术中实施人工智能临床决策支持系统的伦理考量:一项范围综述
Ethical considerations for intraoperative implementation of artificial intelligence clinical decision support systems: a scoping review
一项范围综述梳理术中人工智能临床决策支持系统的伦理与实施问题,检索5个来源共3,020条记录,2,535条去重记录进入筛选。综述评估验证方式、部署成熟度与测试多样性,识别知情同意、监督、透明度、监管合规、偏倚和数据治理等报告缺口。作者提出 Ethics-Implementation Scorecard 概念框架,用于刻画术中 AI 临床决策支持系统的伦理准备度。
Introduction
Artificial intelligence (AI) is rapidly transitioning from retrospective analytics to real-time clinical decision support (CDS) in procedural and intraoperative care1. Across surgical and interventional specialties, AI systems are now capable of reliable intraoperative image interpretation, physiologic prediction, and live therapeutic recommendations2. Predominantly deployed in a “human-in-the-loop” structure, these AI-based CDS tools provide advisory outputs while clinicians retain the ultimate decision authority and a promise of enhanced situational awareness and improved physiologic stability in high-risk environments3,4.
Recent research demonstrates that AI improves detection accuracy, physiologic control, and workflow across specialties, including oncology, anesthesiology, and minimally invasive surgery5. Yet, there is currently no consensus on best performing modality types, clinical domains, and validation strategies6. Reported performance metrics are heterogeneous, with variable comparator standards, and limited multi-site or device-diverse evaluations6,7. Testing in live procedures demands complex logistics (e.g., methodological rigor, population diversity) for real-world validity, which is not consistently attained across multiple systems and testing conditions3,8.
Parallel to technical development, concerns regarding governance, transparency, accountability, and bias in surgical AI have intensified9. While regulatory frameworks continue to address safety and performance, characterization of ethical readiness intraoperatively remains poorly standardized10. Consent processes for AI-supported procedures are inconsistently described, subgroup performance reporting is uncommon, auditability mechanisms are rarely detailed, and data governance practices for high-volume surgical video or physiologic data are often summarized at a high level without operational specificity11. This requires a need for how video and physiologic data are collected, stored, de-identified, accessed, retained, and securely governed throughout AI development and clinical deployment, rather than broad statements that data governance procedures exist. Furthermore, establishing standardized ethical considerations for intraoperative AI-based CDS systems presents an important opportunity to guide safe, transparent, and equitable implementation at a global scale. This is timely as AI-based CDS systems are transitioning from algorithmic development toward prospective clinical evaluation and early implementation within real-world surgical workflows, although sustained routine clinical adoption remains limited. However, despite growing recognition of this priority, key elements such as clinician trust in AI systems and adherence patterns remain infrequently incorporated into current reporting standards10.
The intraoperative environment presents distinct ethical challenges, as medical decisions are made in real time under cognitive load, with minimal tolerance for latency, uncertainty, or system failure12. Unlike outpatient or retrospective AI applications, intraoperative errors may be irreversible, offer limited opportunity for retrospective correction, and unfold within tightly synchronized team-based workflows13. Real-time human-AI interactions in this setting introduce additional risks related to ambiguity, overreliance, and automation bias10. AI systems should therefore communicate the confidence associated with their recommendations, incorporate predefined safety responses when predictions are unreliable or system failures occur, and define clear clinician-led escalation pathways specifying when AI recommendations should be overridden or deferred to clinical judgment. Without these safeguards, AI may introduce new failure modes, including inappropriate clinician reliance on inaccurate recommendations, delayed recognition of erroneous outputs, or ambiguity regarding responsibility when AI recommendations conflict with clinical judgment14. Additionally, geographic concentration of early deployments and limited reporting on patient and device diversity raise concerns regarding generalizability and equity15. Despite these challenges, AI-related guidelines and consensus statements in surgery remain limited with no unified framework integrating ethical, regulatory, and implementation considerations for intraoperative AI-based CDS systems. In this context, technical performance metrics alone do not establish ethically responsible clinical deployment9.
This review aimed to: (1) map intraoperative AI-based CDS systems that have reached clinical use or prospective evaluation; (2) assess validation approaches, deployment maturity, and testing diversity; (3) identify gaps in ethical governance and reporting, including consent, oversight, transparency, regulatory compliance, bias, and data governance; and (4) develop an Ethics-Implementation Scorecard as a conceptual framework to characterize ethical readiness for intraoperative AI-based CDS deployment. The review is reported in accordance with the PRISMA-ScR checklist16 with key concepts and definitions summarized in Table 1.
By integrating technical validation characteristics with operational ethical domains, this work advances a translational framework for intraoperative AI-based CDS implementation. Rather than equating performance metrics with deployment readiness, we propose that ethical maturity and governance infrastructure are integral components of clinical validation in intraoperative care.
Results
Identifying relevant studies
The database search identified 3,020 records across five electronic sources. After the removal of 485 duplicates, 2535 unique records underwent title and abstract screening. Of these, 2524 were excluded, primarily due to non-intraoperative context or absence of actionable clinical decision support (CDS). 11 full-text articles and national clinical trials were assessed for eligibility, with six excluded (inappropriate setting, n = 4; inappropriate intervention, n = 2), and five meeting all inclusion criteria (Fig. 1).
The PRISMA-ScR flow diagram taken from Covidence summarizes study identification, screening, eligibility assessment, and inclusion. Records identified through database searching (n = 3020); duplicates removed (n = 485); records screened (n = 2535); full text articles and clinical trial registry records assessed for eligibility (n = 11); full texts excluded (n = 6: inappropriate setting n = 4; inappropriate intervention n = 2); studies included in final review (n = 5).
Study characteristics and AI modalities
Five intraoperative AI-based CDS systems were identified across five studies spanning four clinical specialties: interventional radiology17, gynecologic oncology (SLN deep learning registry; NCT06619002)18, and three anesthesiology-focused studies evaluating intraoperative physiologic decision support, including the RL-PRAIS proof-of-concept randomized trial (NCT07123675)19, the SEASCAPE observational registry (NCT07253012)20, and the FOCUS-AFM randomized controlled trial (NCT07314853)21. Study characteristics are summarized in Supplementary Table 2. Despite broad search parameters and inclusion of registered prospective studies, the included studies represent a highly restricted subset of clinically evaluated and prospective intraoperative AI-based CDS systems.
Geographically, two studies originated from Italy18,21, while one study each was conducted in Greece, China, and Chile17,19,20. Overall, the studies represent early clinical implementation across Europe, Asia, and South America. However, given the small study sample, these findings should be interpreted descriptively and do not necessarily reflect the broader global distribution of intraoperative AI-based CDS research or implementation. No eligible studies meeting the inclusion criteria were identified from North America, Africa, or Oceania (Fig. 2).
World map (created using mapchart.net) illustrating the countries in which intraoperative AI-based CDS systems were evaluated or registered. Color coding represents the associated clinical specialty: interventional radiology (Red; Greece;17), gynecologic oncology and anesthesiology/surgical oncology (Orange; Italy; NCT0661900218 and NCT0731485321), anesthesiology (Yellow; China; NCT0712367519), and anesthesiology (Green; Chile; NCT0725301220).
AI modalities reflected heterogenous intraoperative data streams and decision-support applications across multiple surgical specialties. Using a sample of five patients, Kagadis et al. applied a semi-automated perfusion quantification algorithm to digital subtraction angiography (DSA) sequences to generate quantitative perfusion metrics following infrapopliteal angioplasty17. The SLN Deep Learning Registry with a prospective recruitment targeting 100 participants, described development of a deep learning surgical video analysis model integrating minimally invasive surgical video with indocyanine green (ICG) fluorescence for surgical phase detection and structure identification18. Three anesthesiology-focused systems leveraged continuous physiologic monitoring to support intraoperative decision-making. RL-PRAIS employed a reinforcement learning-based CDS to generate anesthetic dosing recommendations aimed at optimizing intraoperative mean arterial pressure with planned enrollment of 40 participants19. SEASCAPE integrated multimodal physiologic data, including hemodynamic variables, electroencephalography (EEG), Analgesia Nociception Index (ANI), Train-of-Four (TOF) monitoring, ventilator parameters, and pharmacokinetic models to generate real-time remifentanil dosing recommendations20. The study aims to recruit 150 participants across three development and validation phases20. The FOCUS-AFM trial, with a plan to enroll 176 participants, utilized an AI-based hemodynamic CDS system that continuously analyzed invasive arterial blood pressure and other intraoperative hemodynamic data to recommend intravenous fluid administration during major cancer surgery21.
Across all five AI modalities, autonomy was limited to advisory decision support. Four studies explicitly described human-in-the-loop operations17,19,20,21, with clinicians retaining full authority to accept, modify, or override AI-generated recommendations, while the SLN Deep Learning Registry described a planned decision-support platform without explicitly defining autonomous functionality18. None of the included studies evaluated fully autonomous or closed-loop intraoperative decision-making, in which AI systems independently execute clinical actions without clinician confirmation. Instead, AI outputs were designed to augment clinician judgment through procedural assessment17, surgical guidance18, physiological optimization19, opioid titration20, and intraoperative fluid management21.
Importantly, only the feasibility study by Kagadis et al. reportedly analyzed patient outcomes17, whereas the remaining studies represent ongoing prospective registries or randomized trials with planned enrollment totaling 466 participants18,19,20,21.
Validation and deployment maturity
Evidence maturity across included intraoperative AI-based CDS systems spanned early clinical feasibility to planned evaluation. Study designs included one single-center feasibility study17, two prospective observational registries supporting clinical implementation18,20, and two randomized controlled trials evaluating AI-assisted intraoperative decision support19,21. Despite this progression toward more rigorous prospective evaluation, published empirical clinical evidence remained limited17. Across the included studies, clinical validation was operationalized primarily through real-time intraoperative deployment, although methodological rigor varied whether systems were evaluated during live clinical workflows with measurable intraoperative outputs, and whether evaluation occurred across multiple clinical environments.
In the feasibility study from Greece, an open-source semi-automated 2D perfusion DSA tool was used in infrapopliteal angioplasty for critical limb ischemia, with seven enrolled and five analyzed patients after two cases were excluded due to motion artifacts17. Quantitative perfusion outputs demonstrated directional changes following intervention (mean MTT = 19.1%; mean PBF + 446%; mean PBV + 69.5%), supporting the technical feasibility of intraoperative perfusion quantification in a live procedural setting17. Follow-up outcomes were reported at the limb level, with limb salvage and complete wound healing in four of five limbs (80%) at six months, although the study was designed as a feasibility evaluation without external validation17.
Registry-based development and planned translation were represented by two studies18,20. The Italian multicenter SLN deep learning registry (NCT06619002) focused on developing a deep learning surgical video model using minimally invasive surgical video and ICG fluorescence for planned surgical phase detection, structure identification, and metastasis correlation18. Similarly, the SEASCAPE registry (NCT07253012) described development of a machine learning-based clinical copilot integrating multimodal physiologic monitoring to generate real-time nociception classification and remifentanil dosing recommendations during general anesthesia20. Beyond algorithm development, the registry incorporated staged prospective clinical validation and clinician usability assessment to support future implementation20. Neither registry had posted intraoperative performance results at the time of data extraction, and validation remained prospective18,20.
Prospective randomized evaluation was represented by two studies19,21. The dual-center RL-PRAIS proof-of-concept randomized controlled trial in China (NCT07123675) evaluated reinforcement learning-based anesthetic dosing recommendations during elective non-cardiac surgery19. This study used mean arterial pressure time-in-range as the primary endpoint and secondary intraoperative measures, including MAP variability and severe hypotension and hypertension events19. The multicenter FOCUS-AFM randomized controlled trial (NCT07314853) similarly evaluated AI-assisted intraoperative fluid management during major abdominal cancer surgery, with the primary endpoint defined as intraoperative hypotension burden measured using the time-weighted average of hypotension. In both trials, AI-generated recommendations were advisory, clinicians retained full treatment authority, and no study results had been posted at the time of extraction19,21.
Despite intraoperative deployment being central to clinical validation across records, explicit external validation and diversity characterization remained limited. The feasibility study was conducted at a single center with a small sample size, limiting inference regarding transportability17. Although the remaining four studies incorporated multicenter or prospective study designs to improve generalizability18,19,20,21, none had published results available for assessment, and reporting of device diversity, patient subgroup performance, and external validation remained limited. Collectively, these findings indicate that intraoperative AI-based CDS maturity is progressing toward prospective multicentre clinical evaluation. However, evidence supporting clinical effectiveness, external validity, and cross-context generalizability remains constrained.
Ethical reporting landscape
Ethical reporting was inconsistently documented across the records, with substantial variation by publication type (peer-reviewed articles versus registry records) and limited operational detail for governance-relevant characteristics17,18,19,20,21.
Consent for AI-supported care was described in four of the five studies17,18,20,21. The interventional radiology feasibility study reported ethics committee approval and written clinical consent, but did not distinguish AI-specific consent from procedural consent17. The SLN deep learning registry required willingness to participate and provide informed consent as an eligibility criterion18. Similarly, the SEASCAPE registry required written informed consent from participants20, while the FOCUS-AFM randomized controlled trial required written informed consent and explicit consent for data processing prior to enrollement21. In contrast, the RL-PRAIS record did not specify consent procedures for AI-supported decision-making within the extracted fields19.
Reporting of data governance remained limited across all included studies. The feasibility study, SLN Deep Learning Registry, and RL-PRAIS protocol did not describe operational approaches to data governance, including access controls, data retention, auditing, or oversight of intraoperative data streams17,18,19. The SEASCAPE and FOCUS-AFM registry records provided minimal governance-related information20,21. SEASCAPE reported the presence of a Data Monitoring Committee and stated that individual participant data (IPD) sharing was undecided20, whereas FOCUS-AFM indicated planned IPD sharing but did not report a Data Monitoring Committee21. Neither study described operational governance procedures for managing multimodal physiologic monitoring or continuous intraoperative hemodynamic data20,21.
Consideration of generalizability was reported in four of the five records17,18,20,21. The feasibility study explicitly acknowledged limitations related to its small sample size17, while the SLN Deep Learning Registry identified multicenter data collection as a strategy to improve future generalizability18. Similarly, SEASCAPE incorporated a multiphase prospective design spanning diverse age groups and anesthetic techniques20, and FOCUS-AFM employed a planned multicenter randomized design in patients undergoing major abdominal oncologic surgery21. Nevertheless, none of the studies reported formal bias auditing, subgroup performance analyses, or algorithmic fairness assessments, and the RL-PRAIS protocol did not specify generalizability considerations within the extracted registry fields19.
Measures of clinician trust or adherence were not reported in any record17,18,19,20,21. Although both SEASCAPE and FOCUS-AFM maintained clinician oversight by permitting anesthesiologists to accept, modify, or override AI-generated recommendations, neither protocol included quantitative assessment of clinical reliance, override behavior, or adoption metrics20,21. Reported implementation challenges were largely technical or workflow-proximal rather than structured evaluation of human-AI interaction, auditability, or governance processes. This included motion artifacts leading to exclusions17, learning curves and mapping failure18, inability to blind anesthesiologists19, distinguishing nociception from anesthetic depth and neuromuscular blockage during algorithmic development20, and variability in clinician adherence to AI-guided fluid recommendations within an open-label trial design21.
Ethics-implementation scorecard findings
To complement technical validation reporting, a conceptual Ethics-Implementation Scorecard framework was developed and subsequently applied to all five intraoperative AI-based CDS systems using extraction fields related to consent reporting, governance descriptions, autonomy configuration, validation, design, and generalizability statements17,18,19,20,21. The scorecard characterizes ethical readiness across seven domains: (i) Safety and Risk Management, (ii) Autonomy and Human Oversight, (iii) Transparency and Explainability, (iv) Regulatory Compliance and Auditability, (v) Equity and Representativeness, (vi) Governance Readiness, and (vii) Clinical Translation Readiness. These domains were informed by commonly cited ethical AI frameworks and clinical AI reporting guidelines, drawing on principles related to safety, transparency, accountability, human oversight, and equity described in the medical AI governance literature13,22,23. Specifically, the framework builds upon the reporting recommendations of DECIDE-AI24, which emphasize early-stage clinical evaluation, safety, human factors, and implementation reporting; the SAGES White Paper25, which outlines governance, liability, regulatory oversight, and institutional responsibilities for surgical AI deployment; and the European guidance26 by Cobianchi et al. which provides recommendations for the safe, transparent, and responsible integration of AI into surgical practices. While these publications establish foundational principles for clinical evaluation, reporting, governance, and responsible deployment, they are primarily intended as reporting governance, or policy guidance rather than practical assessment tools24,25,26. Consequently, they do not provide a structured method for systematically evaluating whether these ethical implementation elements are reported across individual intraoperative AI-based clinical decision support studies. Our conceptual scorecard complements this literature by translating these established principles into seven predefined implementation domains with explicit operational elements that can be consistently applied to characterize reporting of ethical implementation readiness across intraoperative AI systems.
For each system, domains were scored present (+) when the study met at least two of the three predefined operational elements within that domain, and absent (-) when fewer than two operational elements were identified. The illustrative framework underlying the scorecard defines three operational elements (a-c) for each ethical implementation domain to guide consistent assessment. Domain definitions and criteria for scoring are summarized in the framework scorecard (Fig. 3A), and applied results across the included records are presented in the study-level scorecard (Fig. 3B). The total score represents the number of ethical implementation domains explicitly reported within a study and is intended to facilitate structured comparison across studies rather than define a threshold for ethical acceptability or clinical deployment. The proposed scorecard is hypothesis-generating and will require future refinement and external validation before its application for regulatory use.
A Blank scoring template defining seven operational domains used to assess the ethical readiness of intraoperative AI-based CDS systems. Each domain is evaluated using three predefined operational elements (a, b, c). B Application of framework to the five included AI-based CDS systems17,18,19,20,21. Domains were scored as present (+) when at least two out of three operational element was explicitly reported in the study record and absent (-) when one out of three, or none are described. Each study received a total readiness score out of seven. A higher score indicates greater reporting of operational ethical and governance elements supporting responsible clinical deployment of intraoperative AI-based CDS systems.
All five systems had explicit safety objectives within their procedural contexts17,18,19,20,21. This included perfusion improvement assessment following angioplasty17, surgical safety monitoring during sentinel lymph node dissection18, and physiologic stability targets for intraoperative anesthesia management19, optimization of intraoperative nociception management20, and hemodynamic stability during AI-guided fluid therapy21. However, only two records described additional operational safety elements beyond predefined endpoints, such as mitigation of anticipated failure modes or planned human-factors evaluation18,20. None of the studies reported calibrated uncertainty displays or comprehensive operational safety-monitoring frameworks17,18,19,20,21.
Human-in-the-loop oversight was reported in three systems, each described as advisory decision-support systems in which clinicians retained final authority over intraoperative decisions19,20,21. The 2D perfusion DSA feasibility study did not explicitly define the operational autonomy configuration despite functioning as an assistive imaging tool17, while the SLN registry described AI-based CDS development without specifying operational autonomy18. None of the studies reported mechanisms for logging clinician overrides or formal contestability pathways17,18,19,20,21.
Operational transparency was reflected through descriptions of intended system function and clinical outputs. Quantitative perfusion maps were generated in the DSA system17, planned outputs from the SLN registry included surgical phase detection and metastasis correlation signals from operative video streams18, the RL-PRAIS system generated real-time vasoactive dosing recommendations19, SEASCAPE generated analgesic guidance recommendations based on multimodal physiologic integration20, and FOCUS-AFM generated real-time fluid administration recommendations21. Although intended clinical use was well described, no records reported model versioning, subgroup performance evaluation, or algorithm-update procedures17,18,19,20,21.
Regulatory compliance mechanisms remained poorly characterized. While isolated elements such as recommendation traceability19 or identification of responsible investigators were reported20,21, no studies described a comprehensive roadmap incorporating audit logging, traceability of AI-supported decisions, and designated governance oversight for AI-mediated clinical recommendations17,18,19,20,21.
Equity-related reporting was present but limited in scope. All five studies reported basic demographic characteristics and eligibility criteria, and several incorporated multicentre recruitment or explicitly discussed generalizability considerations17,18,19,20,21. However, none of the studies reported subgroup performance evaluation, fairness analyses, or algorithmic bias assessments, and representativeness was considered through study recruitment rather than evaluation of AI performance across diverse patient populations17,18,19,20,21.
Governance elements were variably reported through ethics approval and informed consent processes. Ethics approval and informed consent procedures were described in the 2D perfusion DSA study17, the SLN registry18, and SEASCAPE20, whereas governance reporting was less complete in the RL-PRAIS19 and FOCUS-AFM21 registry records. Operational data-governance measures, including privacy protections, data stewardship, or security frameworks specific to AI deployment were not reported17,18,19,20,21.
Clinical translation readiness was consistently demonstrated. The 2D perfusion DSA study reported intraoperative feasibility outcomes in clinical deployment17, whereas the remaining studies represented prospective clinical evaluations or implementation protocols integrating AI-CDS directly into intraoperative workflows18,19,20,21. Validation approaches included pathology correlation18, randomized controlled evaluation19,21 and prospective multimodal validation20, demonstrating increasing maturity toward clinical implementation despite limited published effectiveness outcomes.
When mapped across the seven domains, transparency and explainability, along with clinical translational readiness, were most consistently reported, followed closely by safety and risk management, and equity and representativeness (Fig. 3). In contrast, regulatory compliance and auditability remained consistently underreported, while governance readiness and human oversight were inconsistently reported17,18,19,20,21. The conceptual scorecard highlights a recurring pattern in intraoperative AI-based CDS literature whereby systems increasingly describe clinically actionable decision-support workflows and intended implementation pathways but continue to provide limited documentation of the regulatory compliance and governance infrastructure necessary for ethically scalable deployment.
Discussion
This scoping review examined clinical applications of intraoperative AI-based CDS systems and identified a limited, but emerging evidence base within real surgical workflows. Across more than three thousand screened records, five studies met the inclusion criteria, highlighting a substantial translational gap between rapid AI development and real-world operative deployment. Importantly, only one of these studies has reported completed prospective clinical outcomes, while the remaining four represent ongoing prospective registries or randomized trial protocols without published results. We found that (1) intraoperative AI-based CDS systems demonstrate early technical feasibility across diverse surgical domains, (2) ethical governance, oversight, and standardized reporting remain underdeveloped relative to technical validation, and (3) clinical evidence is geographically concentrated, raising concerns about generalizability and equitable deployment. Together, these findings suggest that while intraoperative AI-based CDS is beginning to emerge in clinical practice, ethical maturity and governance infrastructure have not progressed at the same pace as technical development.
The included studies illustrate the diversity of clinical domains in which intraoperative AI-based CDS systems are appearing17,18,19,20,21. Applications ranged from imaging-based perfusion quantification during vascular intervention17 and surgical video analysis18, to AI-assisted hemodynamic optimization19, intraoperative nociception management20, and fluid therapy guidance during major cancer surgery21. Collectively, these studies reflect the expanding availability of multimodal intraoperative data streams including fluoroscopic imaging, surgical video, continuous physiologic monitoring, nociception indices, ventilator parameters, and invasive hemodynamic measurements17,18,19,20,21. Such data environments are well-suited to machine learning systems capable of real-time inference and decision support. Previous research has suggested that intraoperative environments represent a promising domain for clinical AI since surgical workflows generate continuous structured signals that can be linked to measurable outcomes and immediate decision points1,27. The findings of the current review support this view. All included systems were designed to provide actionable information during live procedures, and although only one study has reported prospective clinical performance outcomes to date17, the remaining records represent prospective observational registries or randomized clinical trials evaluating AI-supported decision-making within intraoperative workflows18,19,20,21. Nevertheless, the evidence base remains limited in scale, heterogeneous in methodology, and predominantly composed of early-stage feasibility, registry, or protocol-level studies17,18,19,20,21.
The broader perioperative AI landscape further contextualizes these findings, as AI has become increasingly integrated throughout the surgical pathway, with applications spanning preoperative assessment, operative planning, and postoperative care. Recent studies have demonstrated promising applications of AI-assisted preoperative decision support for patient selection, perioperative risk stratification, surgical planning, and resource optimization across multiple surgical specialities28,29,30. This contrast further emphasizes the translational gap identified in the present review, suggesting that, despite advances in perioperative AI, prospective evaluation and governance of AI systems used during live surgical decision-making remain comparatively underdeveloped. This distinction is further illustrated by AI systems for polyp detection during colonoscopy, which were excluded despite their clinical maturity31, as these applications are typically conducted in endoscopy suites rather than operative environments and do not involve intraoperative decision-making under surgical conditions1. Although real-time AI-assisted colonoscopy represents one of the most mature forms of procedural AI-CDS, studies in this domain were excluded due to fundamental differences in workflow complexity, team structure, and clinical risk compared to intraoperative surgical environments32. Evidence suggests that AI performance and usability are highly context-dependent, with integration challenges varying significantly across clinical settings, limiting the transferability of findings from endoscopic procedures to high-acuity operative care31,32. Excluding these studies preserves the intraoperative focus of this review but limits comparison with more mature AI-CDS domains, underscoring the relatively early stage of evidence supporting intraoperative AI-based CDS.
An important observation across the included records is that technical feasibility and procedural integration continue to outpace the operationalization of ethical governance. All five systems defined explicitly safety-relevant procedural objectives within their clinical contexts17,18,19,20,21. Similarly, all five studies clearly described their intended clinical functions and generated actionable outputs designed to support intraoperative decision-making17,18,19,20,21. However, reporting of regulatory compliance and governance mechanisms remained limited. None of the studies described comprehensive audit logging, traceability of AI-supported decisions, or operational governance frameworks, and data governance procedures were largely absent17,18,19,20,21. This discrepancy between technical deployment and governance reporting reflects a broader pattern observed across medical AI research. Multiple systematic analyses of clinical AI studies have shown that algorithmic performance metrics are frequently reported in detail, whereas elements related to regulatory compliance, reproducibility, and governance are inconsistently described3,8,33. Without characterized reporting of these elements, it remains difficult to assess whether systems are ready for safe integration into complex clinical environments such as the operating room.
The geographic distribution of studies identified in this review further highlights challenges related to generalizability and equity. The included studies were conducted across Europe17,18,21, Asia19, and South America20, with no eligible studies originating from North America, Africa, Oceania, or multi-continental collaborations. Although this pattern likely reflects differences in research priorities and regulatory pathways across regions, it underscores the possibility that intraoperative AI-based CDS evidence remains concentrated within a limited set of healthcare environments. Interestingly, this finding contrasts with broader global AI adoption trends, where North America, particularly the United States, represents one of the largest investors and developers of clinical AI technologies34. Despite this leadership in digital health infrastructure, few intraoperative AI-based CDS systems appear to have progressed to prospective surgical deployment. This paradox suggests that regulatory pathways and institutional risk tolerance may influence where AI technologies are first translated into real-world surgical environments. Geographic concentration raises important questions about the transportability of AI systems across diverse clinical infrastructures, patient populations, and resource settings34. Prior work has established that machine learning models trained in narrow institutional or demographic contexts may exhibit reduced performance when applied to external populations, raising concerns about algorithmic bias and inequitable deployment34,35,36. Ensuring that intraoperative AI systems are evaluated across diverse clinical environments is essential for establishing equitable and generalizable deployment pathways.
The results demonstrate that clinical validation presence does not necessarily indicate readiness for ethical deployment. All five studies involved prospective intraoperative clinical evaluation, including one published feasibility study17, three prospective clinical protocols18,19,20, and one multicentre randomized controlled trial21. These designs represent an important step beyond retrospective model validation because they evaluate or plan to evaluate AI systems within authentic procedural workflows. However, even in these settings, key elements of responsible AI deployment remained incompletely reported. Although one protocol planned to assess clinician usablity20 and another included clinician acceptance of AI recommendations as a secondary outcome19, none of the studies described comprehensive mechanisms for audit logging, traceability of AI-supported decisions, or retrospective governance of system outputs17,18,19,20,21. Similarly, while demographic characteristics and generalizability considerations were reported more consistently, none performed subgroup performance analyses or formal bias assessments17,18,19,20,21. These findings suggest that prospective intraoperative evaluation alone should not be interpreted as a proxy for ethical readiness. Responsible deployment requires governance mechanisms that enable oversight, traceability, and continuous monitoring of system performance once integrated into clinical workflows. International reporting and evaluation frameworks such as SPIRIT-AI, CONSORT-AI, and DECIDE-AI emphasize the importance of these elements in the clinical evaluation of AI-based decision support tools3,24,37.
To help operationalize these considerations, this review introduced a conceptual Ethics-Implementation Scorecard designed to assess the reporting of governance-relevant domains in intraoperative AI-based CDS studies (Fig. 3). The framework encompasses seven operational domains that reflect core dimensions of responsible clinical AI implementation: safety and risk management, human oversight, transparency, regulatory compliance, equity, governance readiness, and clinical translation readiness. Application of the framework to the included studies was intended to illustrate how these governance domains are currently reported in the intraoperative AI literature, rather than to formally validate the framework or provide a regulatory quality assessment. When applied to the included systems, the framework revealed scores ranging from 4 to 6 out of 7 (Fig. 3B). Domains related to safety, clinical translation readiness, and governance readiness were frequently reported, whereas regulatory compliance and auditability remained consistently underreported, with transparency and equity showing variable reporting. This suggests that the current literature prioritizes demonstration of technical capability and procedural integration while providing comparatively limited detail regarding the governance structures necessary for responsible deployment. Importantly, the proposed scorecard is not intended to function as a regulatory instrument or formal quality assessment tool. Instead, it offers a practical and simple framework that will assist hospitals, developers, researchers, and regulators in identifying governance elements that must be considered when evaluating intraoperative AI-based CDS systems.
The potential regulatory relevance of such frameworks is increasingly recognized as clinical AI systems move from experimental development to operational deployment. Regulatory bodies including the United States Food and Drug Administrators and European Regulators have emphasized the importance of lifecycle oversight, algorithm, transparency, and post-market performance monitoring for AI-enabled medical technologies10,38,39. This includes mechanisms for post-deployment monitoring, transparency regarding algorithm updates, and clear governance structures defining responsibility for system outputs38,39. Intraoperative AI systems introduce additional considerations because their outputs may influence time-sensitive clinical decisions during surgical procedures. Establishing standardized expectations for reporting governance and oversight elements may therefore support safer implementation pathways and facilitate regulatory evaluation of AI-based decision support systems.
Several limitations should be considered when interpreting the findings. First, only five studies met the inclusion criteria, highlighting the nascent state of the literature surrounding intraoperative AI-based CDS. Although this scarcity represents an important finding, it limits the depth of empirical synthesis and the ability to draw robust conclusions regarding clinical effectiveness, implementation outcomes, or performance comparisons across systems. Second, only the 2D perfusion DSA feasibility study reported completed clinical results, based on five analyzed patients17. The remaining four records were prospective registries or randomized trial protocols with planned enrollment but no posted results18,19,20,21. Although these reports provide valuable insight into emerging approaches and planned clinical evaluation, they do not yet contribute empirical evidence regarding safety, efficacy, workflow integration, clinician interaction, or patient outcomes. Moreover, the only system with published clinical findings used algorithmic image-processing rather than learning, further emphasizing the limited empirical evidence supporting intraoperative AI-based CDS. Third, heterogeneity across surgical specialties, AI modalities, deployment environments, study designs, and measured outcomes constrained direct comparisons and opportunities for standardized synthesis. Additionally, the proposed Ethics-Implementation Scorecard should be interpreted as an illustrative conceptual framework informed by established AI reporting and ethics guidance, rather than a validated evaluation instrument or measure of deployment readiness. Its application to five heterogeneous records demonstrates its potential utility for structuring ethical and governance reporting, but future refinement, stakeholder evaluation, and external validation across a broader evidence base is required. Lastly, although demographic and generalizability information was reported, no studies reported subgroup performance analyses, formal bias assessments, or fairness evaluations, limiting the assessment of equity across patient populations, institutions, and clinical settings.
Despite these limitations, the findings suggest several priorities for future research. First, completed prospective and multicentre evaluations of intraoperative AI systems are needed across diverse clinical environments, patient populations, devices, and institutional workflows. As the four ongoing registries and trials identified in this review mature and additional AI-based CDS systems enter prospective evaluation, their findings may provide the broader external evidence required to refine and validate the proposed conceptual scorecard. Repeated application across emerging systems could determine whether its domains consistently identify meaningful gaps in safety, oversight, regulatory compliance, equity, governance, and clinical translation, and support development of a more reproducible and measurable ethical reporting tool. Second, standardized reporting frameworks tailored to intraoperative AI-based CDS may improve transparency regarding governance, oversight, and human-AI interaction during surgical workflows. Existing reporting guidelines provide valuable foundations but may not fully capture the unique characteristics of intraoperative environments, including real-time decision support, multimodal data streams, and complex team dynamics in the operating room. Third, adoption of intraoperative AI-based CDS may be influenced by concerns among surgeons and medical staff regarding potential elimination of clinical autonomy or replacement of human expertise, despite these systems being designed to function as adjunct decision-support tools with a human-in-the-loop framework. Future evaluations should therefore incorporate measures of human-AI collaboration, including clinician acceptance, adherence, modification or override of recommendations, usability, and appropriate reliance on system outputs.
In conclusion, the current evidence base for intraoperative AI-based CDS systems remains limited but reflects growing prospective evaluation across multiple surgical specialties. Existing studies demonstrate early technical feasibility in one completed feasibility study and increasing clinical interest through ongoing prospective registries and randomized trials17,18,19,20,21. However, governance reporting, ethical oversight, and evaluation across diverse clinical environments remain comparatively underdeveloped and should evolve alongside technical innovation. As additional prospective evidence emerges, structured conceptual frameworks that assess ethical implementation alongside technical performance may support more transparent reporting and responsible translation of intraoperative AI-based CDS into safe and equitable tools that support surgical care.
Methods
Search strategy and criteria for selection
The following databases were searched by an information specialist (M.E.) from inception via the Ovid platform: MEDLINE, MEDLINE ePubs and In-Process Citations (daily), Embase Classic+Embase, Cochrane Central Register of Controlled Trials, the Cochrane Database of Systematic Reviews, and LILACS (WHO PAHO). All the databases were initially searched on November 25, 2025, and the search was rerun on July 6, 2026, to identify newly published or registered studies and ensure the review remained current.
The search process followed the Cochrane Handbook40 and the Cochrane Methodological Expectations of Cochrane Intervention Reviews (MECIR)41 for search conduct. The PRESS guideline for peer-reviewing the search strategies42, drawing upon the PRESS 2015 Guideline Evidence-Based Checklist, was applied to peer-review the search strategy and minimize potential errors. Preliminary searches were conducted, and full-text literature was mined for potential keywords and appropriate controlled vocabulary terms (Medical Subject Headings for MEDLINE and EMTREE descriptors for Embase). The Yale MeSH Analyser43 was used to facilitate the MeSH and text word analysis.
Search strategy concept blocks were structured around Artificial Intelligence, Clinical Decision Support Systems, Validation, Surgery, and Ethics, using both controlled vocabulary and text word searching. Searches were limited to human subjects and publication years ≥2010. For the Ethics component of the MEDLINE search strategy, the following filter was used: CADTH ETHICS—Ethics, Legal and Social Issues (ELSI)—MEDLINE44. The Ovid MEDLINE search strategy is provided in Supplementary Table 1.
Studies were included if they evaluated AI-based CDS systems used intraoperatively in human surgical or interventional settings and reported empirical or planned clinical evaluation in real operative contexts. Eligible studies described systems generating actionable intraoperative outputs, such as risk prediction, alerts, or decision guidance influencing procedural management. We excluded studies conducted outside the intraoperative setting such as in endoscopy suites, automation without decision-support functionality, retrospective studies, non-human studies, and non-original publications including editorials, conference abstracts, or duplicates.
All records were imported into Covidence for screening and duplicate removal. Title and abstract screening were conducted independently by two reviewers (L.P. and A.W.). Studies meeting the inclusion criteria or lacking sufficient information for exclusion proceeded to full-text review. Full-text screening was performed independently by the same two reviewers using pre-specified eligibility criteria. Discrepancies were resolved by consensus, and unresolved conflicts were adjudicated by a third independent reviewer (W.H.S.).
Data analysis
Data extraction was conducted independently by the same two reviewers using a standardized extraction template on an Excel sheet developed a priori. Extracted variables included study characteristics, AI modality, intraoperative integration context, autonomy level, validation approach, diversity of testing environments, ethical governance elements, and reported limitations. The Ethics-Implementation Scorecard was informed as an illustrative conceptual framework by synthesizing governance and reporting domains from established clinical AI reporting guidelines3,8,24 and the broader ethical AI governance literature. Seven domains were selected to reflect key operational components of responsible intraoperative AI implementation, with each domain consisting of three predefined characteristics. The finalized dataset informed descriptive synthesis and application of the Ethics-Implementation Scorecard framework. Consistent with scoping review methodology, no formal risk-of-bias assessment was performed.
Ethics
Not applicable. Ethics approval and informed consent were not required for this study because it was a scoping review based exclusively on publicly available literature and registry records.
Data availability
Data extracted from studies included and data used for all analyses are provided in Supplementary Table 2.
References
Topol, E. J. High-performance medicine: the convergence of human and artificial intelligence. Nat. Med. 25, 44–56 (2019).
Byrd IV, T. F. & Tignanelli, C. J. Artificial intelligence in surgery—a narrative review. J. Med. Artif. Intell. 7, 29 (2024).
Liu, X., Cruz Rivera, S., Moher, D., Calvert, M. J. & Denniston, A. K. Reporting guidelines for clinical trial reports for interventions involving artificial intelligence: the CONSORT-AI extension. Nat. Med. 26, 1364–1374 (2020).
Bera, K. et al. Artificial intelligence in digital pathology — new tools for diagnosis and precision oncology. Mod. Pathol. 38, 100705 (2025).
Kenig, N., Monton Echeverria, J. & Muntaner Vives, A. Artificial intelligence in surgery: a systematic review of use and validation. J. Clin. Med. 13, 7108 (2024).
Wang, Y. et al. Guidelines, consensus statements, and standards for the use of artificial intelligence in medicine: systematic review. J. Med. Internet Res. 25, e46089 (2023).
Ibrahim, H. et al. Reporting guidelines for clinical trials of artificial intelligence interventions: the SPIRIT-AI and CONSORT-AI guidelines. Trials 22, 11 (2021).
Cruz Rivera, S. et al. Guidelines for clinical trial protocols for interventions involving artificial intelligence: the SPIRIT-AI extension. Nat. Med. 26, 1351–1363 (2020).
Malpani, R. et al. Ethics and Governance of Artificial Intelligence for Health: Guidance on Large Multi-Modal Models (World Health Organization, 2024).
U.S. Food and Drug Administration. Artificial Intelligence and Machine Learning (AI/ML)-Based Software as a Medical Device (SaMD) Action Plan (U.S. Food and Drug Administration, 2021).
Weiner, E. B., Dankwa-Mullan, I., Nelson, W. A. & Hassanpour, S. Ethical challenges and evolving strategies in the integration of artificial intelligence into clinical practice. PLOS Digit. Health 4, e0000810 (2025).
Siala, H. & Wang, Y. SHIFTing artificial intelligence to be responsible in healthcare: a systematic review. Soc. Sci. Med. 296, 114782 (2022).
Idárraga Ruiz, N. et al. Accuracy and reliability of artificial intelligence in surgical decision-making: a literature review. Cureus 17, e95337 (2025).
Taylor, R. A. et al. Leveraging artificial intelligence to reduce diagnostic errors in emergency medicine: challenges, opportunities, and future directions. Acad. Emerg. Med. 32, 327–339 (2025).
Cross, J. L., Choma, M. A. & Onofrey, J. A. Bias in medical AI: implications for clinical decision-making. PLOS Digit. Health 3, e0000651 (2024).
Tricco, A. C. et al. PRISMA extension for scoping reviews (PRISMA-ScR): checklist and explanation. Ann. Intern. Med. 169, 467–473 (2018).
Kagadis, G. C. et al. 2D perfusion DSA with an open-source, semi-automated, color-coded software for the quantification of foot perfusion following infrapopliteal angioplasty: a feasibility study. Eur. Radiol. Exp. 4, 47 (2020).
ClinicalTrials.gov. Implementation of surgical safety and intraoperative metastasis identification through deep learning: multicentric video collection for minimally invasive sentinel lymph node dissection in uterine malignancies (LYSE). Identifier: NCT06619002 (ClinicalTrials.gov, 2024).
ClinicalTrials.gov. AI-assisted blood pressure control during anesthesia (RL-PRAIS PoC). Identifier: NCT07123675 (ClinicalTrials.gov, 2025).
ClinicalTrials.gov. AI-assisted Analgesia Copilot System (SEASCAPE). Identifier: NCT07253012 (ClinicalTrials.gov, 2025).
ClinicalTrials.gov. Using artificial intelligence to guide fluid therapy during major cancer surgery: a randomized controlled trial (FOCUS-AFM). Identifier: NCT07314853 (ClinicalTrials.gov, 2025).
Singh, M. P. & Keche, Y. N. Ethical integration of artificial intelligence in healthcare: narrative review of global challenges and strategic solutions. Cureus 17, e84804 (2025).
Radanliev, P. AI ethics: integrating transparency, fairness, and privacy in AI development. Appl. Artif. Intell. 39, 2463722 (2025).
Vasey, B. et al. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nat. Med. 28, 924–933 (2022).
Hashimoto, D. A., Marwaha, J. S., Lee, S. A., Schwaitzberg, S. D. & Duffourc, M. N. Risk and liability in the deployment of AI systems for surgery: a SAGES white paper. Surg. Endosc. https://doi.org/10.1007/s00464-026-12881-8 (2026).
Cobianchi, L. et al. Artificial intelligence and surgery: ethical dilemmas and open issues. J. Am. Coll. Surg. 235, 268–275 (2022).
Hashimoto, D. A., Rosman, G., Rus, D. & Meireles, O. R. Artificial intelligence in surgery: promises and perils. Ann. Surg. 268, 70–76 (2018).
Muelbauer, E. J., Alvi, M. A., Kennedy, D. J. & Fehlings, M. G. The future is now: how AI is reshaping spine care. N. Am. Spine Soc. J. 24, 100825 (2025).
Rouhi, A. D., Menon, S. V., Ghanem, Y. K. & Han, J. J. Concordance of large language model recommendations with multidisciplinary heart team decisions in coronary revascularization and aortic valve intervention: a systematic review and pooled analysis. Cardiol. Ther. https://doi.org/10.1007/s40119-026-00453-9 (2026).
Habboub, G., Berven, S., Ames, C., Peterson, T. & Mroz, T. Real-world implementation of artificial intelligence/machine learning for managing surgical spine patients at 2 academic health care systems. Int. J. Spine Surg. 17, S11–S17 (2023).
Repici, A. et al. Efficacy of real-time computer-aided detection of colorectal neoplasia in a randomized trial. Gastroenterology 159, 512–520.e7 (2020).
Heinlein, M. Artificial intelligence and the clinical gaze: visual practices of AI-assisted colonoscopy. Sociol. Health Illn. 48, e70144 (2026).
Sounderajah, V. et al. Developing specific reporting guidelines for diagnostic accuracy studies assessing AI interventions: the STARD-AI Steering Group. Nat. Med. 26, 807–808 (2020).
Kelly, C. J., Karthikesalingam, A., Suleyman, M., Corrado, G. & King, D. Key challenges for delivering clinical impact with artificial intelligence. BMC Med 17, 195 (2019).
Obermeyer, Z., Powers, B., Vogeli, C. & Mullainathan, S. Dissecting racial bias in an algorithm used to manage the health of populations. Science 366, 447–453 (2019).
Wiens, J. et al. Do no harm: a roadmap for responsible machine learning for health care. Nat. Med. 25, 1337–1340 (2019).
Domalpally, A. & Channa, R. Real-world validation of artificial intelligence algorithms for ophthalmic imaging. Lancet Digit. Health 3, E463–E464 (2021).
Aboy, M., Minssen, T. & Vayena, E. Navigating the EU AI Act: implications for regulated digital medical products. npj Digit. Med. 7, 237 (2024).
Larson, D. B. et al. Regulatory frameworks for development and evaluation of artificial intelligence–based diagnostic imaging algorithms: summary and recommendations. J. Am. Coll. Radiol. 18, 413–424 (2021).
Higgins, J. P. T. & Thomas, J. (eds). Cochrane Handbook for Systematic Reviews of Interventions, version 6.5 (Cochrane, 2024).
Cochrane. Methodological Expectations of Cochrane Intervention Reviews (MECIR) (Cochrane, 2023).
McGowan, J. et al. PRESS peer review of electronic search strategies: 2015 guideline statement. J. Clin. Epidemiol. 75, 40–46 (2016).
Yale University. Yale MeSH Analyzer (Yale University, 2025).
Canada’s Drug Agency. Ethics, Legal and Social Issues (ELSI) Search Filter for MEDLINE (Canada’s Drug Agency, 2020).
Acknowledgements
Not applicable.
Author information
Authors and Affiliations
Surgical Artificial Intelligence Research Academy, University Health Network, Toronto, ON, Canada
Luka Parikh, Ariana Walji, Caterina Masino, Wagner H. Souza & Amin Madani
Faculty of Health Sciences, School of Medicine, Queen’s University Kingston, Ontario, ON, Canada
Luka Parikh
Temerty Faculty of Medicine, University of Toronto, Toronto, ON, Canada
Ariana Walji
Library & Information Services, University Health Network Toronto, Ontario, ON, Canada
Marina Englesakis
Division of General Surgery, University Health Network, Toronto, ON, Canada
Wagner H. Souza
Department of Surgery, University of Toronto, Toronto, ON, Canada
Amin Madani
Authors
- Luka Parikh
- Ariana Walji
- Caterina Masino
- Marina Englesakis
- Wagner H. Souza
- Amin Madani
Contributions
Conceptualization: L.P., A.W., C.M., W.H.S., A.M. Librarian and Information Specialist: M.E. Methodology: L.P., A.W., C.M., M.E., W.H.S., A.M. Screening: L.P., A.W., W.H.S. Data collection (extraction): L.P., A.W., W.H.S. Writing of Original Draft: L.P. Writing - Review and Editing: L.P., A.W., C.M., M.E., W.H.S., A.M. Supervision: W.H.S, A.M. All authors reviewed and approved the final manuscript.
Corresponding author
Correspondence to Amin Madani.
Ethics declarations
Competing interests
AM is a consultant with Johnson and Johnson. The remaining authors declare no competing interests.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
About this article
Cite this article
Parikh, L., Walji, A., Masino, C. et al. Ethical considerations for intraoperative implementation of artificial intelligence clinical decision support systems: a scoping review. npj Digit. Surg. 1, 23 (2026). https://doi.org/10.1038/s44484-026-00024-4
Received: 13 May 2026
Accepted: 11 August 2026
Published: 07 October 2026
Version of record: 07 October 2026
DOI: https://doi.org/10.1038/s44484-026-00024-4
来源:Nature 健康护理主题 · nature.com