Decision-centered artificial intelligence for perioperative care outside the operating room: a practical review for surgeons Authors: Soyul Han Published: 2026 ABSTRACT Artificial intelligence (AI) is increasingly being applied across the spectrum of surgical care; however, most existing review articles have organized prior studies primarily by algorithmic type or predicted outcomes. Consequently, a structured understanding of “when” and “why” AI is integrated into real-world clinical workflows, and how its outputs inform surgical decision-making, remains insufficiently developed. Although both the preoperative and postoperative phases involve risk prediction, they differ fundamentally in their decision contexts, data characteristics, and modes of clinical application. This review sought to reorganize the surgical AI literature using a decision-centered framework. AI applications during the operation itself—including surgical video analysis, robotic automation, and real-time image guidance—are outside the scope of this review. Instead, this review focuses on AI that supports decision-making before and after surgery, emphasizing decision points relevant to minimally invasive surgical practice, including patient selection, treatment planning, surgical extent and approach planning, postoperative monitoring, discharge readiness, and surveillance planning. This review examined PubMed-indexed studies published between 2015 and 2025, analyzing the literature according to the clinical decision points each study was intended to support, rather than emphasizing comparative model performance. Among the studies reviewed, preoperative AI was predominantly applied to support patient selection and treatment planning in the context of diagnostic uncertainty. In contrast, postoperative AI was mainly used to support time-sensitive management and prognostic assessment. This review reframes surgical AI not as a standalone predictive instrument, but as an integral component of phase-specific clinical decision pathways across the surgical care continuum. INTRODUCTION Surgical outcomes are substantially determined by decisions made outside the operating room. Preoperative patient selection, risk stratification, and treatment planning, together with postoperative surveillance and management, define the perioperative decision environment in which surgical care is delivered [1]. Globally, postoperative complications remain a leading cause of surgical morbidity and mortality, and the increasing prevalence of elderly patients with complex comorbidities has intensified the need for individualized decision support across the perioperative continuum [2–4]. Traditionally, perioperative decision-making has relied on clinician judgment and standardized scoring systems such as the American Society of Anesthesiologists (ASA) physical status classification and the Charlson Comorbidity Index [5–7]. While clinically useful, these tools are based on linear additive structures and are therefore limited in their ability to capture complex, nonlinear interactions among multiple concurrent risk factors. This limitation is increasingly evident in contemporary surgical practice, where high-dimensional and time-dependent data—including electronic health records (EHRs), multimodal imaging, and continuous physiological monitoring—are routinely generated [8,9]. Machine learning (ML)-based models may complement conventional statistical approaches by capturing nonlinear relationships among clinical variables that linear methods cannot represent [10]. Artificial intelligence (AI)-based models have demonstrated improved or complementary performance across perioperative risk stratification, complication prediction, early deterioration detection, and long-term prognostic assessment [11,12]. Accordingly, research in surgical AI has expanded rapidly over the past decade [13]. Despite this growth, existing reviews have primarily organized the literature by algorithmic type or prediction targets, emphasizing cross-study performance comparisons [11,13,14]. While informative for methodological development, this approach provides limited insight into how AI is integrated into clinical workflows, which specific decisions it supports, and how its outputs translate into actionable clinical decisions. As a result, the gap between predictive performance and clinical applicability remains insufficiently addressed. This review addresses this gap by reorganizing perioperative AI around six clinical decision points at which AI-based predictions may inform surgical decision-making. These include patient selection and risk stratification, treatment strategy selection, surgical extent and approach planning, postoperative monitoring intensity, discharge readiness and care transition, and surveillance and recurrence risk planning. For each decision point, the synthesis was structured according to the data context, the form of model outputs, and their potential contribution to clinical decision-making. This review also distinguishes tools that have achieved clinical deployment from those that remain at the prototype stage, a distinction that is critical for real-world adoption. Intraoperative AI technologies—including robotic surgical systems, real-time image guidance, and instrument control—are outside the scope of this review. This review focuses exclusively on preoperative and postoperative decision support, as illustrated in Fig. 1. This review provides: (1) a structured understanding of where AI contributes to clinical decision-making across the perioperative continuum; (2) a practical distinction between investigational systems and clinically available tools; and (3) an assessment of current evidence limitations that must be addressed for broader clinical implementation. To operationalize this perspective, this review proposes a decision-centered framework that organizes perioperative AI according to the clinical decisions it is intended to support (Table 1). This framework maps AI applications to six decision points spanning the preoperative and postoperative phases and links input data, model outputs, and corresponding clinical actions within each domain. By doing so, it enables interpretation of existing studies in terms of their relevance to clinical decision-making rather than predictive performance alone. METHODS Review approach This article was designed as a narrative review focused on clinical applicability. Given the heterogeneity of study designs, AI methodologies, patient populations, and outcome definitions, and because the objective was interpretive rather than comparative, quantitative synthesis was not undertaken. Instead, this review examines how AI-based tools are applied to support specific decisions across the perioperative care pathway and narratively organizes the literature according to key surgical decision points. Comparative meta-analytic evaluation of model performance was therefore outside the scope of this review. Literature search To identify relevant literature, PubMed was searched for studies published between 2015 and 2025, a period during which ML and deep learning (DL) methods became increasingly adopted in medical research. Searches combined terms related to AI, ML, or DL with terms related to preoperative or postoperative clinical decision-making and management. These searches served as a starting point for the purposive identification of representative studies. To broaden coverage, the reference lists of key review articles and representative original studies were also examined. Foundational methodological and landmark references predating this period—such as the original Charlson Comorbidity Index [15]—were retained to provide conceptual and historical context and were not treated as primary evidence for specific decision points. Identification of representative studies Publications identified through the literature search were reviewed for their relevance to perioperative clinical decision support. Because the objective of this review was to develop a decision-centered framework rather than to generate an exhaustive evidence map, a purposive approach was used to select representative studies that illustrated clinically meaningful applications of AI in preoperative or postoperative care. Studies were considered for inclusion if they were conducted in preoperative or postoperative settings with a clearly defined clinical use case and involved the development or validation of AI-based models, including ML and DL approaches. Priority was given to studies in which model outputs could plausibly be linked to specific clinical actions, such as patient selection, treatment strategy selection, or postoperative management. To capture variation in evidentiary support and clinical implementation, prototype-stage models as well as externally validated, clinically implemented, or commercially available tools were included. Studies focused primarily on intraoperative technologies, including robotic systems and real-time image analysis, were excluded because they fell outside the scope of this review. Studies with unclear relevance to clinical decision-making (e.g., purely computational simulations) were also excluded. Non-original articles, including reviews, editorials, and conference abstracts, were not treated as primary evidence for the decision-point synthesis, although selected review articles were used to contextualize the field and to identify additional representative original studies. Full-text review was performed for studies deemed thematically relevant after initial screening. Following the removal of duplicate references, the final set of included studies was categorized into six clinical decision points spanning the preoperative and postoperative phases based on their primary clinical decision context. Decision-centered synthesis framework The selected representative studies were synthesized according to the clinical decisions they were intended to support rather than by algorithmic architecture or performance metrics alone. Clinical decision points were defined as stages within the perioperative care pathway at which AI-derived information could inform a specific clinical action, such as patient selection, treatment strategy selection, postoperative monitoring, discharge coordination, or surveillance planning. Each study was examined from three perspectives: the information available at the time of decision-making, the characteristics of the decision itself, and the extent to which AI-derived predictions were linked to actionable clinical use. When a study demonstrated model development or validation without clear integration into clinical workflow, this was interpreted as a gap between predictive performance and decision support. This approach was intended to provide an interpretive framework for clinical applicability, not to rank all available models or provide an exhaustive assessment of the literature. PREOPERATIVE ARTIFICIAL INTELLIGENCE DECISION SUPPORT In the preoperative phase, surgeons are often required to make high-stakes decisions before definitive pathological or treatment response information becomes available. Decisions regarding operative eligibility, treatment sequencing, and surgical extent therefore rely on indirect information, including imaging, laboratory data, biopsy findings, comorbidities, and functional status. Accordingly, the clinical value of AI in preoperative decision-making lies not merely in improving diagnostic accuracy, but in providing calibrated and interpretable predictions that enable the translation of uncertainty into actionable clinical decisions. Within this context, AI-assisted preoperative decision support can be organized into three key decision points: patient selection and risk stratification, treatment strategy selection, and surgical extent and approach planning. Patient selection and risk stratification Preoperative patient selection involves determining whether and when to proceed with surgery, with risk stratification serving as a central basis for this decision. Surgeons integrate patient comorbidities, functional status, laboratory abnormalities, and the urgency and invasiveness of the planned procedure to decide whether to proceed as scheduled, delay surgery for preoperative optimization, or consider nonoperative alternatives. Conventional risk assessment tools, including the ASA physical status classification, the Charlson Comorbidity Index, and the American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) Surgical Risk Calculator, have been widely used to support these decisions; however, they are limited in their ability to capture complex interactions between patient- and procedure-specific factors and to incorporate dynamically updated clinical data [15]. ML-based preoperative risk prediction models have been developed to address these limitations, although the strength of supporting evidence and the level of clinical implementation vary across tools. In emergency surgery, POTTER (Predictive OpTimal Trees in Emergency Surgery Risk) applies interpretable ML to estimate 30-day mortality and postoperative complication risk and demonstrated improved discrimination compared with conventional tools in its original development and validation study [16]. MySurgeryRisk uses EHR-derived variables to generate individualized estimates of major postoperative complications and mortality risk; however, the original study primarily represents development and validation of an automated preoperative risk algorithm, with real-time clinical workflow implementation requiring further testing [17]. The FLEX (Flexible Surgical Set Embedding) score uses routinely collected EHR data, including patient characteristics, diagnosis codes, and procedure codes, to predict postoperative mortality, readmission, discharge to long-term care, and length of stay, and has undergone multicenter validation [18]. In a large-scale health-system study, an EHR-based preoperative ML model was prospectively evaluated and clinically deployed to identify patients at high risk of 30-day postoperative mortality or major adverse cardiac and cerebrovascular events, with superior predictive performance compared with the ACS-NSQIP Surgical Risk Calculator [19]. These models are not intended to automate patient selection but to support structured preoperative risk stratification. Lower estimated risk may support proceeding with planned surgery, whereas higher risk estimates may prompt anemia correction, nutritional or exercise-based prehabilitation, cardiopulmonary evaluation, anesthesiology or multidisciplinary consultation, and intensive care unit (ICU) resource planning [20]. In selected cases, surgery may be delayed, the operative plan modified, or nonoperative alternatives reconsidered. Risk estimates should therefore be interpreted in relation to the predicted outcome, model calibration, surgical urgency, institutional resources, and patient preferences, rather than as fixed decision thresholds. The level of clinical deployment varies by tool and use case. Available preoperative AI models range from research-stage risk calculators and EHR-based algorithms to externally validated models and prospectively implemented health-system tools. Some disease-specific AI applications have progressed further toward routine clinical use. For example, the Optellum Virtual Nodule Clinic is a commercially available AI radiomics-based computer-aided diagnosis tool that estimates malignancy risk in pulmonary nodules from computed tomography (CT) imaging. Its primary role is to support diagnostic risk stratification and downstream management decisions, including surveillance, biopsy, and referral for specialist evaluation [21]. More broadly, AI currently serves as a decision-support resource in preoperative patient selection and risk stratification by facilitating the early identification of high-risk patients, informing preoperative optimization and resource planning, and supporting shared decision-making regarding operative and nonoperative management. Treatment strategy selection Once surgical eligibility is established, the multidisciplinary team must determine the optimal treatment sequence, including upfront surgery, neoadjuvant therapy, treatment intensification, or organ preservation strategies. These decisions have a direct impact on long-term oncologic outcomes; however, accurately predicting tumor response prior to treatment remains inherently challenging. Recent advances in radiomics and DL have enabled imaging-based prediction of treatment response before surgery. In locally advanced rectal cancer, post-chemoradiotherapy magnetic resonance imaging (MRI)-based radiomics models have demonstrated the ability to predict pathological complete response (pCR), which may help identify candidates for organ-preserving strategies such as watch-and-wait [22]. In esophageal squamous cell carcinoma, pretreatment CT-based DL radiomics models have been developed to predict major pathological response to neoadjuvant chemoimmunotherapy [23]. However, these models remain largely investigational, derived mainly from retrospective, disease-specific cohorts, with limited prospective evidence that their integration into clinical workflows improves treatment selection or patient outcomes. In contrast, more clinically mature examples have begun to emerge in urologic oncology. In localized prostate cancer, a multimodal AI biomarker integrating digitized biopsy histopathology with baseline clinical variables was developed and validated using multiple NRG Oncology phase III randomized controlled trials of radiotherapy-based treatment, with or without systemic therapy [24]. Subsequent work translated this biomarker into a clinically usable framework that provided more refined risk stratification than conventional National Comprehensive Cancer Network risk groups [25]. Although this evidence derives from radiotherapy rather than surgical cohorts and therefore does not constitute direct surgical decision-support evidence, it illustrates a more mature direction for AI in oncologic treatment selection: the translation of multimodal prognostic modeling into clinically interpretable risk stratification that may support individualized treatment planning. Surgical extent and approach planning Determination of surgical extent is a preoperative decision aimed at achieving adequate oncologic resection while minimizing injury to normal tissue and preserving organ function. Surgeons integrate tumor location, the likelihood of local invasion, spatial relationships with major vessels and organs, and patient-specific anatomical variation to define the resection margins and operative approach. In this domain, AI supports decision-making through two main mechanisms. First, imaging-based predictive models can estimate the probability of lymph node metastasis or disease spread, thereby informing the extent of resection. For example, MRI-based radiomics models for predicting axillary lymph node involvement in breast cancer may inform the scope of nodal dissection, although most remain at the prototype-to-translational stage [26]. Second, AI-enabled three-dimensional (3D) anatomical reconstruction provides patient-specific visualization to support decisions regarding surgical extent and approach. By delineating the spatial relationships among tumors, vessels, bronchi, and adjacent organs, these tools can be directly incorporated into operative planning. In thoracic surgery, an AI-based 3D system (InferOperate) has been reported to improve the identification of bronchial and pulmonary vascular variations, the accuracy of procedure selection, and planning efficiency compared with conventional 2D CT-based planning [27]. In addition, 3D-assisted planning increased surgeons’ confidence and interobserver consistency. However, these evaluations primarily assessed planning accuracy and efficiency rather than patient-level outcomes. As an illustrative example outside core general surgery, patient-specific 3D virtual-reality models changed operative plans in approximately 30% of robot-assisted partial nephrectomy cases; plan modification was associated with shorter operative time, lower estimated blood loss, and shorter clamp time [28]. Whether improvements in planning accuracy and decision consistency translate into fewer complications, better functional preservation, or improved survival remains to be established prospectively. POSTOPERATIVE ARTIFICIAL INTELLIGENCE-ASSISTED DECISION SUPPORT In the postoperative phase, AI-based decision support extends beyond discrete, one-time decisions toward trajectory-based management driven by time-series clinical data. Whereas preoperative decision-making is centered on static choices under conditions of limited diagnostic information, postoperative decision-making involves continuous analysis of incoming clinical data to enable real-time adjustment of management strategies and optimization of patient outcomes. This section examines the clinical role of AI-based tools across three key postoperative decision points: monitoring intensity, discharge readiness, and surveillance planning. Postoperative monitoring intensity Adjustment of postoperative monitoring intensity is intended to facilitate early recognition of clinical deterioration and timely escalation of care. Conventional early warning scores, such as the Modified Early Warning Score and National Early Warning Score (NEWS), apply fixed thresholds to snapshots of a limited set of physiological variables. This approach may generate nonactionable alerts in some patients while failing to capture clinically important changes in physiological trajectories over time [29,30]. Dynamic monitoring approaches have been developed to address this limitation by incorporating temporal changes in physiological measurements rather than relying solely on static observations. In postoperative cardiac surgical patients, the Dynamic Individual Vital Sign Trajectory Early Warning Score incorporated recent physiological trends, variability, and observation frequency and outperformed NEWS in predicting cardiac arrest, unplanned ICU readmission, or in-hospital death within 24 hours [30]. These findings highlight the potential value of continuously updated risk assessment for postoperative monitoring. Building on this concept, AI-based early warning systems extend dynamic monitoring by integrating longitudinal EHR data—including vital signs, laboratory results, medications, procedures, and clinical documentation—to generate continuously updated individualized risk estimates. Evidence from clinically deployed systems is derived largely from mixed medical–surgical or general inpatient populations rather than exclusively postoperative cohorts. Targeted Real-time Early Warning System (TREWS), an EHR-integrated ML system for sepsis detection, has been prospectively implemented across multiple hospitals, and timely provider evaluation of alerts was associated with earlier antibiotic treatment and improved clinical outcomes [31]. However, deployment does not necessarily imply reliable performance. In a large multicenter comparison, the Epic Deterioration Index showed lower discrimination, positive predictive value, and lead time than several alternative early warning scores [32]. A meta-analysis of prospectively evaluated AI-based systems reported associations with lower mortality and shorter hospital stay, whereas effects on ICU transfer remained uncertain [33]. Key implementation challenges therefore include false-alert burden, variable clinician response, and uncertainty regarding downstream clinical benefit [29, 30, 32, 33]. Because patient populations, workflows, and data-generating processes change over time, deployed models also require continuous evaluation for dataset shift, performance degradation, unintended consequences, and patient-level outcomes [34]. Future studies should determine whether AI-guided monitoring improves timely escalation of care, postoperative complication rates, ICU utilization, and mortality specifically in postoperative populations. Discharge readiness and care transition Discharge timing requires balancing recovery, safety, and institutional capacity: premature discharge may shift unresolved clinical problems outside the hospital, whereas unnecessary hospitalization exposes patients to hospital-acquired complications and inefficient resource use. Conventional discharge decisions rest on clinical criteria such as oral intake, pain control, mobility, laboratory stability, and the absence of acute complications. Although essential, these indicators do not always capture the overall trajectory of postoperative recovery or the practical barriers to a safe transition of care. ERAS (Enhanced Recovery After Surgery) pathways provide a clinical framework for evaluating discharge readiness according to procedure-specific recovery milestones and the availability of appropriate post-discharge support rather than length of stay alone. Depending on the operation, relevant domains may include oral intake, pain control, mobility, the absence of acute complications, patient education, and follow-up planning [35,36]. AI-based models may complement this clinical assessment. Safavi et al. [37] developed and validated a feedforward neural-network model that predicted discharge within 24 hours and identified clinical, practice-related, and nonclinical causes of delayed discharge, allowing teams to prioritize likely discharge candidates and address potentially modifiable barriers. Similarly, the Adelaide Score uses routinely collected vital signs and laboratory parameters to estimate the likelihood of discharge within the subsequent 12 and 24 hours after general surgery [38]. In a prospective implementation study, the score was combined with an existing natural language processing system to prioritize patients for review by an interprofessional discharge-support team [39]. The intervention was associated with shorter length of stay and lower 7-day readmission; however, its single-center, historical-control design and bundled workflow preclude attribution of these outcomes to the Adelaide Score alone. Current evidence therefore supports AI as a clinician-facing aid for anticipating discharge and coordinating multidisciplinary care, with medical fitness for discharge remaining the responsibility of treating teams. Prospective implementation evidence is still limited, particularly regarding generalizability across surgical populations and effects on patient safety, continuity of care, and longer-term outcomes. Surveillance and recurrence risk planning Postoperative surveillance aims to detect recurrence while additional treatment remains feasible. Current surveillance schedules are generally determined on the basis of disease-specific guidelines and broad clinicopathological risk categories. However, the risk and timing of recurrence vary according to tumor location, pathological stage, pathological and surgery-related risk factors, and the time elapsed since surgery. In a retrospective cohort study of 2,622 patients with stage I–III colorectal cancer, recurrence after curative resection peaked at approximately 11 months, and the temporal pattern of recurrence differed according to patient-level risk factors [40]. This heterogeneity underscores the need for more personalized, risk-based surveillance strategies. At present, AI has been investigated primarily as a tool for predicting individual recurrence risk and for refining risk stratification. ML models can integrate routinely collected demographic, clinicopathological, immunohistochemical, and laboratory data to predict recurrence risk, whereas DL models can extract prognosis-related features directly from digital pathology images. A web-based explainable ML model predicted 5-year colorectal cancer recurrence with an area under the curve (AUC) of 0.83–0.84 on internal validation of a retrospective cohort [41]. In addition, a weakly supervised DL model based on histopathological whole-slide images was externally validated in patients with nonmetastatic colorectal cancer, achieving an AUC of 0.715 and stratifying patients into high- and low-risk groups with 5-year recurrence-free survival rates of 45.7% and 82.5%, respectively [42]. Current evidence therefore supports AI mainly for baseline recurrence-risk stratification rather than adaptive surveillance. No prospective study identified in this review used AI to modify surveillance intervals or imaging schedules or demonstrated improved recurrence-related outcomes. Future studies should test whether risk-adapted surveillance improves detection of treatable recurrence while reducing unnecessary testing. Evidence and implementation status vary substantially across perioperative AI decision points. Table 2 summarizes representative decision-support examples by evidence/use status and clinical role. The evidence/use status descriptors indicate the highest reported level of validation, workflow integration, or implementation and should not be interpreted as a single ordinal ranking of clinical maturity. DISCUSSION Organizing perioperative AI by the decisions it supports, rather than by algorithm or prediction target, shows that the clinical value of a predictive output depends less on its discrimination than on the decision it enters. Across the six decision points examined, predictive capability has consistently outpaced its integration into surgical decision-making. A decision-centered reframing Prior reviews have grouped surgical AI by algorithmic architecture or prediction target and emphasized cross-study performance comparison [11,13,14]. This orientation obscures features that determine clinical value: whether a model’s output maps onto a defined action, whether it was validated in the population where it will be used, and whether research effort matches clinical need. Reorganizing the literature by decision point makes these features explicit. It shows, for example, that postoperative surveillance research is dominated by static baseline risk stratification [41,42] although recurrence hazard is time-varying [40]—a mismatch that is not apparent when studies are grouped by algorithm. Unlike prior performance-centered reviews, this decision-centered synthesis shows that translational gaps are not uniform across perioperative care but differ according to the characteristics of each clinical decision. Time horizon and reversibility as organizing axes Perioperative decisions differ along two axes—time horizon and reversibility—and the properties that make an AI tool useful shift accordingly. Preoperative decisions (candidacy, treatment sequencing, resection extent) are made once, under incomplete information, and are difficult to reverse [16–19,22–27]; here calibration and interpretability are paramount, because outputs must be weighed against surgical urgency, treatment intent, institutional capability, and patient preference. Acute postoperative monitoring operates over minutes to hours on continuously updated data to support reversible escalation [30–33], where positive predictive value, lead time, and alert actionability matter more than discrimination alone. Discharge support is a day-scale problem integrating recovery trajectory and logistical barriers [37–39], and surveillance unfolds over months to years, presently relying on baseline rather than continuously updated estimates [40–42]. Evaluating all postoperative AI as “real-time” therefore conflates fundamentally different tasks; evaluation metrics should be matched to each decision’s horizon. From predictive performance to clinical utility Clinical maturity varies markedly across decision points (Table 2). Some AI systems have already been incorporated into routine clinical practice or commercial platforms, and others have undergone prospective clinical implementation. Most models, however, remain at the research stage, having been developed from retrospective, single-center studies or disease-specific cohorts. Moreover, clinical deployment does not guarantee clinical benefit. The widely used Epic Deterioration Index underperformed comparator models in a multicenter validation, whereas TREWS improved outcomes only when clinicians acted on its alerts. These findings indicate that clinical benefit requires both adequate model performance and an effective clinician response. The challenge of clinician response ultimately stems from the gap between prediction and clinical action. Although many models report high discrimination, they often fail to link their predictions to specific clinical actions or to prospectively demonstrate that acting on these predictions improves patient outcomes [43]. Because probabilistic predictions permit interpretive variability, the same predicted value may elicit different responses from different clinicians in the absence of explicit thresholds or action criteria, thereby undermining consistency [44]. In contrast, outputs expressed as risk categories or classifications directly linked to specific actions can be applied more consistently. Similarly, frequent alerts arising from low specificity may raise clinicians’ response threshold and delay recognition of clinically important signals. Alert-based AI systems should therefore be evaluated primarily on the basis of positive predictive value and the appropriateness of the clinical response following an alert, rather than sensitivity alone. Ultimately, achieving clinically meaningful benefit depends less on further improving discrimination than on closing the gap between predictive output and clinical action. Criteria for clinical integration Taken together, these considerations suggest that a minimum set of criteria must be met before perioperative AI can be adopted in clinical practice. First, model predictions should be directly linked to a specific clinical action. If a prediction does not alter actual clinical management, its value as a decision-support tool remains limited regardless of how high its accuracy may be. Second, predicted probabilities should be adequately calibrated. It is calibration, rather than discrimination, that determines the safety of threshold-based decision-making, yet calibration remains underreported in most studies. Third, models should be externally validated in patient populations and clinical settings comparable to those in which they are intended to be applied. Models developed at a single institution frequently fail to generalize because of differences in case-mix or healthcare system context. Fourth, reported performance should be carefully examined for potential data leakage, which can produce inflated performance estimates that are not reproducible in real-world clinical practice. Finally, because preoperative predictions are inherently based on incomplete information, AI outputs should serve to support, rather than replace, clinical judgment. In addition, the literature search focused primarily on studies published between 2015 and 2025. Therefore, relevant studies published outside this period may have been missed. Limitations and future directions This review was designed as a purposive narrative review rather than a systematic review, and representative studies were selected to illustrate the proposed framework. Consequently, the selection and interpretation of the included studies may have been influenced by reviewer judgment. Because study selection and classification of the level of evidence were conducted by a single reviewer, the possibility of selection and interpretive bias cannot be excluded. Direct comparisons across studies were further limited by heterogeneity in patient populations, outcome measures, validation methods, and clinical implementation settings. In addition, the literature search focused primarily on studies published between 2015 and 2025. Therefore, relevant studies published outside this period may not have been included. Future reviews should improve the comprehensiveness and reproducibility of the evidence base through systematic literature searches and independent study selection and assessment by multiple reviewers. In perioperative AI research, prospective studies are needed to determine whether AI-supported decision-making improves clinical actions and patient outcomes. Calibration, external validation, and the specific clinical responses intended to follow each model output should also be consistently evaluated and reported. CONCLUSION This review reorganized perioperative AI around the clinical decisions it is intended to support, rather than by algorithmic type or predictive performance. In doing so, it demonstrated that the role of AI and the criteria by which it should be evaluated must vary according to contextual factors, including the completeness of available information, the reversibility of the decision, and the time available for response. Across the six decision points examined, preoperative AI was applied mainly to support risk-informed patient selection, treatment strategy selection, and surgical planning under conditions of incomplete information. In contrast, postoperative AI was used primarily to support time-sensitive patient monitoring, discharge coordination, and recurrence-risk assessment. Clinical adoption was most advanced in acute deterioration monitoring, whereas most AI applications for treatment strategy selection, surgical planning, discharge support, and surveillance remained at the research stage or were supported by only limited prospective evidence. From this perspective, the central challenge for perioperative AI research lies not in further improving algorithmic performance, but in connecting predictive outputs to real-world clinical judgment and decision-making. Future research should therefore focus on how this connection is realized in practice and whether such use leads to meaningful improvements in clinical processes and patient outcomes. Only through this shift can perioperative AI evolve beyond a collection of isolated predictive models into a practical clinical decision-support framework spanning the entire continuum of surgical care.