus: machine learning (ML), deep learning (DL), computer vision (CV), and natural language processing (NLP) [2]. In this study, we define AI as computer systems or algorithms capable of performing tasks that typically require human cognitive functions, such as learning, recognition, and decision-making. The technologies discussed in this paper primarily include generative large language models (LLMs) for interactive teaching and assessment, as well as machine learning (ML) and computer vision (CV) algorithms for surgical video analysis and skill assessment. Although this review includes studies on virtual reality (VR) simulations, we have excluded traditional VR systems that merely provide digital immersive environments; the VR simulators included in this analysis are all AI-assisted.

Currently, AI has been applied in various domains of surgical education. For instance, in the field of neurosurgery, Winkler-Schwartz et al. utilized machine learning algorithms to conduct an in-depth analysis of trainees’ micromechanical and instrument kinematics data on a virtual reality (VR) brain tumor resection simulator [3], successfully establishing an automated objective skill grading model with an accuracy of up to 90%. In the context of postoperative debriefing, recent advancements in CV have enabled the automated recognition of surgical phases and tracking of instrument kinematics from video recordings. Such intelligent systems can now automatically segment core procedural steps, allowing medical students and surgical trainees to efficiently review complex anatomical maneuvers without the need for manual video editing [4].

However, in the field of urological education, research exploring the application and efficacy of AI in teaching remains extremely scarce. The traditional lecture-based learning (LBL) method in urology is teacher-centered. It adopts a “spoon-feeding” teaching method, wherein students passively listen to theoretical knowledge. This approach lacks teacher-student interaction and neglects students’ self-directed learning abilities [5]. Consequently, it possesses obvious limitations: it overemphasizes theory over practice, restricts students’ thinking, and makes it difficult for them to translate textbook knowledge into genuine clinical practical skills [6]. An increasing number of novel teaching methods are being attempted, such as problem-based learning (PBL) and case-based learning (CBL) [7]. These methods have been proven to enhance students’ interest, self-learning ability, clinical analytical skills, comprehension, and memory, while simultaneously making the classroom more dynamic [8]. Currently, preliminary research indicates that AI is beneficial to teaching, but its specific role in urological education is not yet fully elucidated [9].

Therefore, this paper aims to systematically review and evaluate the application and effects of AI in urological teaching. Within the framework of competency-based medical education (CBME), objective and validated assessment tools are considered an essential component of educational interventions. Therefore, in order to address the conceptual heterogeneity of AI applications, this review explicitly categorizes the included articles into two distinct areas: direct educational interventions applied to learners, and enabling technologies focusing on technology validation. This will lay the foundation for the future application and promotion of AI in the field of urological education.

Methods

The search strategy and selection criteria

The implementation and reporting of this systematic review were conducted entirely following Preferred Reporting Items on Systematic Reviews and Meta-Analyses (PRISMA) 2020 recommendations [10]. The PubMed, Scopus, and Web of Science databases were searched comprehensively from January 1, 2016, to February 10, 2026. No automatic database filters (such as language or article type filters) were used during the search. Scientifically, the search strategy incorporated controlled vocabularies (Medical Subject Headings, MeSH) and free-text keywords. The main search concepts tended to be very close to the notions of AI, Machine Learning, Large Language Models, Virtual Reality, Urological Education and the related synonyms. All the electronic search queries with every combination of Boolean operators and particular MeSH terms are listed in Supplementary Table 1.

Prior to the literature screening, our research team had internally developed a standardized review protocol; however, this protocol was not prospectively registered in a public database. This decision was made to avoid lengthy registration wait times, given the rapid pace of iteration in AI technology. Nevertheless, all procedures strictly adhered to the PRISMA guidelines to ensure transparency in our research methodology.

Eligibility criteria

To ensure that literature which is strongly aligned to the goals of this work is filtered effectively, inclusion and exclusion criteria were set prior to the screening process. Inclusion criteria stipulated that articles must simultaneously include all four elements below: (1) Study population: the population encompasses both ends of the medical education spectrum. This includes trainees (medical students, urology interns, trained residents), fellows and educators (attending physicians, clinical faculty, surgical instructors), and human evaluators involved in skill assessment. (2) Interventions: use of AI-based large language models, machine learning algorithms, computer vision systems or intelligent VR simulators (Eligible articles that include virtual reality must integrate AI components or algorithms. These AI components primarily involve computer vision algorithms for automated extraction and analysis of kinematic data and for real-time surgical task recognition.) These studies fall into two main categories: educational interventions in which AI is involved in the “teaching and learning” or “feedback” aspects of teaching. The other is AI enabling technologies that test the intelligence of AI and the accuracy of the data measured to provide a technological foundation for future teaching and learning. (3) Comparators: For direct educational interventions, traditional clinical teaching or no intervention was used as a control. For the AI technology validation and feasibility study, traditional manual scoring or assessment by a senior human expert was defined as the standard control group. (4) Outcomes: quantifying educational assessment measures (e.g. objective exam scores, OSATS/GOALS ratings, operation time or validated satisfaction scores). Non-English literature, review papers, case reports, and multidisciplinary medical journals that exclusively addressed AI-assisted clinical condition diagnosis but did not provide a clear pedagogy aim were presented in the exclusion criteria, as well as traditional virtual reality (VR) or computer-aided simulation platforms that lack clear AI integration.

Selection process

Each reference identified during the search was imported into the EndNote citation management software, and duplicates were automatically removed by the software. In the first stage of screening, two reviewers independently assessed the titles and abstracts of all identified records based on predefined eligibility criteria. In the second stage, potentially relevant articles were retrieved in full text and re-evaluated independently by the same two reviewers. Any disagreements or inconsistencies arising at either screening stage were resolved through detailed discussion to reach consensus, or by consulting a third senior and experienced researcher for final arbitration. Following the screening, the two reviewers independently and in duplicate extracted data using Microsoft Excel. To best adhere to the PRISMA recommendations, we systematically extracted the following key data elements: (1) Initial study features (e.g. lead author, year of publication, country of origin and detailed experimental design, e.g. randomized controlled or prospective educational studies); (2) Characteristics of the participants (e.g. overall sample size, clinical training level, and previous surgical or AI experience); (3) Intervention and control information (including the exact type of AI technology, particular applications in cognitive/ psychomotor or assessment domains, and the pedagogical structure of the control group); (4) Key educational results (obtaining their respective specific Kirkpatrick evaluation, main outcome measure devices like the objective time metric on simulator, and overall statistics like mean differences and p-values).

Synthesis of results

Considering the extreme heterogeneity of the literature in terms of methods of design and clinical application settings- as is especially evident in the variety of AI technology types (as in the use of both LLMs and VR alongside computer vision) and large variation in target population, highly inconsistent measures of educational outcomes - the forced aggregation of such data into a meta-analysis of quantitative data was considered as not methodologically sound. However, the review used a narrative synthesis approach that was informed by reflexive thematic analysis [11]. We were able to present the shared value of AI technologies in the various aspects of education in urology through systematic synthesis of the study findings in a logical flow. In order to avoid overestimating the reliability of the various pieces of evidence, we have divided our analysis into two parts: direct educational interventions, and feasibility validation of the relevant technologies. Reporting practices as suggested by PRISMA and the latest recommendation on reporting of systematic reviews of medical education were used. Duplicate screening and extraction were done to reduce errors. Structured thematic synthesis was used to synthesize findings. Two reviewers (G.C.L. and L.X.) independently applied the Medical Education Research Study Quality Instrument (MERSQI) in the quality methodological control of quantitative studies [12]. Any discrepancies in scoring were resolved through discussion; if no consensus was reached, a senior reviewer (P.Y.) made the final decision. The MERSQI is a multidimensional scale that examines the literature in six main areas: study designs, sampling methods, types of data (subjective or objective), validity of assessment measures, and complexity of data analysis, and outcome measures, and it has a maximum score of 18 [13]. MERSQI is essentially an educational research quality indicator rather than a formal risk of bias (RoB) assessment tool. However, due to the methodological heterogeneity of the included studies—ranging from randomized controlled trials to AI technology validation studies— Applying a single standard RoB framework (such as Cochrane RoB 2 or ROBINS-I) uniformly across all studies is methodologically inappropriate. Therefore, MERSQI provides a framework capable of standardizing the quality of these studies, which is used in this article to encompass these diverse study designs. Also, in order to stratify the strength of evidence of these educational interventions in a more orderly and standardized manner, we adopted an innovative approach to apply a modified Kirkpatrick structure [14]. This theoretical framework strictly divides the effects of pedagogical results into four levels as follows: Level I (reaction and satisfaction of trainees); Level II (performance of learning which is further divided into level IIa, changes in attitude and level IIb, acquisition of knowledge/skills); Level III (transfer of skills to the clinical setting and behavior change); Level IV (results, i.e., improvement in real post-operative health care of patients or health indicators) [15].

Results

After searching the PubMed, Scopus and Web of Science databases systematically, we have found a cumulative number of 1,199 articles published as of February 10, 2026. Automated removal of duplicates left us with 1,171 unique citations. Eventually, there were 15 quantitative studies that met all the criteria and incorporated into the last synthesis (Fig. 1).

Fig. 1.

PRISMA flow diagram

Study populations included medical students, urological residents and urologist trainees. Major findings on details and major outcomes of the studies involved were presented in Table 1. The methodological quality of the 15 included studies was evaluated using the MERSQI, with a mean score of 11.33 (range: 9.5–13.0). Detailed scores for each domain are presented in Table 2.

Table 2.

MERSQI score of 15 included studies

Study (Author, Year)	Study Design	Sampling	Type of Data	Validity of Instrument	Data Analysis	Outcomes	Total MERSQI score
Digiacomo (2025) [17]	3	1.5	3	1	3	1.5	13
Hui (2025) [18]	3	1.5	3	2	2	1.5	13
Cowan (2021) [29]	2	1	3	2	2	1.5	11.5
Aloosh (2016) [30]	2	1	3	3	2	1.5	12.5
Russo (2025) [20]	1.5	1	3	2	2	3	12.5
Ebina (2022) [28]	2	1.5	3	2	2	1.5	12
Khanna (2024) [24]	1.5	1.5	3	1	2	1.5	10.5
Sato (2025) [21]	2	1	3	2	2	1.5	11.5
Pears (2024) [25]	2	1	3	1	2	1.5	10.5
Kim (2025) [19]	1.5	1	3	2	2	1.5	11
Başaranoğlu (2025) [16]	1.5	1	3	2	2	1.5	11
Hetz (2024) [23]	1	1	3	2	2	1.5	10.5
Touma (2024) [26]	1	1	3	2	2	1.5	10.5
Yudovich (2024) [27]	1	1	3	2	2	1.5	10.5
Henning (2024) [22]	1.5	1	2	1	2	2	9.5

MERSQI Medical Education Research Study Quality Instrument

Table 1.

Summary of the studies included and characteristics of AI models in urology education

Author/ Year	Source	Population	Study design	Intervention Details	Controls	Outcomes	AI Modality	Kirkpatrick Level	Evidence for Kirkpatrick Level	Findings
Başaranoğlu et al. (2025) [16]	BMC Med Educ	Medical students	Non-randomized Controlled Trial	ChatGPT and Gemini were utilized to automatically generate 300 multiple-choice questions for evaluating students’ urological competency.	Same cohort performance compared across specific intervals: baseline, mid-rotation, and final assessment.	Final vs. baseline mean score: 78.4% vs. 45.2% (P < 0.001).	Generative AI (ChatGPT & Gemini)	Level 2b: Acquisition of knowledge/skills	Demonstrated acquisition of skills and knowledge by showing that a hybrid AI teaching model significantly improved trainees’ objective performance.	Improving learning efficiency & Reducing teacher workload. Improving learning efficiency & Reducing teacher workload. AI accurately generates test questions and significantly raises students’ scores. It significantly reduces teachers’ manual preparation time and effectively boosts students’ knowledge acquisition.
Digiacomo et al. (2025) [17]	World J Urol	Medical students	Randomized Controlled Trial	Students utilized ChatGPT to independently study assigned urological topics by actively generating queries and interacting with the AI.	Students were randomly divided into three study classes with different learning methods: pure ChatGPT, pure Lecture, and ChatGPT + Lecture.	ChatGPT+Lecture vs. pure ChatGPT median score: 12 vs. 10 (P = 0.007). Multivariate regression estimate: 2.6 (P = 0.002).	Generative AI (ChatGPT)	Level 2b: Acquisition of knowledge/skills	Confirmed knowledge acquisition, as the AI-assisted PBL method enabled students to achieve objectively higher theoretical scores at the end of the course.	Enhancing theoretical knowledge. Combining ChatGPT with traditional lectures yields the highest scores, and it shows AI consequentially acts as a supplementary tool to deepen theoretical understanding for students.
Hui et al. (2025) [18]	BMC Med Educ	Medical students	Randomized Controlled Trial	ChatGPT was integrated into Problem-Based Learning sessions as an AI assistant to help students explore and solve clinical cases.	The experimental group received PBL and ChatGPT-assisted teaching methods, while the control group received traditional teaching methods.	PBL-ChatGPT vs. traditional theoretical score: 93.90 vs. 90.33 (P < 0.01). Overall clinical competence (Mini-CEX): 6.14 vs. 5.14 (p < 0.001).	Generative AI (ChatGPT)	Level 2b: Acquisition of knowledge/skills	Evidenced knowledge acquisition by demonstrating that students in the AI-assisted PBL cohort attained superior theoretical scores compared to traditional groups.	Enhancing theoretical knowledge & Enhancing clinical skills. ChatGPT-assisted learning improves both the performance of written exams and clinical evaluation (Mini-CEX) scores, and it demonstrates a comprehensive enhancement on students’ capabilities.
Kim et al. (2025) [19]	Can Urol Assoc J	Urology residents	Cross-sectional Study	A GPT-4 model, enhanced with retrieval-augmented generation (RAG), was instructed to formulate 60 multiple-choice questions mimicking board exams.	The proportion of questions answered correctly was evaluated by year of training (PGY1-4) to test discriminatory value.	Median score: 37/60. Significant correlation between scores and training year (Pearson r = -0.759, P = 0.004).	Generative AI (GPT-4)	Level 1: Reaction	Assessed learner reaction, specifically evaluating the perceived usability, user engagement, and acceptability of AI-integrated educational topics.	Optimizing assessment methods & Reducing teacher workload. AI generates questions that accurately distinguished residents by seniority, and it provides a reliable, automated evaluation tool that contributes to reduce teachers’ actual working burden.
Russo et al. (2025) [20]	Front Urol	Urology residents	Non-randomized Controlled Trial	Residents utilized an AI-enabled automated semen quality analyzer (LensHooke X1 PRO) featuring AI algorithms to assess patient semen parameters.	Each patient underwent semen analysis the day before and 3 months after surgery (self-controlled baseline vs. follow-up).	Improved total sperm concentration MD = 13.30 (P = 0.03). Total sperm count MD = 27.05 (P = 0.03).	AI-assisted Device (CASA)	Level 2b: Acquisition of knowledge/skills	Indicated skill acquisition by demonstrating that the intervention equipped trainees with the ability to effectively manipulate advanced clinical tools.	Enhancing clinical skills. AI provides rapid, standardized clinical readouts, and it familiarizes residents with advanced diagnostic tools, ultimately translating to better clinical decision-making.
Sato et al. (2025) [21]	Langenbecks Arch Surg	Urologist trainees	Cohort Study	An AI-based computer vision model automatically extracted instrument status from surgical videos to measure dissection and exposure times.	Novices and experts were compared by measuring dissection and exposure times automatically derived by the AI model.	Exposure time (novice vs. expert): 195.7s vs. 89.7s (P < 0.001); D/E ratio: 0.174 vs. 0.315 (P = 0.003).	Computer Vision	Level 1: Reaction	Measured learner reaction by investigating how AI-driven surgical interface recognition positively influenced novices’ subjective training experiences and self-perception.	Optimizing assessment methods. AI can extract specific surgical metrics (exposure time) to differentiate skill levels, and it offers an objective, data-driven method for evaluating trainees’ surgical proficiency.
Henning et al. (2024) [22]	J Surg Educ	Urologist trainees	Non-randomized Controlled Trial	Implementation of an AI-driven surgical intelligence platform (Theator) that automatically titled, indexed, and annotated surgical videos in real-time.	Platform usage and engagement by level of training were compared prior to and following department in-service training.	Active junior resident users increased from 3 to 10; monthly comments per user increased from 2.3 to 7.8.	AI Video Platform (Theator)	Level 2b: Acquisition of knowledge/skills	Facilitated skill acquisition by utilizing AI-labeled videos to guide trainee self-reflection and instructor-led improvements in specific surgical maneuvers.	Reducing teacher workload & Improving learning efficiency. AI platform automates the tedious task of video annotation, and it saves faculty time and makes continuous surgical coaching more accessible to trainees.
Hetz et al. (2024) [23]	ESMO Real World Data Digit Oncol	Urologist trainees	Cross-sectional Study	UroBot-4o (a customized RAG-enhanced GPT model) was tasked with answering 200 European Board of Urology in-service assessment questions.	Benchmarked against the zero-shot performance of standard LLMs and historical average performance of human urologists.	Average correct accuracy UroBot-4o vs. GPT-4o: 88.4% vs. 77.6% (P < 0.001).	Generative AI (GPT-4o)	Level 2a: Modification of attitudes/perceptions	Prompted a modification of attitudes toward AI’s clinical utility by proving its average accuracy on board exams surpassed that of human urologists.	Enhancing theoretical knowledge. AI outperforms trainees in answering specialty questions, and it proves to be a highly capable on-demand knowledge retrieval tool for independent study.
Khanna et al. (2024) [24]	J Urol	Urologist trainees	Cross-sectional Study	A novel AI computer-vision algorithm was applied to automatically identify and annotate specific surgical steps in robotic-assisted radical prostatectomy videos.	Accuracy of automated video annotation was determined by comparing it to manual human expert annotations as the reference standard.	Overall full-length video step recognition accuracy was 92.8%.	Computer Vision	Level 2a: Modification of attitudes/perceptions	Drove a modification of perceptions regarding AI’s reliability by validating its capacity to autonomously and accurately recognize specific surgical steps in RARP videos.	Optimizing assessment methods. AI algorithm accurately identifies surgical steps, and it paves the way for the automated tracking and objective assessment of trainees’ operative progression.
Pears et al. (2024) [25]	J Healthc Inform Res	Urologist trainees	Randomized Controlled Trial	ChatGPT-4 was deployed to provide real-time, personalized feedback on trainees’ Non-Technical Skills based on simulated surgical scenarios.	Trainees interacted with two human consultant urologists and the AI model simultaneously, allowing for double-blind comparisons of the feedback.	AI feedback was significantly better in response length (P = 0.003) and promoting critical thinking (P = 0.02).	Generative AI (GPT-4)	Level 2b: Acquisition of knowledge/skills	Demonstrated knowledge and skill acquisition through simulated clinical encounters where trainees received and applied targeted AI feedback.	Optimizing assessment methods & Reducing teacher workload. AI feedback on non-technical skills matches human experts, and it can automate routine evaluations on trainees’ performance, freeing up teachers for more complex mentoring.
Touma et al. (2024) [26]	Can Urol Assoc J	Urology residents	Cross-sectional Study	ChatGPT-4 was administered a mock Canadian urology board examination to assess its medical knowledge acquisition.	Compared the performance of the AI model to a cohort of 29 Canadian urology PGY-5 human test takers.	Oncology sub-score (ChatGPT-4 vs. human average): 35% vs. 71% (P < 0.05).	Generative AI (ChatGPT-4)	Level 1: Reaction	Established baseline metrics for tool acceptability (a prerequisite for favorable learner reaction) by comparing current AI exam performance against real graduating residents.	Enhancing theoretical knowledge. AI currently contributes to improve residents’ learning ability to obtain the specialized knowledge.
Yudovich et al. (2024) [27]	J Educ Eval Health Prof	Urologist residents	Cross-sectional Study	GPT-3.5 and GPT-4 were directly evaluated on their accuracy in answering 700 multiple-choice urology board exam-style items.	The accuracy of GPT-3.5 and GPT-4 was laterally compared across item complexity types (recall, interpretation, problem-solving).	Total accuracy (GPT-4 vs. GPT-3.5): 44.4% vs. 30.9% (P < 0.001); Interpretation accuracy: 45.6% vs. 31.5% (P < 0.001).	Generative AI (GPT-3.5 & GPT-4)	Level 1: Reaction	Evaluated the tool’s baseline educational validity (informing user reaction and trust) by measuring the AI’s knowledge against the standard passing grade of a human doctor.	Optimizing assessment methods. AI is beneficial to improve the capability and accuracy of the standardized urology knowledge assessment systems which are used by residents to prepare for examinations.
Ebina et al. (2022) [28]	Langenbecks Arch Surg	Urologist trainees	Cross-sectional Study	Machine learning regression algorithms were applied to motion capture data to automatically calculate and estimate trainees’ surgical GOALS scores.	Correlations between the AI-derived metrics and the manual GOALS scores evaluated by two human experts.	High algorithmic accuracy confirmed via MAE metrics: SVR MAE = 2.2352; PCA-SVR MAE = 1.2714.	Machine Learning (Mocap)	Level 2b: Acquisition of knowledge/skills	Validated the reliable measurement of skill acquisition by showing that AI-predicted GOALS scores highly concorded with manual evaluations by senior experts.	Optimizing assessment methods. Machine learning effectively estimates standard scores and outcomes based on physical motion data, and it provides an automated, objective tool for evaluation on trainees’ clinical skill.
Cowan et al. (2021) [29]	J Endourol	Urologist trainees	Randomized Controlled Trial	Trainees completed surgical tasks in a Virtual Reality (VR) simulator, which utilized built-in algorithms to generate automated performance metrics.	Performance during analogous tasks was compared across two environments: virtual reality (VR) and dry laboratory (DL).	DL completion time (experts vs. trainees): 10 min vs. 15 min (p = 0.005). VR load correlated with DL cognitive load (P < 0.001).	VR Simulator	Level 2b: Acquisition of knowledge/skills	Supported skill acquisition by allowing subjects to actively train and refine surgical techniques within an AI-integrated VR environment.	Enhancing clinical skills & Optimizing assessment methods. VR metrics correlates with actual performance and distinguished expertise, and it can serve as both a safe training ground for reinforcing trainees’ skills and an objective tool for the trainees’ evaluation.
Aloosh et al. (2016) [30]	J Endourol	Urology residents	Non-randomized Controlled Trial	Trainees practiced ureteroscopic stone extraction using the UroMentor VR simulator, which provided objective algorithm-based performance assessments.	URS-GRS scores obtained by participants on the simulator were correlated with scores assessed in the real operating theatre.	Significant positive correlation between simulator and operating theatre scores: r = 0.9 (P = 0.03).	VR Simulator (UroMentor)	Level 2b: Acquisition of knowledge/skills	A significant positive correlation was established between simulator proficiency and operating room performance, thus demonstrating skill acquisition and validating the effectiveness of the simulation tool.	Enhancing clinical skills & Optimizing assessment methods.The simulator score is able to predict residents’ real surgical performance, and it validates AI-driven VR as a reliable tool for optimizing as well as assessing residents’ clinical skills.

AI Artificial intelligence, CASA Computer-assisted semen analysis, D/E ratio Dissection-to-exposure ratio, DL Dry laboratory, GPT Generative pre-trained transformer, LLMs Large language models, MAE Mean absolute error, MCQs Multiple choice questions, MD Mean difference, Mini-CEX Mini-clinical evaluation exercise, NTS Non-technical skills, PCA Principal component analysis, PGY Post-graduate year, RAG Retrieval-augmented generation, SVR Support vector regression, VR Virtual reality

In the field of urology education, the use of AI and its impacts can be summarized in five main areas: The enhancement of student learning efficiency through AI-assisted teaching, the enhancement of medical trainees’ theoretical knowledge and feedback behavior, the feasibility of enhancing clinical skills, the performance validation of AI models in professional examinations, and the optimization of competency assessment and the potential to alleviate faculty workload. To ensure conceptual clarity, these findings were synthesized into two overarching domains: educational interventions and learner outcomes which synthesize the direct impacts on trainees and encompassing concepts as well as enabling technologies and performance validation which evaluate the basic functionality and automated assessment capabilities of AI systems, and encompassing concepts.

Educational interventions and learner outcomes.

1. The enhancement of student learning efficiency through. AI-assisted teaching. A randomized controlled trial by Digiacomo et al. demonstrated that a blended teaching model combining ChatGPT with traditional lectures can improve learning outcomes [17]. The median test score of the blended group (12 points) was higher than that of the pure ChatGPT group (10 points, p = 0.007). In addition, Hui et al. found that ChatGPT-assisted PBL (Problem-Based Learning) teaching enabled students to reach a theoretical score of 93.90 ± 3.65 at the end of the course [18], which was superior to the traditional teaching group’s 90.33 ± 4.08 (p < 0.01).

2. Enhance medical trainees’ theoretical knowledge and feedback behavior. Furthermore, Pears et al. explored the application of GPT-4 in non-technical skills (NTS) training [25], finding that AI can provide accurate, evidence-based, and empathetic feedback, showing advantages in reinforcing information gathering (p = 0.04) and adjusting explanation complexity according to the learner’s level (p = 0.002). Additionally, the study by Başaranoğlu et al. showed that after systemic training, the objective theoretical assessment scores of medical interns steadily and improved from a baseline of 45.2% to a final assessment score of 78.4% [16].

3. The feasibility of enhancing clinical skills. Research by Aloosh et al. showed that trainees can reach proficiency in ureteroscopy and lithotripsy after 7 training sessions on the UroMentor simulator [30], and the evaluation scores on the simulator demonstrated a positive correlation with performance in the real operating room (r = 0.9, p = 0.03).

Foundational enabling technologies and performance validation.

1. Performance validation of AI models in professional examinations. Hetz et al. confirmed that the dedicated language model UroBot-4o [23], optimized through Retrieval-Augmented Generation (RAG) combined with the European Association of Urology (EAU) guidelines, achieved an average accuracy rate of 88.4% when answering urology board exam questions, surpassing the average level of human urologists (68.7%). However, studies by Touma et al. and Yudovich et al. indicated that generative models like ChatGPT-4 still struggle to consistently pass rigorous board-style urology examinations [26, 27], scoring lower than human resident averages.

2. Optimization of competency assessment and the potential to alleviate faculty workload. A study by Başaranoğlu et al. pointed out that traditional test question development typically requires 45–60 min per question [16], whereas using AI-assisted development can shorten this to 15–20 min, reducing the time consumption and cost per validated question by 65%. Similarly, Henning et al. observed that AI video platforms (e.g., Theator) can automate the tedious task of surgical video annotation [22]. Concurrently, Cowan et al. also verified that the VR simulator has a good correlation with the Dry Lab in terms of evaluation metrics [29]. Furthermore, the integration of AI-assisted diagnostic devices, such as the computer-assisted semen analysis (CASA) system, helps trainees familiarize themselves with advanced clinical tools [20]. Sato et al. utilized an AI-based interface recognition model to automatically measure the “dissection” and “exposure” times in Robot-Assisted Radical Prostatectomy (RARP) [21], finding that the exposure time of the expert group was shorter than that of the novice group (89.7 s vs. 195.7 s, p < 0.001). Research by Khanna et al. indicated that the overall accuracy of AI algorithms in automatically recognizing the steps of full-length RARP surgical videos reached up to 92.8% [24]. In addition, Ebina et al. applied machine learning combined with a motion capture system to assess laparoscopic skills [28]. The total GOALS scores predicted by their models (such as SVR and PCA-SVR) were consistent with the manual scoring by senior experts, with a median absolute error of only 1.27 to 2.23 points. Kim et al. [19] validated that GPT-4 could automatically generate high-quality multiple-choice questions that discriminated urological residents according to their training year.

Discussion

The primary objective of this study is to evaluate the impact of introducing AI into medical education on both students and clinical faculty. It should be emphasized that AI in urology teaching can act as a direct educational intervention for the learner to improve the educational effectiveness. Our research results indicate that AI-assisted teaching tools have the potential to enhance students’ learning efficiency, mastery of theoretical knowledge, and acquisition of clinical skills [31]. The improvement in student learning efficiency and theoretical knowledge observed in this study may be theoretically attributed to the personalized features of AI algorithms. Unlike traditional large-class lectures, intelligent tutoring systems are designed to analyze students’ individual performance in real-time, push targeted learning resources, and provide immediate feedback [32]. In principle, AI has the capacity to meet the needs of different learning styles by generating multiple types of learning methods, such as combining charts, videos, and text. Furthermore, theoretical models suggest that the spaced repetition algorithm underlying the system could combat the knowledge forgetting curve, which also aligns with the Ebbinghaus forgetting curve to help students better remember knowledge [33]. If fully realized, these technical features offer the potential to enhance students’ learning efficiency, enabling them to focus their energy on filling specific knowledge blind spots, thereby accelerating the learning process [34].

In terms of learning efficiency and theoretical knowledge, our conclusions align with one previous study [35] and the results of multiple recent systematic reviews. They similarly found that AI-driven adaptive learning platforms can improve medical students’ learning efficiency and knowledge retention rates through personalized feedback. In addition, urology trainers using an AI-assisted virtual reality training system showed improved clinical skills. By providing an immersive, high-fidelity simulated environment, students can safely engage in repeated hands-on practice under the premise of reduced ethical burden and patient risk [36].

In terms of clinical skill training, the positive impact observed in this study corroborates research conclusions regarding AI combined with virtual reality (VR) simulators, such as the recently published article by De Luca et al. [37], which explored how immersive environments hold the potential to promote the standardization of clinical procedures and the formation of muscle memory. However, although this study observed positive effects within a short-term intervention, some existing literature has also proposed findings that differ from our results. This study emphasized the advantages of AI in improving clinical operational skills, but some scholars have raised opposite concerns. For example, a recent survey targeting clinicians highlighted that if there is an over-reliance on AI’s diagnostic recommendations and automated guidance during early training, it may weaken young doctors’ independent clinical reasoning abilities and basic physical examination skills [38]. Furthermore, as pointed out by scholars such as Sriram, AI simulators excel in enhancing procedural operations, but they cannot completely replicate real doctor-patient communication, empathy, and the adaptability to face complex clinical environments. These are blind spots that AI-assisted teaching currently cannot reach [39].

Besides direct educational impact and benefit (e.g., improvement in knowledge, skills, or clinical competence), AI may be also potential in utilized as an enabling technology that provides technical support and infrastructure for the educational process. Regarding technical capabilities in surgical environments, the AI system possesses the capability to identify surgical steps and provide automated micro-skill feedback. In future applications, this technology has the potential to play a crucial role in perfecting operational standards before trainees truly enter the operating room. Traditional clinical assessments are inherently subjective, time-consuming, and prone to scoring inconsistencies. By utilizing AI to generate theoretical test questions and relying on machine learning models for objective clinical skill evaluation, the reliance on continuous, direct observation by teachers could be minimized to the greatest extent [40].

In addition, in terms of optimizing assessment and reducing teacher burden, our research results resonate with the latest studies exploring AI-assisted scoring for Objective Structured Clinical Examinations (OSCE), such as the latest research in JMIR by Yokose et al. [41]. These validation studies confirm that utilizing Natural Language Processing (NLP) and machine learning algorithms for real-time objective scoring shows promise for enhancing the standardization degree of assessment and potentially cutting the administrative and grading burden of clinical educators. At the same time, the future implementation of automated, AI-driven assessment methods offers a pathway to optimize the evaluation process, thereby holding the potential to reduce the administrative and routine teaching workload of clinical teachers [42]. From the perspective of educators, the potential reduction in workload is a projected result of the optimization of assessment methods. This envisioned shift would enable clinical faculty to reallocate their precious time to higher-order clinical mentoring and complex case discussions, rather than being bogged down in administrative work such as routine grading.

Furthermore, regarding technical accuracy, it is also important to note that although some studies, such as Hetz et al. have reported “superhuman” performance in professional exams, a recent study by Katsimperis et al. showed that the accuracy of a large-scale language model on urology board-style questions are currently limited to 63% to 77% accuracy. These validation metrics demonstrate that while large language models may be useful as a quick reference tool for doctors in training, their error rates are still too high to replace traditional learning methods or well-validated educational resources, and they are still a long way from being fully reliable [43]. And their study also mentions that the problem of AI ‘illusion’ also poses a serious threat to academic integrity. AI will falsify citations that seem plausible but don’t actually exist. While AI can significantly improve writing efficiency, its tendency to invent citations poses a threat to research rigor. For trainees, if used uncritically, it can undermine trust in scientific work and damage developing academic reputations. Thus, AI is an excellent assistant, but it is not yet the primary source of clinical evidence. Rigorous manual verification remains the most preferred option to ensure the authenticity and reliability of all cited literature in medical education.

The core findings of this study largely remain consistent with recent literature exploring the application of AI in medical education. The positive results of this study confirm the potential of AI as a medical education tool, consistent with the optimistic expectations of mainstream literature [44]. However, comparing the prudent findings of other scholars also reminds us: AI improves learning efficiency and holds the potential to enhance the degree of standardization, but it cannot completely replace the teaching of human experts and real doctor-patient interactions [45].

This study also provides new supplements to several core theories in the field of medical education: First, it promotes the theoretical implementation of “Competency-Based Medical Education (CBME)”: For a long time, the medical education community has continuously called for a transition from being “time-oriented” to “competency-oriented,” but the theoretical bottleneck lies in the lack of objective, continuous competency assessment indicators [46]. This study demonstrates the technical capacity of AI and machine learning models to quantify clinical skills and provide automated assessments. This theoretically indicates how AI could fill the gap in the CBME assessment system, suggesting that “data-driven competency evaluation” is technically feasible for future applications.

Despite achieving the above positive results, this study still has certain limitations. First, the sample size is relatively limited, and the research subjects are confined to a specific trainee group, which may limit the universal applicability of the conclusions of this study in broader medical disciplines. Second, although AI tools demonstrated high technical accuracy in validation studies, and showed positive educational effects on short-term skill acquisition, this study did not evaluate students’ long-term knowledge retention rate, nor did it track whether these skills obtained on the simulator can be fully translated into patient benefits in real clinical environments. Third, there may be uncontrolled potential confounding variables that could influence their interactions with AI platforms. Finally, a more fundamental issue lies in the heterogeneity of the evidence and the fact that its applicability to the field of education remains unclear. Some of the included studies evaluated the performance of AI rather than that of learners. Therefore, while these studies demonstrate the technical feasibility of AI tools, the lack of direct learner data prevents them from fully substantiating the long-term impact of these tools on trainees’ actual clinical behavior. It is also important to note that the MERSQI is designed to assess the overall methodological rigor of the included studies, rather than to formally quantify the risk of bias. This limitation suggests that, while the included studies are of acceptable methodological quality, the underlying evidence regarding specific experimental biases should be interpreted with caution.

Furthermore, while the future of AI in urological education looks bright, greater attention must be paid to ethical considerations. Since AI technology is still in its early stages of development, we need to devote more effort to ensuring that the technology is safer and more robust. Regarding ethics, the first point of concern is algorithmic bias. If the datasets used to train AI-driven surgical simulators or automated assessment tools lack diversity or are insufficient in volume, the resulting models may produce biased assessment results or reinforce existing clinical biases. Second, data security and privacy must be strengthened, particularly when AI platforms handle sensitive information such as intraoperative videos of urological surgeries, patient personal information, or trainees’ surgical performance metrics. Strict data encryption and access controls must be ensured, as these are crucial for protecting intellectual property and personal data security. Third, the “black box” nature of many complex AI models leads to a lack of transparency in the results produced, which may hinder learning efficiency. For AI-driven feedback to have educational value, the feedback must be interpretable and reproducible. Fourth, learner surveillance is a double-edged sword: while it allows for more detailed recording and evaluation of trainees’ procedures, it also places psychological pressure on those being observed and restricts their autonomy. Fifth, accountability for AI-generated educational content must be addressed as soon as possible. Questions such as who should bear responsibility when AI provides incorrect information or guidance to urology trainees, leading to adverse consequences, and who should oversee the accuracy of the generated content, must be clarified in the near future. Addressing these ethical issues is crucial for building trust between educators and learners, as well as for ensuring the feasibility and sustainability of AI-assisted urology training.

In the future, it is recommended to conduct large-sample, multi-center longitudinal cohort studies to further verify whether AI-trained students can more effectively learn theoretical knowledge and improve clinical skills, and whether teachers (clinical instructors and professors teaching theory) have reduced their workload and improved teaching efficiency. At the same time, researchers should explore the effectiveness of AI assessment tools’ evaluation capabilities at different stages of standardized residency training. On a practical level, medical school administrators and curriculum planners should no longer merely view AI as an option; they should formulate clear institutional guidelines to evaluate and thoughtfully incorporate validated AI-assisted simulators and automated feedback mechanisms into the core curriculum framework of clinical clerkships and residency training, and ensure fair access rights and comprehensive technical training are provided for all teachers and students.

Conclusion

In summary, this systematic review clarifies the role of AI in urology education. AI can be used as an adjunct tool to improve students’ learning efficiency, theoretical knowledge level and clinical skills through assisted teaching. Moreover, as the enabling technology, the application of AI may be potential in providing support during the teaching process. But AI should be utilized in a responsible and ethical manner, and further studies are needed to determine these educational outcomes.

Abbreviations

AI

Artificial intelligence

CASA

Computer-assisted semen analysis

CBL

Case-based learning

CBME

Competency-based medical education

CV

Computer vision

D/E

Dissection/exposure

DL

Deep learning

EAU

European Association of Urology

EBU

European Board of Urology

GOALS

Global Operative Assessment of Laparoscopic Skills

GPT

Generative pre-trained transformer

LBL

Lecture-based learning

LLMs

Large language models

MCQs

Multiple-choice questions

MERSQI

Medical Education Research Study Quality Instrument

Mini-CEX

Mini-clinical evaluation exercise

ML

Machine learning

NLP

Natural language processing

NTS

Non-technical skills

OSATS

Objective structured assessment of technical skills

OSCE

Objective structured clinical examinations

PBL

Problem-based learning

PCA

Principal component analysis

PRISMA

Preferred Reporting Items for Systematic Reviews and Meta-Analyses

RAG

Retrieval-augmented generation

RARP

Robot-assisted radical prostatectomy

RCTs

Randomized controlled trials

SVR

Support vector regression

VR

Virtual reality

Authors’ contributions

P.Y. and G.M.Y. contributed to the conception and design of this study; G.C.L. , L.X., and J.G. performed the search strategy development and screening; G.C.L. , L.X., and J.G. extracted data and assessed quality; G.C.L. and L.X. performed data analysis and interpretation; G.C.L. , J.G. , and P.Y. drafted and wrote the manuscript; P.Y. and G.M.Y. participated in revision; P.Y. contributed to supervision. All authors reviewed the manuscript and participated in final approvals.

Funding

This study was supported by Fund for the Reform and Development of Higher Education of Hunan Provincial Department of Finance (No.159990012).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This systematic review is a secondary research analysing publicly available data that does not require ethics approval, consent to participate or the collection of new data.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.