Abstract
Artificial intelligence (AI) and machine learning (ML) techniques are rapidly advancing in anesthesiology, showing promise in patient monitoring, outcome prediction, clinical decision support, and automated drug delivery. However, a substantial gap remains between algorithmic capability and practical implementation at the bedside. This narrative review examines the current state of AI/ML applications in anesthesia, including predictive analytics, closed-loop control systems, AI-assisted imaging, workflow optimization, and anesthesia planning, and explores the translational barriers that have limited routine clinical adoption. We discuss technical, organizational, regulatory, and cultural challenges impeding translation, including data quality issues, EHR interoperability constraints, lack of outcome-oriented clinical evidence, business model uncertainty, interpretability concerns, alarm fatigue, and regulatory ambiguity. Strategies to close this gap are proposed, including rigorous prospective validation, interdisciplinary collaboration with industry and payers, post-deployment model surveillance, training data transparency, user-centered design, and implementation science principles. Ethical and legal considerations, encompassing algorithmic bias, accountability for autonomous AI recommendations, privacy beyond de-identification, and equitable access, are also reviewed. A conceptual framework, summary table of applications, and practical implementation checklist are provided. Bridging the translational divide is essential for AI to fulfill its potential in improving anesthesia care, and will require coordinated action from clinicians, researchers, technologists, regulators, and healthcare institutions.
Keywords: Artificial intelligence, Machine learning, Anesthesiology, Clinical decision support, Technology translation, Closed-loop drug delivery, Alarm fatigue, Post-market surveillance
Introduction
Artificial intelligence (AI) has emerged as a powerful force in medicine, capable of analyzing vast datasets and uncovering complex patterns beyond human recognition [1]. In anesthesiology, AI-driven systems have been developed to assist in nearly every phase of perioperative care, from predicting adverse events to automating drug delivery. Despite rapid advances and numerous proof-of-concept studies, a persistent gap remains between AI-based research and its translation into clinical anesthesia practice [2]. Many algorithms demonstrate impressive accuracy in retrospective analyses or simulations, yet few have been implemented as standard tools in the operating room (OR) or intensive care unit (ICU). This translational gap has become a focal point for researchers and clinicians alike [2].
As used in this review, AI and machine learning (ML) refer to computational systems that learn from data to identify patterns, make predictions, or generate clinical recommendations, as distinguished from traditional rule-based algorithms or physiologically derived indices. Closed-loop control systems built on classical control engineering (e.g., proportional–integral–derivative controllers) are discussed where they inform the broader translational picture, but the distinction from ML-based approaches is noted where it matters. The “prediction-to-practice gap” serves as shorthand for the divide between technical capability and bedside adoption, encompassing AI functions beyond prediction, including pattern recognition, risk stratification, decision support, and autonomous or semi-autonomous action.
The promise of AI in improving anesthetic monitoring and patient outcomes will only be realized if we can bridge this divide. This review provides an overview of current AI applications in anesthesia, examines the barriers preventing widespread clinical adoption, and proposes strategies to move from prediction to practice. We also address the ethical and legal dimensions that must accompany responsible deployment.
Methods
We conducted a narrative literature review, searching PubMed, Embase, and IEEE Xplore using combinations of terms including “artificial intelligence,” “machine learning,” “anesthesia,” “anesthesiology,” “closed-loop,” “predictive analytics,” and “perioperative.” The primary focus was on publications from 2015 to 2025, with seminal earlier works included to provide historical context. Reference lists of key articles and recent reviews were hand-searched. This is not a systematic review; our aim was comprehensive and representative coverage of the translational literature rather than exhaustive enumeration of all published studies.
AI in anesthesiology: current landscape
AI applications in anesthesiology span a broad range of domains, reflecting the specialty’s data-rich and time-sensitive nature. A 2020 scoping review identified six major categories of AI research in anesthesia, namely (a) depth of anesthesia monitoring, (b) control of anesthetic delivery, (c) event and risk prediction, (d) ultrasound-guided procedure assistance, (e) pain management, and (f) operating room logistics [3]. A seventh category is beginning to emerge. AI-assisted anesthesia planning, in which large language models and other generative AI systems are being explored for preoperative assessment, anesthetic plan generation, documentation, and clinical note summarization, though published performance data in anesthesiology remain preliminary. Since 2019, the field has seen exponential growth in publications, particularly in areas like real-time hemodynamic risk prediction, AI-assisted ultrasound for regional blocks, and intelligent monitoring systems [1]. Table 1 summarizes the current state of development across the established categories.
Table 1.
AI Applications in Anesthesiology: Categories, Examples, and Adoption Status
| Category | Representative examples | Selected performance/findings | Outcome type | Clinical adoption status |
|---|---|---|---|---|
| Hemodynamic prediction | HPI (Acumen); ML mortality models (Lee 2018) | HPI reduced hypotension duration in HYPE trial; DNN AUROC 0.88 for mortality | Process/Patient | Limited commercial use; most remain investigational |
| Depth of anesthesia monitoring | BIS-based ML classifiers; EEG pattern recognition | Comparable accuracy to commercial BIS in target range maintenance | Process | Research prototypes; BIS itself widely adopted but ML overlay is experimental |
| Closed-loop drug delivery | End-tidal agent control (Dräger Zeus, GE Aisys); BIS-targeted propofol controllers | End-tidal control in routine use; BIS controllers maintain target with less drug use | Process/Patient | End-tidal agent in routine clinical use. IV anesthetic approved in select markets (CN, IN); experimental elsewhere |
| AI-assisted imaging | ScanNav Anatomy (ultrasound guidance); AI nerve identification | 93.5% correct structure identification; experts rated reduced trauma risk | Process/Patient | ScanNav received FDA marketing approval (2022); growing adoption |
| Risk & event prediction | AKI prediction; delirium scoring; difficult airway models | Varied; most report AUROC 0.75–0.90 in retrospective validation | Process | Research stage; no routine clinical deployment |
| OR logistics & scheduling | Case duration prediction; resource optimization; NLP decision support | 15–30% improvement in scheduling accuracy reported | System | Pilot implementations at select centers |
Predictive analytics have been at the forefront of AI research in anesthesia. Machine learning models using preoperative and intraoperative data can forecast outcomes such as postoperative mortality and intraoperative hypotension with notable accuracy [2]. Lee et al. (2018) developed a deep neural network that predicted in-hospital post-surgery mortality better than conventional risk indices [4]. Kendale et al. (2018) demonstrated that supervised learning could identify patients at risk of hypotension after induction, outperforming traditional regression models [5]. AI-driven risk scores have also been explored for complications such as acute kidney injury and postoperative delirium [1, 6], aiming to alert clinicians early.
Closed-loop control systems represent another major area of development, though the distinction between systems built on classical control engineering and those incorporating ML matters. End-tidal anesthetic agent closed-loop control has been in routine clinical use for decades, implemented commercially in machines such as the Dräger Zeus and, more recently, GE’s Aisys anesthesia workstation. These systems use physiological feedback loops governed by classical control principles, not machine learning. More experimental closed-loop systems target intravenous anesthetic agents, using the Bispectral Index (BIS) as a control variable. Studies as early as 2000 demonstrated feasibility of maintaining BIS within target range via automated propofol dosing [7], and Struys et al. (2001) found an automated propofol-BIS controller kept patients at target depth with less drug use than standard practice [8]. Subsequent work on muscle relaxation control yielded comparable findings [9]. At least two regulatorily approved closed-loop intravenous systems exist in clinical practice in China and India, though the extent of their routine use remains unclear. Newer systems leverage reinforcement learning and adaptive algorithms to manage multiple parameters simultaneously [10], but these remain largely investigational. The distinction between classical control and ML-based approaches matters, because the translational barriers for each differ in character and severity.
AI-assisted medical imaging and monitoring represents a domain where clinical translation is further advanced. In regional anesthesia, computer vision algorithms can interpret ultrasound images to identify anatomical structures for nerve blocks. The ScanNav Anatomy Peripheral Nerve Block device, which received FDA marketing approval in 2022, uses deep learning to overlay colored highlights on live ultrasound, helping clinicians visualize nerves and other landmarks in real time [11]. In a validation study, this AI system correctly identified target structures in approximately 93.5% of cases, with low false-negative and false-positive rates [12]. Experts judged that the technology could reduce the risk of needle trauma and block failure in the majority of procedures [12]. This illustrates how AI can augment less experienced anesthesia providers in performing ultrasound-guided blocks, potentially expanding access to regional anesthesia. AI-driven analysis of vital sign trends is also being used to detect patterns that precede hemodynamic events. An early warning system that continuously analyzes arterial pressure waveforms can forewarn of hypotension 5–15 min before it occurs [13], allowing anesthesia providers to intervene proactively rather than reacting to an established blood pressure drop.
Perioperative logistics and decision support have also benefited from AI innovations. Machine learning models have been developed to optimize operating room schedules, predict surgical case durations, and flag patients who might benefit from specific interventions. Natural language processing is being applied to anesthesia records and literature to provide context-specific decision support [14]. These applications underscore that AI’s role in anesthesiology extends beyond direct patient monitoring to workflow efficiency and knowledge management.
The range of AI applications in anesthesia encompasses monitoring, drug delivery, prediction, imaging, pain management, planning, and logistics. Across these domains, multiple studies have demonstrated that AI can match or exceed clinician performance on specific, well-defined tasks under controlled conditions [1, 3]. Yet translating these successes into everyday clinical practice has proven challenging, for reasons that the next sections examine in detail.
The prediction-to-practice gap
Despite the plethora of promising results described above, the reality is that very few AI tools are in routine use by anesthesia providers today. The FDA’s publicly available database of AI/ML-enabled medical devices shows that anesthesiology accounts for a small fraction of approved tools relative to specialties such as radiology and cardiology, with only nine devices categorized under anesthesiology as of mid-2024 out of hundreds of total AI/ML-approved medical devices [15]. There is a clear discrepancy between what has been demonstrated in studies and what is available at the bedside [16]. Anesthesia providers largely continue to rely on traditional vital signs and clinical judgment, not algorithmic predictions, during surgery. Advanced closed-loop systems for intravenous agents remain confined to investigational settings outside of a few international markets.
A case in point is the Hypotension Prediction Index (HPI) algorithm. This ML model was trained on arterial waveform features to forecast hypotension before it occurs. In the HYPE randomized trial, use of the HPI early warning system significantly reduced the duration and severity of intraoperative hypotension compared to standard care [17]. One might expect such results to spur rapid adoption. However, uptake has been gradual. Many clinicians remain hesitant, citing concerns about false alarms, the need for additional monitors, and uncertainty about whether preventing brief hypotensive episodes improves meaningful patient outcomes [13]. Some analyses have also questioned whether the reported performance of the HPI system reflects actual clinical performance, given concerns about biased sampling in validation studies [13]. No hemodynamic prediction algorithm, moreover, can anticipate hypotension caused by acute surgical events, such as a surgeon accidentally injuring a major vessel, for instance, which limits the scenarios in which predictive systems provide actionable advance warning.
The translational lag is worth placing in historical context. The gap between technical capability and routine clinical adoption is not unique to AI. Ultrasound-guided regional anesthesia took more than a decade to move from published feasibility to widespread training and practice. Target-controlled infusion systems, well established in Europe and Asia, have never achieved routine adoption in the United States. Closed-loop ventilation modes were available long before most clinicians used them routinely. The pattern suggests that some barriers to AI adoption, including institutional inertia, training deficits, lack of reimbursement incentives, reflect the general challenge of technology diffusion in anesthesia practice. What distinguishes AI, however, is the additional burden of model opacity, the distributional nature of ML failures, and the pace at which the technology itself evolves.
That pace introduces a distinct challenge. The cycle time for AI/ML development (where new model architectures, training approaches, and capabilities emerge on a near-weekly basis) is fundamentally mismatched with the timeline required for clinical validation, regulatory review, and workflow integration, which spans years. By the time a specific AI tool has completed prospective trials and regulatory submission, the underlying technology may already have been superseded. The mismatch contributes to institutional hesitancy to invest in any single system and raises difficult questions about how to validate a technology category that is continuously in flux.
The translational resistance facing AI also raises a question about the consistency of the evidence standards being applied. Pulse oximetry was adopted universally into anesthesia practice without randomized trial evidence demonstrating mortality benefit; its clinical value was considered self-evident from the physiological information it provided. The evidentiary bar for AI appears materially higher. Several factors may explain this asymmetry. Pulse oximetry offered a transparent, physiologically intuitive measurement, while AI models operate through learned associations that clinicians cannot directly inspect. The failure modes of a pulse oximeter are predictable and bounded, while the failure modes of an ML system may be opaque and context-dependent. The differential threshold of acceptance is worth acknowledging explicitly, because it shapes which translational strategies are likely to succeed and which barriers are specific to AI rather than to medical technology adoption in general.
The reasons for this prediction-to-practice gap are complex. Some are inherent to the technology and data, while others relate to organizational, regulatory, economic, and cultural dimensions. Understanding these challenges is a prerequisite to overcoming them.
Barriers to clinical translation of AI
Multiple barriers have hindered the adoption of AI tools in anesthesia practice. These challenges range from technical issues with algorithms and data, to practical concerns of workflow integration and infrastructure, to broader regulatory, economic, and cultural resistance.
Data quality and generalizability
AI models are only as good as the data they are trained on. In anesthesiology, many published AI models have been developed and validated on single-center retrospective data, despite the existence of large multi-institutional registries such as the Multicenter Perioperative Outcomes Group (MPOG) that contain granular data from tens of millions of cases. Access barriers, data use agreements, and the technical demands of working with registry-scale data may explain why many researchers have not leveraged these resources. Models trained on homogeneous populations can perform poorly when deployed elsewhere due to differences in patient demographics, surgical practices, or monitoring equipment. Data biases (for example, under-representation of certain ethnic groups or comorbidities) can lead to biased algorithms [18].
The challenge of rare events deserves particular emphasis. Clinically critical but uncommon occurrences (malignant hyperthermia, anaphylaxis, catastrophic hemorrhage) are underrepresented in training distributions. The problem is more insidious than mere omission. Standard ML optimization objectives (such as gradient descent with least-mean-square cost functions) systematically deprioritize performance on rare cases because the aggregate cost of misclassifying them is small relative to common cases. Special care must be taken during model training, for instance through oversampling of rare events, cost-sensitive learning, or stratified evaluation, to ensure that AI systems perform adequately on the cases where accurate prediction matters most. Ensuring external validation and robustness across hospitals remains a significant challenge.
Lack of clinical validation and outcome-oriented evidence
Many AI solutions in anesthesia have not undergone rigorous prospective clinical trials. Promising retrospective accuracy does not always translate into real-world clinical impact. Practitioners and hospital decision-makers are reluctant to adopt new technology without clear evidence of benefit in the clinical setting. Three tiers of outcome evidence can be distinguished. Process outcomes such as reduced hypotension duration, patient-centered outcomes such as morbidity and mortality, and system-level outcomes such as OR efficiency and cost reduction. The literature still lacks large multi-center studies demonstrating patient-centered outcome improvement for most AI applications. Until such evidence is generated, AI tools may be seen as experimental rather than essential.
A conceptual point deserves attention here. No monitoring technology or predictive algorithm, however accurate, improves patient outcomes unless it is coupled with an actionable therapeutic intervention. An algorithm that predicts hypotension confers no benefit if clinicians are already managing blood pressure effectively, or if the prediction does not trigger a specific and timely clinical response. The prediction-to-intervention linkage is fundamental to the value proposition of any AI system in anesthesia. As noted in the preceding section, the evidentiary bar for AI tools appears higher than what has historically been required for other perioperative technologies, a discrepancy worth examining as the field defines its validation standards.
Workflow integration, usability, and alarm fatigue
Introducing AI systems into the OR or ICU workflow is operationally demanding. Anesthesia providers already manage multiple monitors and tasks, and an AI tool that adds complexity or requires extra steps can become a hindrance rather than a help. If alerts or recommendations are not seamlessly presented and intuitively actionable, they may be ignored, a phenomenon well documented in the patient safety literature as alarm fatigue. The proliferation of monitoring alarms in the OR is already a recognized threat to vigilance, and poorly designed AI alert systems risk compounding this problem. Integrating AI outputs into anesthesia information management systems (AIMS) and electronic health records (EHRs) presents further challenges. Some AI models may demand computational resources and network infrastructure not readily available in all ORs.
Data infrastructure and interoperability
The difficulty of connecting AI systems to real-time clinical data streams and displaying their outputs to anesthesia providers at the point of care. EHR systems are primarily designed to support billing and regulatory documentation compliance rather than real-time clinical decision support. EHR vendors are typically large, risk-averse organizations that are protective of data access, and since EHR installations are often highly customized from institution to institution, even within the same vendor, it is difficult for third-party AI developers to scale their solutions. Similarly, physiological monitors generate proprietary data streams that may not be easily accessible for real-time AI processing. Without standardized and open data pipelines, even a well-validated AI model may have no practical means of reaching the clinician who needs its output.
Interpretability, trust, and informational asymmetry
Clinicians are trained to understand and trust physiological reasoning. When an AI algorithm provides a prediction or prescription, anesthesia providers naturally ask why. Black-box models that cannot explain their reasoning face skepticism [19]. Anesthesia professionals need confidence that an AI’s advice is sound and will not harm the patient. Lack of transparency in how a model works, especially when it occasionally errs, undermines trust. The concern ties closely to medicolegal questions, because if something goes wrong, the provider is accountable, not the algorithm. Many anesthesia providers are therefore hesitant to delegate critical decisions to an AI unless they understand the rationale and feel in control of the final decision.
An additional dimension of this challenge is that AI systems operate exclusively within the boundaries of their input data streams. An experienced anesthesia provider integrates multimodal sensory information (visual observation of the surgical field, auditory cues from the monitors and the operative environment, tactile feedback, and contextual knowledge about the surgical plan) that current AI systems cannot access. Such informational asymmetry limits what any algorithmic system can achieve and reinforces the argument for human-in-the-loop design. Anesthesia providers will also need to recognize a priori whether a given AI system is likely to perform reliably in a specific clinical context, an expectation analogous to understanding the known limitations of conventional monitoring devices (e.g., non-invasive blood pressure measurement challenges in patients with conically shaped upper arms).
Regulatory and liability hurdles
Bringing an AI device to market for medical use requires navigating regulatory pathways (e.g., 510(k), De Novo classification, or Premarket Approval) that can be lengthy. Regulators demand proof of safety and effectiveness, but current frameworks are better suited to traditional devices than self-learning algorithms. Issues such as how to validate continuously learning systems or handle software updates create uncertainty [20]. Liability is another concern. If an AI-driven pump causes an overdose, the question of who bears responsibility (the clinician, the hospital, or the software manufacturer) has no settled answer. While liability questions also arise for conventional medical devices (a malfunctioning IV pump, for instance), the AI case introduces distinct complexity. For conventional devices, failure modes are generally deterministic and traceable to hardware malfunction or user error. For AI/ML systems, failure may arise from opaque learned associations, distribution shift, or emergent behavior that is difficult to attribute or predict. The legal system’s traditional tools for allocating responsibility (establishing causation and foreseeability) are challenged when the mechanism of failure is not transparent [18]. Malpractice insurers have scant actuarial data on AI-related adverse events, and the difficulty of defining a standard of care when AI recommendations may vary by model version or training data compounds the uncertainty.
Cultural, economic, and business model barriers
The adoption of any innovation can be slowed by cultural resistance. Some anesthesia providers may fear that AI could diminish their role or threaten jobs [21]. Surveys show a significant fraction are concerned about AI’s impact on their career and income [21], though most remain optimistic it will improve care in the long run [22]. If AI enables providers to manage more cases efficiently or extend their capabilities by reducing low-value cognitive labor, the productivity gains could sustain income levels, but this connection is not yet empirically established. Hospitals must also justify the cost of new technology. Purchasing AI-enabled systems is a financial decision; if the return on investment (better outcomes, efficiency) is unclear, hospital administrators may hold off. Beyond the upfront cost of acquisition, AI systems demand ongoing technical infrastructure (data engineers, clinical informaticists, and vendor support contracts) that most anesthesia departments do not currently maintain. These recurring operational costs may disproportionately burden smaller or resource-limited institutions.
Business model viability is a barrier that the academic literature largely overlooks. No technology will reach the point of care unless every key stakeholder along the value chain (developers, distributors, healthcare purchasers, and payers) has a viable economic model. Reimbursement mechanisms for AI-assisted care remain largely undefined, and without them, the commercial incentive to develop, validate, and support AI tools in anesthesiology is limited. Industry partners and payers must be included as essential stakeholders in the translational effort.
Each of these barriers is substantial on its own. Combined, they explain why many anesthesiology AI innovations have not yet crossed the threshold into everyday practice [23]. Figure 1 maps the translational pathway from AI development through clinical adoption, illustrating how each barrier category corresponds to a proposed mitigation strategy. Recognizing these challenges provides a roadmap for what needs to change.
Fig. 1.
Conceptual overview of the prediction-to-practice gap in AI for anesthesiology. Top row: translational pathway from AI/ML development through internal validation, prospective clinical testing, and regulatory approval to clinical adoption, with the translation gap highlighted between regulatory approval and routine adoption. Middle rows: key barriers impeding translation, including both primary barriers (data quality, clinical evidence, workflow integration, interpretability, and regulatory gaps) and secondary barriers (business model uncertainty, EHR/data infrastructure, pace-of-AI mismatch, and cultural resistance). Bottom row: proposed strategies to address each barrier category
Bridging the gap: strategies for implementation
Closing the translation gap will require concerted efforts on multiple fronts. Below we outline key strategies, and Table 2 provides a practical implementation checklist for institutions and research teams.
Table 2.
Practical Implementation Checklist for AI in Anesthesiology
| Step | Phase | Key Considerations | Responsible parties |
|---|---|---|---|
| 1 | Problem Selection | Identify clinical need with measurable endpoints; confirm that AI adds value over existing approaches | Clinicians, data scientists |
| 2 | Stakeholder Engagement | Engage anesthesia providers, nursing, IT, administration, patients, payers, and industry partners | Project leadership, department chairs |
| 3 | Data Assessment | Evaluate data availability, quality, diversity, and access (EHR integration, real-time feeds); address privacy requirements | Data engineers, compliance officers |
| 4 | Model Development | Select appropriate ML architecture; address class imbalance for rare events; build interpretability features | Data scientists, clinical domain experts |
| 5 | Internal Validation | Test on held-out data; external validation at second site; evaluate across demographic subgroups | Research team, biostatisticians |
| 6 | Prospective Clinical Testing | Register trial; test in live clinical environment with defined endpoints (process, patient-centered, system-level outcomes) | Clinical investigators, IRB |
| 7 | Regulatory Submission | Prepare submission per applicable pathway (510(k), De Novo, PMA, or equivalent); include training data transparency per FDA guidance | Regulatory affairs, legal counsel |
| 8 | Workflow Integration | Design user interface with human factors principles; integrate with existing monitors and EHR; conduct usability testing; address alarm fatigue | IT, human factors engineers, end users |
| 9 | Training & Deployment | Train all anesthesia providers (physicians, CRNAs) on appropriate use, limitations, and override procedures; phased rollout | Education leads, department leadership |
| 10 | Post-Deployment Monitoring | Monitor for model drift, track performance metrics, maintain version audit trail, establish failure reporting system, define update approval governance | Clinical informatics, quality assurance, vendor |
| 11 | Performance Review & Decommissioning | Define performance thresholds below which the tool should be suspended or withdrawn; schedule periodic review against clinical endpoints; establish criteria and process for decommissioning underperforming systems | Department leadership, quality assurance, vendor |
Robust clinical validation
There is no substitute for high-quality clinical evidence. AI tools should be tested in prospective, ideally multi-center trials demonstrating improved patient outcomes or workflow efficiency. Expanding trials such as the HYPE study for hypotension prediction to diverse settings can build confidence in real-world utility [17]. Integrating AI decision support into practice and measuring impacts on complication rates or cost can help justify adoption. The more an AI intervention shows tangible benefits under pragmatic conditions, the more likely clinicians and regulators will endorse it. Publishing negative or neutral findings is equally important to temper hype and refine approaches [13]. A culture of real-world validation must become standard for medical AI development.
Interdisciplinary collaboration
To ensure AI solutions address actual clinical needs, collaboration between data scientists and frontline anesthesia providers is essential [1]. Early involvement of clinicians in the design phase can target relevant problems and incorporate domain knowledge, for example determining which predictions would actually change management and what thresholds are clinically meaningful. Involving human-factors experts can improve user interface design. Collaboration should also extend to patients (for perspectives on acceptability), regulators (to align with safety requirements), industry partners (for commercial sustainability), and payers (for reimbursement pathway development). AI tools designed with the correct user population in mind, whether physician anesthesiologists, CRNAs, or anesthesia assistants, will be more effectively adopted [18].
Improving explainability and reliability communication
Building explainable AI (XAI) is a high priority in healthcare. Techniques such as simplified predictive models, feature importance indicators, or rule-based outputs can demystify black-box algorithms. An AI that predicts hypotension might display the top contributing factors (e.g., “trend of decreasing systolic pressure and high heart rate variability”) to justify its warning. Equally important are real-time reliability indicators, that is, mechanisms by which an AI system communicates its confidence level and the conditions under which its performance may degrade, so that anesthesia providers can recognize a priori whether the system is likely to be reliable in a given clinical context. Providing training on how an AI works, and when it does not, can further build appropriate trust [19].
Data quality initiatives
Efforts to gather larger, more diverse, and well-annotated datasets will directly improve AI performance and generalizability. National or international anesthesia data registries, common data standards, and data-sharing collaborations can surmount the limitations of single-center data. Techniques like federated learning, where models are trained on distributed data without centralizing it, could leverage multi-hospital data while respecting privacy. Continuously monitoring AI performance after deployment (post-market surveillance) can identify when models start to drift or fail on new subpopulations, prompting timely updates.
Workflow integration, interoperability, and training
AI solutions should be designed for smooth integration into existing monitoring systems and clinical workflows. This requires attention not only from AI developers but also from the manufacturers of patient monitors, anesthesia machines, and EHR systems, whose platforms must support data exchange. Existing interoperability standards, including ISO/IEEE 11,073 and its specific parts, ASTM F2761, HL7, and FHIR, provide a foundation, but their adoption remains incomplete and inconsistent across institutions. Without broader uptake of open interoperability frameworks, AI solutions will remain difficult to integrate at scale. Usability testing with clinicians can refine how alerts are presented to minimize distraction and alarm fatigue. Anesthesiology training programs and continuing education should incorporate AI literacy, teaching providers how to interpret AI outputs and use them as an adjunct to clinical judgment.
Post-deployment surveillance and governance
The lifecycle of a clinical AI tool does not end at deployment. Robust post-deployment surveillance is essential and includes several components. Model drift monitoring (tracking performance degradation as patient populations, surgical practices, or documentation patterns change over time), version control and audit trails (maintaining records of which model version was active during any clinical encounter), structured failure reporting (analogous to existing adverse event reporting systems), and governance processes for update approval (determining who authorizes model updates, what validation is required before a new version goes live, and how transitions are managed). The FDA’s evolving framework for Software as a Medical Device (SaMD), including the predetermined change control plan concept [20], provides emerging regulatory structure for these processes, but institutional governance must keep pace. The speed at which AI technologies evolve creates a mismatch with traditional endorsement and vetting timelines. If endorsement of a specific tool takes months, the tool may be obsolescent by the time it is approved. Framework-level validation, in which regulatory and institutional review is applied to the development process and validation methodology rather than to each individual model version, may offer a path through this tension.
Implementation science principles
The strategies outlined above can be strengthened by drawing on implementation science, the study of methods to promote the systematic uptake of evidence-based practices into routine care [24]. Key implementation constructs relevant to AI adoption include readiness for change (whether the institution, its leadership, and its front-line staff are prepared for the new technology), fidelity of use (whether the AI tool is used as intended by its developers, or whether workarounds and modifications alter its effectiveness), sustainability (whether the tool continues to deliver benefit over time, or whether engagement wanes after initial enthusiasm), and clinician behavior change (whether providers actually modify their decision-making in response to AI outputs, or ignore them). These constructs provide a framework for diagnosing adoption failures and designing implementation strategies that go beyond the technical performance of the AI model itself.
Through these strategies, the aim is to create conditions under which effective AI innovations can reliably cross the threshold into everyday anesthesiology practice. Early signs of progress are evident. The FDA’s marketing approval of the first AI ultrasound guidance device signals regulatory willingness to approve anesthesia AI tools when evidence supports them [11]. Professional conferences now frequently feature sessions on AI, reflecting growing clinician engagement. By systematically addressing the challenges outlined here, the current trickle of AI into anesthesia can become a steady stream.
Ethical and legal considerations
Any discussion of implementing AI in anesthesia must include the ethical and legal dimensions. The introduction of AI decision support or automation raises important questions about patient safety, privacy, and the role of the clinician.
First is the principle of non-maleficence, that AI systems should do no harm. For AI, this obligation is qualitatively different from device safety in the conventional sense. A stethoscope or a syringe must be safe, but neither acts or recommends independently; the clinician is the agent. An AI system, by contrast, can generate autonomous recommendations or initiate actions that influence patient care independently of the provider’s reasoning. This capacity for agency, however limited, demands an ethical framework that extends beyond passive device safety to active non-maleficence, accounting for the system’s potential to cause harm through its own outputs. Ensuring this requires thorough testing and vigilant monitoring for unintended consequences. An algorithm might inadvertently reinforce a bias, for example under-treating pain in certain ethnic groups if trained on biased data [18]. Such disparities in algorithmic performance across patient subpopulations are not unique to AI; the well-documented racial bias in pulse oximetry, where devices overestimate arterial oxygen saturation in patients with pigmented skin [25], demonstrates that this is a longstanding concern in medical device design. Transparency in how AI models are developed and validated is essential for maintaining trust.
That transparency must extend specifically to the data on which an AI system was trained and tested. The FDA’s guidance on transparency for machine learning-enabled devices calls for disclosure of training dataset composition, demographic representation, and known limitations [15]. Without this information, neither clinicians nor patients can assess whether a given AI tool is likely to perform reliably for a specific patient population. Some ethicists argue that patients (or at least hospital oversight committees) should be informed when AI is being used in their care, especially if it influences significant decisions.
Privacy is another concern. AI often relies on large volumes of patient data. Robust de-identification and cybersecurity measures are needed to protect sensitive information when aggregating data for model training or when AI systems are integrated with medical records. Regulatory compliance with HIPAA is mandatory, but standard de-identification (removing the 18 Safe Harbor identifiers) may be insufficient when high-dimensional clinical data (physiologic waveforms, genetic information, rare diagnoses) enables re-identification through linkage attacks. Good data stewardship requires additional measures including differential privacy techniques, data use agreements with audit provisions, and institutional review of secondary data use for model training [26].
Accountability is a central concern. In anesthesia, adverse events can be life-threatening. If an AI misleads a clinician or malfunctions (for instance, an autonomous system inappropriately increases a drug), how is accountability assigned? The anesthesia provider overseeing care is ultimately responsible for patient safety, whether AI is involved or not. Accordingly, AI tools must be designed to support clinician oversight, and anesthesia providers must be able to detect malfunction or performance degradation, intervene, or override the system at any point during patient care. Comprehensive training and clear protocols should govern AI use, including guidelines on when to accept or reject AI recommendations, and fail-safes if an AI output is clearly erroneous. Legally, institutions might require informed consent for use of certain AI-driven devices, though generally this is covered under consent for standard anesthesia care if the AI is considered a monitoring modality [27].
Ethical practice also extends to ensuring that AI benefits are equitably distributed. We must avoid scenarios where only well-resourced centers have access to advanced AI systems, potentially widening care disparities. As AI becomes commercialized, cost and access need consideration so that it improves outcomes broadly. Involvement of clinicians in AI development also raises potential conflicts of interest that should be transparently managed [28]. Many of these issues are already under active discussion. The American Society of Anesthesiologists has convened panels on AI, and journals have published primers on ethics in anesthesia AI [18]. Successful integration of AI will require not only technical solutions but also ethical frameworks and updated professional guidelines. Anesthesiology has an opportunity to lead responsible AI adoption in medicine by confronting these questions directly rather than deferring them.
Future directions
Several developments already underway are likely to shape the next phase of AI adoption in anesthesiology. Anesthesia machine manufacturers are beginning to integrate AI-assisted features into commercial platforms, following the trajectory of capnography and processed EEG monitoring, both of which transitioned from specialized research tools to standard equipment over one to two decades. The concept of a “learning health system,” in which data from each surgical case feeds back to improve algorithms for subsequent patients, is being explored in multi-institutional collaborations, though the governance and privacy frameworks required for continuous learning remain underdeveloped [1].
Multimodal data integration is an active area of research with concrete early results. Current AI models typically operate on single data streams (arterial waveforms, EEG signals, or ultrasound images in isolation). Combining information from physiologic monitors, laboratory results, medical imaging, and wearable device data could enable more precise and individualized anesthesia management, moving from prediction of single outcomes toward integrated recommendations that balance hemodynamic optimization with organ protection [1].
Large language models (LLMs) and generative AI have introduced a distinct set of capabilities to the field. Early applications include automated clinical documentation, preoperative patient communication, generation of anesthetic plans from structured patient data, and literature synthesis for clinical decision support. While published performance data in anesthesiology remain limited, LLMs differ from prior AI tools in their capacity for natural language interaction, which may lower the barrier to clinician adoption. Their limitations, including hallucination, inconsistent performance on domain-specific reasoning, and the absence of real-time physiological data integration, are well documented and require careful attention before clinical deployment.
Collaborative AI design, in which systems augment rather than replace clinical judgment, offers a model that may resolve several translational tensions simultaneously. An AI that presents relevant data, suggests interventions, and communicates its uncertainty while learning from clinician responses keeps the human in the loop and addresses both interpretability and trust within a single design framework. Progress along these lines will likely require new roles within the anesthesia care team, including specialists in perioperative data science and technicians responsible for maintaining AI systems in the operating room.
Conclusion
The body of research on AI in anesthesiology has grown rapidly, producing systems that can predict hemodynamic instability, guide regional anesthesia, automate drug delivery, and optimize perioperative logistics. The gap between what these systems can demonstrate in controlled settings and what has been adopted at the bedside remains wide. Closing it will require more than technical refinement.
The barriers documented in this review share a common feature. Each involves a mismatch between the assumptions embedded in AI development and the conditions of actual clinical practice. Models are trained on retrospective data but must perform in real time. Algorithms optimize for prediction accuracy but clinicians need actionable interventions. Developers iterate rapidly but regulatory and institutional processes move at the pace of evidence generation. Business models for AI-assisted care remain undefined while the costs of implementation continue to accrue. Addressing these mismatches will require coordinated effort across disciplines, from data scientists and clinical investigators to EHR vendors, device manufacturers, regulators, payers, and front-line anesthesia providers of all training backgrounds.
The strategies outlined here, including rigorous prospective validation, post-deployment surveillance, explainability and reliability communication, interdisciplinary collaboration, and engagement with implementation science, are not aspirational. They describe work that is already beginning at individual institutions and within professional societies. What remains is the harder task of scaling these efforts, building the infrastructure to support them, and sustaining the commitment to honest evaluation of both successes and failures. If anesthesiology can meet that challenge, the field will have demonstrated something valuable beyond any individual AI tool, a model for how a clinical specialty can integrate computational intelligence into patient care without sacrificing the judgment, accountability, and human presence that define the profession.
Acknowledgements
None.
Author contributions
Janardhan Baliga, MD: Conceptualization; Literature review; Drafting of manuscript; Critical revision of content; Final approval of the version to be submitted.Niranjan Seshadri, MD, JD: Conceptualization; Supervision; Thematic framing; Critical review and editing; Policy and governance analysis; Final approval of the version to be submitted.
Funding
Open access funding provided by Amrita Vishwa Vidyapeetham.
Data availability
No datasets were generated or analysed during the current study.
Declarations
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Liu K, Qiu W, Wang X. Exploring the growth and impact of artificial intelligence in anaesthesiology: a bibliometric study from 2004 to 2024. Front Med. 2025;12:1595060. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Singh M, Nath G. Artificial intelligence and anesthesia: A narrative review. Saudi J Anaesth. 2022;16(1):86–93. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Hashimoto DA, Witkowski E, Gao L, Meireles O, Rosman G. Artificial Intelligence in Anesthesiology: Current Techniques, Clinical Applications, and Limitations. Anesthesiology. 2020;132(2):379–94. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Lee CK, Hofer I, Gabel E, Baldi P, Cannesson M. Development and validation of a deep neural network model for prediction of postoperative in hospital mortality. Anesthesiology. 2018;129:649–62. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Kendale S, Kulkarni P, Rosenberg AD, Wang J. Supervised machine-learning predictive analytics for prediction of postinduction hypotension. Anesthesiology. 2018;129:675–88. [DOI] [PubMed] [Google Scholar]
- 6.Hatib F, Jian Z, Buddi S, et al. Machine-learning Algorithm to Predict Hypotension Based on High-fidelity Arterial Pressure Waveform Analysis. Anesthesiology. 2018;129(4):663–74. [DOI] [PubMed] [Google Scholar]
- 7.Morley A, Derrick J, Mainland P. Closed loop control of anesthesia: An assessment of the bispectral index as the target of control. Anesthesia. 2000;55(10):953–9. [DOI] [PubMed] [Google Scholar]
- 8.Struys MM, De Smet T, Versichelen LF, et al. Comparison of closed-loop controlled administration of propofol using Bispectral Index as the controlled variable versus standard practice controlled administration. Anesthesiology. 2001;95:6–17. [DOI] [PubMed] [Google Scholar]
- 9.Eleveld DJ, Proost JH, Wierda JM. Evaluation of a closed-loop muscle relaxation control system. Anesth Analg. 2005;101:758–64. [DOI] [PubMed] [Google Scholar]
- 10.Cai X, Wang X, Zhu Y, et al. Advances in automated anesthesia: a comprehensive review. Anesthesiol Perioper Sci. 2025;3:3. [Google Scholar]
- 11.Larkin HD. FDA Approves Artificial Intelligence Device for Guiding Regional Anesthesia. JAMA. 2022;328(21):2101. [DOI] [PubMed] [Google Scholar]
- 12.Bowness JS, Burckett-St Laurent D, Hernandez N, et al. Assistive artificial intelligence for ultrasound image interpretation in regional anaesthesia: an external validation study. Br J Anaesth. 2023;130(2):217–25. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Michard F, Futier E. Predicting intraoperative hypotension: from hope to hype and back to reality. Br J Anaesth. 2023;131(2):199–201. [DOI] [PubMed] [Google Scholar]
- 14.Bellini V, Russo M, Domenichetti T, et al. Artificial Intelligence in Operating Room Management. J Med Syst. 2024;48(1):19. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.U.S. Food and Drug Administration. Artificial Intelligence and Machine Learning (AI/ML)–Enabled Medical Devices. FDA website. Updated 3-4-2026.
- 16.Lonsdale H, Jalali A, Gálvez JA, Ahumada LM, Simpao AF. Artificial intelligence in anesthesiology: Hype, hope, and hurdles. Anesth Analg. 2020;130:1111–3. [DOI] [PubMed] [Google Scholar]
- 17.Wijnberge M, Geerts BF, Hol L, et al. Effect of a Machine Learning–Derived Early Warning System for Intraoperative Hypotension vs Standard Care on Depth and Duration of Intraoperative Hypotension During Elective Noncardiac Surgery: The HYPE Randomized Clinical Trial. JAMA. 2020;323(11):1052–60. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Cascella M, Tracey MC, Petrucci E, Bignami EG. Exploring Artificial Intelligence in Anesthesia: A Primer on Ethics, and Clinical Applications. Surgeries. 2023;4(2):264–74. [Google Scholar]
- 19.Mathis MR, Kheterpal S, Najarian K. Artificial intelligence for anesthesia: What the practicing clinician needs to know. Anesthesiology. 2018;129:619–22. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.U.S. Food and Drug Administration. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence/Machine Learning (AI/ML)-Enabled Device Software Functions. FDA Guidance. 8-18-2025.
- 21.Alamo CEE, Diatta F, Monsell SE, Lane-Fall MB. Artificial intelligence in anesthetic care: A survey of physician anesthesiologists. Anesth Analg. 2024;138:938–50. [DOI] [PubMed] [Google Scholar]
- 22.Erel S, Kılıç AG. The use of artificial intelligence in anesthesiology: Attitudes and ethical concerns of anesthesiologists. Saudi J Anaesth. 2025;19(4):498–504. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Bellini V, Rafano Carnà E, Russo M et al. Artificial intelligence and anesthesia: a narrative review. Ann Transl Med. 2022;10(9). [DOI] [PMC free article] [PubMed]
- 24.Proctor EK, Landsverk J, Aarons G, et al. Implementation research in mental health services: an emerging science with conceptual, methodological, and training challenges. Adm Policy Ment Health. 2009;36(1):24–34. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Sjoding MW, Dickson RP, Iwashyna TJ, Gay SE, Valley TS. Racial Bias in Pulse Oximetry Measurement. N Engl J Med. 2020;383(25):2477–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Yadav N, Pandey S, Gupta A, et al. Data Privacy in Healthcare: In the Era of Artificial Intelligence. Indian Dermatol Online J. 2023;14(6):788–92. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Maloney JA, Johnson B, Harbell MW. The double-edged sword: artificial intelligence’s promise and perils in anesthesia patient safety. Curr Opin Anaesthesiol. 2025;38(6):776–85. [DOI] [PubMed] [Google Scholar]
- 28.Weiner EB, Dankwa-Mullan I, Nelson WA, Hassanpour S. Ethical challenges and evolving strategies in the integration of artificial intelligence into clinical practice. PLOS Digit Health. 2025;4(4):e0000810. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Bellini V, Rafano Carnà E, Russo M, Di Vincenzo F, Berghenti M, Baciarello M, Bignami E. Artificial intelligence and anesthesia: a narrative review. Ann Transl Med. 2022 May;10(9):528. [DOI] [PMC free article] [PubMed]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Data Availability Statement
No datasets were generated or analysed during the current study.

