Introduction
Scientific advances have significantly influenced the evolution of education and training in recent decades. Emerging technologies such as technology-enhanced learning and simulation-based training have played a crucial role in improving the learning experience of practitioners and have become essential in modern education systems [1].
Traditionally, surgical training has mainly focused on gaining experience through a significant number of surgeries and direct involvement, in which trainees receive less supervision from experienced surgeons as they gain competence and eventually become capable of doing surgeries independently [2]. This model embodies the “see one, do one, teach one” approach [3]. An experienced surgeon first executes a procedure, which the trainee observes. Then, under supervision, the trainee replicates the process. Finally, upon achieving competence, the trainee is expected to instruct others on how to perform it. This approach underscores the importance of direct observation, practical experience, and the ability to transmit information and expertise to future generations of medical practitioners. However, it also raises inquiries regarding the diversity of learning experiences, the consistency of the skills acquired, and the stress that it places on seasoned surgeons and trainees to quickly comprehend and transmit complex procedures involving inherent risks [4,5]. Acquiring and improving skills in the field of medicine are complex processes that last throughout a physician’s career. Since the 1990s, ongoing discussions have focused on enhancing teaching practices [6].
Researchers have developed various simulators and training platforms to address these challenges and the demands of an expanding spectrum of surgical operations [7]. These tools enable trainees to develop expertise in different surgical procedures and provide the benefit of unlimited practice opportunities, customizable difficulty levels, and cost-effective solutions that emulate the difficulties of actual surgery procedures [8,9]. Furthermore, these platforms offer a secure and interactive setting that promotes learning through experimentation, enabling risk-free practice. Nevertheless, there remains considerable potential to improve the effectiveness of these training setups [10,11].
As technological advancements continue, interest in incorporating artificial intelligence (AI) into medical training has also increased [12]. AI, with its capacity to emulate certain aspects of human cognition, has the potential to enhance educational outcomes and transform traditional methods of training and teaching [13]. It enables the creation of the next generation of autonomous systems to execute tasks usually performed by individuals, representing a substantial advancement in computer science. Furthermore, AI algorithms could assist in enhancing conceptual understanding, facilitating virtual practice, and offering analytical feedback on performance. Through the use of data-driven insights and predictive analytics, AI has the potential to revolutionize surgical training, offering customized and efficient learning pathways.
This scoping review aims to map and analyze current applications of AI in surgical training, assessment, and evaluation, identifying the most common surgical procedures, AI techniques, and training setups while highlighting gaps and opportunities for future research. The following research questions guided this study:
What are the specific surgical procedures where AI algorithms are most frequently applied in surgical training?
Which AI techniques have been used in surgical training and evaluation?
How are AI techniques being used to assess and improve surgical training?
How do AI applications in surgical training affect the learning curve of surgical residents and fellows?
The paper is organized as follows: the “Methods” section outlines the methodology used to carry out this scoping review. The “Results” section provides a comprehensive overview of the findings, shows additional findings, and identifies potential areas for opportunity. The “Discussion” section presents an outline of the research questions, shows additional findings, identifies potential areas for opportunity, acknowledges the limits of the current review, and concludes with final thoughts and directions for future research in the realm of AI in surgical education.
Although there are different definitions and approaches to what AI is, this study is particularly interested in Russell and Norvig’s [14] approach to systems that act rationally, that is, systems that act intelligently and rationally, ideally in the best possible way given the available information. AI is a disruptive technology that is reshaping education, facilitating a shift toward more efficient teaching protocols [15]. It enables machines to imitate various complex human skills, and AI-based techniques are typically employed in the following areas:
Expert systems “emulate the behavior of a human expert within a well-defined, narrow domain of knowledge” [16].
Intelligent tutoring systems emulate “model learners’ psychological states to provide individualized instruction. They… help learners acquire domain-specific, cognitive, and metacognitive knowledge” [17].
AI can be subdivided into machine learning (ML), which further includes deep learning (DL). ML aims to “perform intelligent predictions based on a data set” [18]. It uses statistical, data mining, and optimization methods to design models that can identify patterns and make predictions with higher precision than human experts. In this field, there are 3 fundamental ML paradigms:
Supervised learning uses input data and their matching labeled output to train models [19]. A labeled output is data that has been assigned labels to add context; consequently, the objective of supervised learning is to learn and predict outputs for unseen data based on the initial input-output pairs.
Unsupervised learning involves working with unlabeled data [20]. The algorithms autonomously attempt to discern patterns and relationships within the data.
Reinforcement learning uses an autonomous entity known as an agent, which learns to make decisions by performing activities inside an environment to reach a specific objective [21]. The feedback the agent receives in the form of rewards or penalties serves as a guide as it iteratively refines its strategy to achieve optimal performance.
Finally, DL is a branch of machine learning that uses artificial neural networks to replicate the sophisticated processes of the human brain [22]. Algorithms in this category learn to identify patterns and comprehend large datasets. DL is highly efficient because it can automatically extract and learn high-level characteristics from data, reducing the need for manual feature selection. It excels at handling complex tasks such as image and audio recognition, natural language processing, image generation, and data-driven prediction.
Numerous models have been developed within AI to address challenging problems and tasks across different sectors and research fields. Each approach provides certain advantages specific to the type of data to be processed and the analytical needs (see Textbox 1).
Approaches and advantages specific to the type of data to be processed and analytical needs.
Regression analysis forecasts a continuous output by considering one or more predictor variables [23].
Cluster analysis methods group similar items based on shared characteristics. These algorithms help identify patterns within the data [24].
Support vector machine (SVM) categorizes data by identifying the optimal boundary that divides distinct groups [25].
Decision trees analyze data by using a series of questions and rules, resulting in the generation of predictions or classifications [26].
Random forest (RF) uses a set of decision trees to enhance predictive precision and mitigate overfitting, a phenomenon in which predictions are accurate for training data but not for new data [27].
Bayesian networks model the relationships and dependencies among variables using probability theory [28]. They are represented through a directed acyclic graph. This approach facilitates the prediction of outcomes based on established conditions.
Markov models represent the transitions between states in a system using probabilities [29]. They are characterized by the Markov property, where the future state depends only on the current state and not on the sequence of events that preceded it.
Fuzzy systems are based on fuzzy logic, which extends classical Boolean logic to handle the concept of partial truth, where truth values can range between completely true and completely false [30].
Neural networks (NNs) are inspired by the human brain. They rely on interconnected nodes to process data and detect connections [31]. This model can be subdivided based on its specific use.
Convolutional neural networks (CNNs) process data that displays a grid-like structure, such as images [32].
Recurrent neural networks (RNNs) predict sequences [33]. They use their internal state (memory) to process sequences of inputs, such as language or time series data.
Long short-term memory (LSTM) networks are a type of RNN that can learn long-term dependencies [34]. They are ideal for activities that require comprehension of long sequences.
Deep neural networks (DNNs) consist of multiple interconnected layers of neurons [35]. These networks can learn from extensive amounts of data and detect complex patterns.
Transformers are a type of network that relies on self-attention mechanisms, allowing it to weigh the importance of different parts of the input data [36].
Large language models (LLMs) are advanced types of networks that have been trained on vast datasets of words and sentences [37]. They produce coherent, human-like responses to written text by selecting the most probable next words.
These AI models highlight the potential of this technology in educational contexts. The United Nations Educational, Scientific and Cultural Organization indicates that digital technologies have the potential to complement, enrich, and transform education, aligning with the United Nations’ Sustainable Development Goal 4 (SDG 4) for education and providing universal access to learning [38]. Consequently, the integration of AI in surgical training could boost independence, self-study, engagement, and motivation.
Discussion
Principal Findings
This section discusses the study’s implications and contributions to the field. The review maps and analyzes current applications of AI in surgical training, assessment, and evaluation, identifying the most common surgical procedures, AI techniques, training setups, and highlighting gaps and opportunities for future research. The results show that AI is most frequently reported in data-rich, risk-mitigated environments, notably simulation training and box-trainer setups, and that ML (unspecified) and DL (unspecified) approaches dominate model choices.
Within these settings, many studies report models that leverage synchronized inputs, for example, kinematics, video, and other performance metrics, to classify technical skill using consistent criteria, to characterize learning trajectories across repeated attempts, and to localize performance-limiting behaviors at the level of gestures, steps, or procedural phases. When embedded in iterative practice, these capabilities may enable individualized training pathways that adjust task parameters and feedback density to a trainee’s evolving competence, with the potential to shorten time to proficiency and to reduce instructor workload. These implications are consistent with the results, in which simulation training accounted for 36% (20/56) and box trainer setups for 23% (13/56) of the included studies.
Findings in Relation to the Research Questions
Regarding the first research question aimed at identifying the specific surgical procedures where AI algorithms are most frequently applied in surgical training, AI use concentrates on MIS skills [40-54], neurosurgery [55-65], and laparoscopy [66-74]. Rather than simple frequency, the common thread across these areas is structured, high-signal data capture and well-specified tasks. Endoscopic and robotic workflows generate synchronized video, robotic kinematics, and simulator logs, which enable reproducible labels such as phase boundaries, gesture events, and Objective Structured Assessment of Technical Skills–aligned rubrics. This ecosystem lowers barriers to annotation and validation, thereby accelerating method development. Beyond these clusters, activity in ophthalmology [78-80], open surgery [84,85], robot-assisted surgery [81-83], and single-study specialties including radiology [93], urology [94], and vascular surgery [95] signals a widening scope. However, these domains often face less standardized capture or a more variable field-of-view, which complicates model training and external validation. The overall distribution, therefore, appears to reflect data tractability and curricular formalization more than inherent differences in educational need.
The second research question investigated which AI techniques have been used in surgical training and evaluation. Studies use ML (unspecified) [44,46,60,67,69,73,80,86,87,89-91] and DL (unspecified) [48,66,70,82,84,92] as broad families, with task-appropriate specializations such as CNNs for video [42,52,65,68,85,88] and SVMs for lower-dimensional kinematics or hand-crafted features [47,56,58,64,71]. NNs [59,75,76,93] support competency modeling when feature engineering is feasible, and CNN+LSTM hybrids [51,54] target temporal dynamics for suturing and task segmentation. DNNs are explicitly mentioned in [62,79]. Single-study categories (fuzzy systems [40,41], regression analysis [45], Markov chains [53], tutoring system (unspecified) [63], Bayesian network [83], transformers [94], and SVM+RF [95]) illustrate exploratory breadth rather than established consensus. Consistent with coding, CNN+LSTM is treated as a distinct class and not double-counted under CNNs. No single approach emerges as universally optimal; instead, methods align with task structure (classification vs sequence prediction), signal characteristics (video and kinematics), and assessment granularity (summative scores versus frame- or gesture-level feedback).
The third research question investigated how AI techniques are being used to assess and improve surgical training. Across setups, a common pattern is the move from retrospective, manual scoring to prospective, automated analytics that are both standardized and timely. In simulation training, synchronized streams enable immediate feedback and progression gating, which supports deliberate practice cycles grounded in objective metrics. This aligns with the preponderance of simulation studies in the dataset and the consistent application of ML and DL to transform kinematics and video into competency-linked outputs. In box trainers, models quantify motion economy, tool path quality, and task efficiency, enabling skill stratification and targeted coaching [40-43,46-50,52,66,70,71]. In robotic systems on the da Vinci platform, studies demonstrate automated assessment, uncertainty-aware feedback, and domain adaptation for cross-site or cross-task transfer [44,45,51,67,82,83]. In surgical video pipelines, investigators focus on procedural understanding, ergonomics, and fine-grained performance analytics [73,74,78-80,84,85,91,94]. The unifying mechanism across these contexts is measurement at scale that reduces feedback latency, increases consistency, and enables adaptive progression rules without displacing instructor oversight.
Finally, the last research question investigated the way in which AI applications in surgical training affect the learning curve of surgical residents and fellows. Multiple studies report outcomes consistent with accelerated learning and improved technical performance under AI-enabled training. This includes predictive modeling of progression [49], metric selection and learning-curve characterization in simulation [55], a randomized comparison of feedback modalities [57], competency-based training backed by neural models [59], continuous monitoring of bimanual expertise with deep models [62], and competency estimation in laparoscopic training [69]. Evidence from robotic contexts shows that automated assessment can structure practice with short feedback loops [45]. That said, effect sizes remain difficult to aggregate due to heterogeneous study designs, small sample sizes, nonstandard outcome measures, and limited external validation. The most defensible interpretation is that personalized, data-driven feedback and objective, repeated measurement are plausible mechanisms for the observed gains, with further multicenter validation needed to establish generalizability and durability.
The findings suggest that current AI deployment in surgical training follows data availability and standardization, that ML/DL with video and kinematics are dominant because they best match that data, and that automated, timely feedback is the primary lever through which AI influences performance and learning. Where capture is less standardized or external validation is sparse, adoption tends to lag. This synthesis directly motivates the recommendations presented later in the Discussion section on common benchmarks, transparent reporting, and SDG 4–aligned scalability.
Comparison With Previous Work
Systematic literature reviews in surgical training found in the literature have focused on specific training methods (eg, simulation-based training) or on specific types of surgery (eg, plastic surgery and orthopedic surgery) rather than providing a cross-specialty map of AI methods for training, assessment, and evaluation. Reviews focused on simulation-based training within specific domains underscore this pattern. Lawaetz et al [96] examined simulation-based training and assessment in open vascular surgery, cataloguing common methods and commenting on effectiveness within that context. Abelleyra Lastoria et al [97] surveyed simulation-based tools in plastic surgery and concluded that the validity of many approaches requires further investigation. Woodward et al [98] reached a similar conclusion in orthopedic surgery, noting concerns about the construct validity and methodological rigor of simulation studies. Reviews centered on robotic-assisted surgery also reflect divergent emphases: Rahimi et al [99] provided a descriptive overview of training modalities and assessment practices, whereas Boal et al [100] explicitly scrutinized AI methods for technical skills in robotic surgery and highlighted that both manual and automated assessment tools are often insufficiently validated.
Closer to the scope of the present scoping review, several analyses have examined automation and AI across surgical training tasks. Levin et al [101] identified families of automated technical skill assessment methods, including computer vision, motion tracking, ML and DL, and performance classification, but did not synthesize evidence on educational effectiveness. Lam et al [102] focused specifically on ML methods and reported accuracy rates that generally exceeded 80 percent across included studies, offering a performance-oriented view rather than a training-context analysis. Pedrett et al [103] emphasized the central role of video-derived motion and robotic kinematic data as inputs to AI models for technical skill assessment in minimally invasive surgery, reinforcing the importance of structured, high-signal data streams.
Findings from the present review are consistent with these previous observations in several respects. First, the centrality of simulation and other risk-managed environments recurs across literature, reflecting where ground truth is tractable and measurement can be standardized. Second, many reviews identify validation gaps, noting that reported metrics, dataset partitions, and labeling practices vary widely, which complicates comparison across sites and inhibits external generalizability [96-100]. Third, there is broad agreement that AI-assisted assessment is advancing rapidly in robotic and minimally invasive settings; yet, many frameworks remain descriptive or single-center, and their educational impact is not consistently established with robust designs [99-103].
At the same time, this review differs from earlier work in several ways. The scope extends across specialties and across training setups, linking procedures, techniques, and use cases in a single comparative framework. Rather than isolating a single algorithm family or specialty, the analysis connects the dominant AI techniques to the data modalities they exploit and to the assessment functions they serve. This mapping clarifies why ML and DL approaches, particularly CNN-based and hybrid temporal models, are prevalent where high-quality video and kinematics are available, and why adoption is slower where capture is less standardized. In addition, the review integrates signals relevant to learning curves, highlighting studies that associate AI-enabled feedback with improvements in proficiency trajectories, while also acknowledging heterogeneity and the need for external validation. By taking this comparative perspective, the review identifies shared deficiencies that cut across specialties, including nonstandard outcome measures, limited transparency in algorithmic reporting, and sparse multicenter testing, and points toward future work on benchmarks, interoperable data schemas, and scalable deployment aligned with SDG 4.
Whereas previous reviews have been primarily domain-specific or method-specific, this scoping review offers a cross-specialty synthesis that links where AI is used, which techniques are used, and how they are used to support training and assessment. This perspective complements existing literature by emphasizing comparability across contexts, illuminating mechanisms by which AI influences learning, and articulating the methodological steps needed to translate promising prototypes into reproducible, generalizable, and educationally meaningful tools.
Strengths and Limitations
This scoping review offers a broad, cross-specialty perspective on the application of AI in surgical training, assessment, and evaluation. It maps procedures, techniques, and training setups within a single comparative framework, which supports interpretation across contexts rather than within a single specialty. The review adheres to PRISMA-ScR guidance, applies explicit inclusion and exclusion criteria, and uses transparent counting rules that assign each study a primary AI technique and a primary setup to avoid double-counting. Results are presented as both narrative synthesis and structured summaries. The Discussion integrates an SDG 4 perspective, offering concrete implementation considerations related to access, scalability, and equity. Together, these elements provide a panoramic view of where AI is currently deployed, why certain methods dominate in specific data environments, and how these choices influence assessment and feedback in practice.
Several constraints should be considered. First, the search was limited to English-language publications and to the period ending March 18, 2024, which may omit relevant work outside this window. Second, many articles describe methods only at a general label level (AI, ML, and DL) without specifying architectures or training details, which limits interpretability and reproducibility. Third, the evidence base is concentrated in simulation, box-trainer, and video-centric settings, which may not fully capture transfer to live clinical performance, patient outcomes, or longer-term retention. Fourth, external validation is limited, as relatively few studies report multicenter testing, performance under domain shift, subgroup analyses, or calibration, which constrains confidence in portability.
To address these limitations, educational outcomes should also be mapped to recognized competency frameworks and reported with standardized metrics that enable replication and meta-synthesis. When multisetup or multi-technique pipelines are used, authors should specify proportional attribution. Reporting on access, resource requirements, and cost per trainee hour will support the deployment and equity assessment of SDG 4. Multicenter collaborations that release shared benchmarks and interoperable datasets will be necessary to improve reproducibility and to allow fair comparisons across techniques and settings.
Future Work Recommendations
This scoping review identified current applications of AI in surgical education and highlighted priority areas for further work. As summarized in Table 6 and visualized in Figure 2, a large proportion of studies focus on simulation training [53-64,75-77,81,86,88,90,93], representing 36% (20/56) of the included articles. This concentration reflects the suitability of simulation for controlled data capture and iterative practice. Building on this foundation, AI can enhance simulation-based training with realistic, adaptive, and personalized learning experiences [104,105], while also enabling standardized and rapid feedback that supports deliberate practice.
Advances in computer vision are particularly significant where high-quality video and kinematic data are accessible, which aligns with the prevalence of simulation and box-trainer studies in the included literature. In these regulated, risk-mitigated environments, AI systems can produce timely and structured feedback linked to defined competency frameworks, including economy of motion, bimanual coordination, camera control, tissue handling, and ergonomics, thereby facilitating deliberate practice. Although natural language processing technologies are less represented in the current review, their growing maturity suggests near-term opportunities to integrate narrative guidance, rubric-based feedback, and reflective prompts alongside quantitative metrics, provided such outputs are aligned with curricular objectives and are appropriately validated.
Future efforts should pursue 5 complementary directions.
First, strengthen external validity. Studies should include multi-institution cohorts, predefined external test sets, and reporting of performance under domain shift, including different camera views, instruments, and case difficulty. Where feasible, researchers should evaluate the transfer from simulation or bench-top tasks to higher-fidelity or clinical settings with clearly specified outcome measures and follow-up intervals.
Second, standardize educational outcomes. Investigators should map AI outputs to recognized competency frameworks and report validity, reliability, learning curve parameters, and time to competency with consistent definitions. Agreement on core outcome sets will enable comparison across techniques and facilitate meta-synthesis.
Third, expand the breadth and transparency of data. New work should prioritize multimodal capture that combines video, kinematics, tool telemetry, where appropriate, eye tracking or physiological signals. Public or data-sharing consortia should release interoperable schemas, labeling protocols, and benchmark tasks that are specific to procedures and skill elements. Clear descriptions of models and training and validation splits will improve reproducibility.
Fourth, improve usability, equity, and scalability in alignment with SDG 4. Models should operate on standard hardware, interoperate with existing simulators and video platforms, and function reliably in low-bandwidth or offline environments. Reporting of access, installation steps, resource needs, and cost per trainee hour will support adoption in diverse settings. Interfaces should disclose uncertainty, make feedback interpretable, and integrate into educator workflows without adding undue burden.
Fifth, broaden methodological scope responsibly. There is an opportunity to study natural language technologies for rubric-based guidance, structured debriefs, and reflective prompts, provided outputs are aligned with curricular objectives and validated for educational use. Prospective trials that compare feedback modalities and density, and that measure downstream retention and transfer, will clarify how AI should be integrated pedagogically.
Together, these directions could move the field from promising prototypes toward reproducible, generalizable, and educationally meaningful tools that improve surgeon training while supporting equitable access to high-quality education.
Conclusions
This scoping review maps current applications of AI in surgical training, assessment, and evaluation across procedures, techniques, and training setups. From 1400 records, 56 studies met the inclusion criteria, with activity concentrated in minimally invasive surgery, neurosurgery, and laparoscopy. AI is most frequently deployed in data-rich, risk-mitigated environments, particularly simulation training and box trainers, where synchronized video and kinematic streams support objective measurement and timely feedback. Technique choices reflect these data conditions, with ML (unspecified) and DL (unspecified) methods predominating and task-specific variants, such as CNNs and hybrid temporal models, applied to video-centric problems.
Across settings, studies describe automated skill assessment, structured formative feedback, and adaptive progression, with several reporting improvements consistent with accelerated learning curves. At the same time, heterogeneity in study design, small samples, nonstandard outcome measures, and limited external validation constrain strong inferences about effect sizes and generalizability. The evidence, therefore, supports cautious optimism that AI-enabled feedback can enhance skill acquisition, while underscoring the need for more rigorous evaluation.
Future work should prioritize precise reporting of models and datasets, multicenter validation, and standardized educational outcomes linked to recognized competency frameworks. Interoperable data schemes, shared benchmarks, and transparent methods will be essential to enable comparison across sites and techniques. Attention to scalability, access, and usability will support alignment with SDG 4, ensuring that benefits extend beyond well-resourced centers. With these elements in place, AI has the potential to deliver reproducible, equitable, and educationally meaningful gains in surgical training.