With more than 320 million surgical procedures performed worldwide annually, there is a global responsibility to enhance the quality of surgical care [1]. Complications such as surgical site infections and stroke or heart and lung complications, whether minor or severe, often lead to reoperations, morbidity, and prolonged hospital stay [2,3]. Death following surgery approaches 5% and every tenth patient experiences preventable surgical complications [1]. Adverse events also impact surgeons as a second victim [4], and health care resource use rises notably [5]. The added cost of surgical complications ranges from US $3.5 to $10 billion yearly and is associated with an average increase in hospital stay of 11 days [6]. Despite attempts to optimize adherence to clinical pathways to reduce the frequency of surgical complications, complications persist [7,8].

Artificial intelligence (AI) is transforming the field of surgery, offering unprecedented advancements in precision, efficiency, and patient outcomes that could possibly reduce surgical complications [9]. For example, AI-assisted frame reviews in neuroscience demonstrate that AI can help both junior and senior clinicians perform better [10]. In educational platforms, AI is outpacing traditional coaching programs as demonstrated by a gallbladder surgery program [11]. By leveraging the power of machine learning, natural language processing, and computer vision, AI can enhance various aspects of surgical practice from preoperative planning to intraoperative guidance to reduce surgical complications [12].

During preoperative discussions, the surgeon and patient must weigh the benefits of surgery against the risks. AI can rapidly process large amounts of health data, unburdening health personnel and allowing them to better inform their patients [13]. These AI models analyze patterns and their relationship to determine complex combinations that can indicate the patient’s risk for surgical complications. There is a gap in studies focusing on the adoption and clinical validation of these AI models [14-21]. Existing literature is focused on the retrospective development of models aimed at preventing surgical complications [9]. There are several reasons why these models have not been widely adopted in actual clinical use, including the lack of validation, lack of supporting data, differences in culture and behavior, and organizational structure [22,23]. Most common problems are regulatory and ethical constraints given that AI in surgery is considered high risk, time-consuming, and expensive. Randomized controlled trials (RCTs) are generally needed, which do not exist [24]. Also, there is currently a gap in the literature regarding the evaluation of AI models in clinical practice. This review aims to provide an overview of the published studies. Previous reviews have focused exclusively on studies involving the development and validation of models conducted using retrospective data [25-27]. Our focus is to uncover AI models that have been prospectively tested with real-time data at the bedside and to pinpoint the barriers and facilitators to implementing these models for the prevention of surgical complications.

The review was conducted in accordance with the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines (Checklist 1) [28]. We used a scoping review protocol that addressed key concepts and types of evidence. The aim was to map the existing literature by systematically searching, selecting, and synthesizing current evidence on AI models designed to prevent surgical complications using real-time data [29]. In this review, our definition of AI models also encompasses traditional statistical models, such as the National Surgical Quality Improvement Program (NSQIP), as these were tested prospectively. We relied on the five-stage framework proposed by Pollock et al [30]: (1) developing the review objective, (2) applying the eligibility criteria, (3) selecting the articles, (4) extracting and analyzing the data, and (5) reporting the results. The inclusion criteria were implemented using the population-concept-context framework, where population represents the patients undergoing surgery, concept includes the use of AI models to prevent surgical complications, and context includes studies that are conducted pre-, peri-, and intraoperative with both real-time and retrospective data excluding retrospective validation studies without prospective evaluation. The research team comprised both surgeons and machine learning engineers. The primary research question was “Which AI models have been clinically tested for the prediction of surgical complications?” The secondary research question was “For AI models not yet implemented in routine clinical practice, what barriers hinder their implementation, and are there any models on the horizon that could readily be adopted? ” We included prospective, observational, and interventional peer-reviewed studies in English that included model development, validation, and implementation. A comprehensive search strategy was developed in collaboration with a medical librarian to identify peer-reviewed original studies. We searched the following databases from October 2024 to January 2025: Scopus, CINAHL, the Cochrane Library, PubMed, MEDLINE, Web of Science, Embase, Epistemonikos, and IEEE Xplore [28]. The search strategy combined controlled vocabulary (eg, Medical Subject Headings and Emtree) and natural language keywords using Boolean operators and truncation to capture variations in terminology. The specific search terms and keywords were defined following iterative literature searches and several rounds of discussion among the authors.

The search query was structured around 3 key concepts:

The search was restricted to articles published in English. Studies were included if they described prospective model validation or clinical implementation.

A reference list of selected articles was used to extract additional articles to get a complete overview of the field. Detailed information on the search strategy can be found in Multimedia Appendix 1. The librarian vetted the initial search, using Mendeley (version 2.129.0; Elsevier). Eligibility assessment and screening were independently conducted by the primary investigator (KM) and co-investigator (ELJ) based on the established inclusion and exclusion criteria. After the initial screening, a full-text assessment was carried out. Disagreements were arbitrated by a third reviewer (CT-O). The reporting quality of the included studies was assessed by using the TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) with the AI statement [31,32]. The checklist includes 22 items (27 for the AI statement) with the potential answer options: “yes,” “no,” and “not applicable.” The 2 reviewers assessed the included studies for compliance with the items described in the TRIPOD+AI checklist. Furthermore, the following information was extracted: bibliographic details, study design and setting, surgical specialty and procedure type, AI model and technical details, predicted outcome or complications, stage of implementation, validation, and reported barriers and facilitators to clinical implementation. Quantitative characteristics derived from the included studies were summarized using tables and figures. Thematic analysis was performed on qualitative data related to barriers and facilitators. For this scoping review, we developed an analytical categorization framework to systematically classify the included studies according to the primary applications of the AI models they used (Table 1). This framework served to structure the evidence by grouping studies into conceptually coherent domains, thereby facilitating a clearer understanding of the thematic focus, methodological approaches, and applied contexts represented across the studies.

A PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) [33] flow diagram, as shown in Figure 1, illustrates the study selection process. The initial search yielded 199 articles, and 76 additional articles were gleaned from reference lists, for a total of 275 records screened. Of these 275 records, 19 studies met the inclusion criteria for this scoping review. The majority of studies were conducted in high-income countries, with the United States (n=7) being the most frequent contributor. The studies used a prospective study design, including RCTs, pilot interventional studies, and prospective cohorts. There were 8 RCTs and 11 prospective studies with a strong trend toward pilot-scale prospective studies. Large-scale validation or postdeployment studies were lacking. Few studies were evaluated outside of controlled research settings. External validation of the AI models was infrequent and adherence to TRIPOD+AI was poor. No study fully met the criteria for transparent reporting (Table 2).

To organize the heterogeneous literature, we developed a thematic categorization framework based on the primary intended function of each model as described in the original publications. The categorization was not based on the underlying statistical methodology, but on how the model output was framed and intended to be used in the clinical setting. Models categorized as “risk for complications” primarily focused on estimating the probability of specific postoperative outcomes, with performance evaluation as the main objective and limited emphasis on downstream clinical use. In contrast, models categorized as “decision support tools” were explicitly presented as supporting clinical decision-making processes, such as preoperative planning, shared decision-making, or patient counseling. We acknowledge that these categories are not mutually exclusive and that several models could reasonably fit more than one theme. In such cases, classification was based on the dominant emphasis in the study objectives and presentation (Table 1).

Study demographics and their respective funding are presented in Table 3. A total of 11 studies tested intraoperative hemodynamic monitoring and complication prediction using the Hypotension Prediction Index (HPI), while the remaining studies included AI models that addressed predicting general complications (n=7) and image analysis (n=1). While this review identified diversity of AI model applications, the majority of studies evaluated HPI, potentially skewing the findings through over-representation of intraoperative hemodynamic monitoring as the primary area of AI use in surgery. As such, the generalizability of results across surgical domains is narrowed. The predominance of HPI-related research may reflect a commercially available and well-integrated AI model, with greater funding and dissemination of pathways than other early-phase AI models. This may introduce publication and funding biases, where more rigorously tested, industry-supported models are over-represented compared to academic, noncommercial exploratory models. However, most of the studies testing the HPI had funding from the manufacturer Edwards Lifesciences. The authors stated that manufacturers were not involved in the conduct of the studies and did not approve or disapprove of the manuscript. Moreover, the clinical end points targeted by HPI are relatively narrow compared to the diverse risks associated with surgery. This limits the scope of AI use in surgery in terms of complication prediction, workflow optimization, and personalized surgical planning. Few studies evaluated AI models for long-term surgical complications, and the under-representation of other AI models narrows applicability. The main population studied was high-risk patients in 4 studies, with a subspecialty lens of noncardiac surgical patients. Table 4 delineated performance metrics of the AI models.

The American College of Surgeons (ACS) NSQIP Surgical Risk calculator is a widely adopted quality improvement tool used globally, and the HPI is a commercially available, regulatory-approved medical device (Acumen IQ) deployed in operating rooms. However, few of the other tools are deployed in surgical practice. In this regard, some of the common barriers included lack of external validation, limited generalizability, and black box model opacity (Table 5). Clinicians reported low trust in AI models (AI illiteracy and workflow issues as barriers) but noted real-time performance benefits (integration with existing platforms and clinical support as facilitators). Most of the AI models required high implementation costs, and together with the lack of financial incentives and reimbursement structures, these represented the greatest challenges to implementation.