Artificial Intelligence-Assisted Monitoring in Anaesthesia: A Systematic Review and Meta-analysis of Diagnostic Accuracy and Clinical Impact AI in Anaesthesia: Monitoring and Outcomes Review

Key Message

1. AI-based perioperative monitoring systems, encompassing machine learning and biomarker-guided methodologies, exhibit enhanced diagnostic precision relative to conventional statistical techniques in various surgical environments.

2. Randomised controlled evidence demonstrates the clinical efficacy of AI-assisted haemodynamic management, resulting in substantial decreases in the occurrence, severity, and duration of intraoperative hypotension.

3. The transition of AI from predictive accuracy to enhanced patient-centered outcomes is dependent on extensive, multicenter randomised trials, external validation, and transparent AI frameworks.

Introduction

Artificial intelligence (AI) has been instrumental in changing the global healthcare scene over the last 20 years, paving the way for new developments in clinical care, diagnosis, individualised treatment, and healthcare data management. Because anaesthesiology requires the real-time interpretation and processing of vast volumes of physiological and pharmacological data, it is a particularly promising field for the adoption of AI-based technologies. Therefore, perioperative safety and the standard of anaesthesia care could be significantly improved by integrating AI-assisted predictive and decision support tools.

The precise administration of fast-acting medications, the timely adaptation to significant physiological changes, and the continuous, high-precision monitoring of multiple vital parameters are all necessary for modern anaesthesiology. However, many clinical decisions are still based on doctors' subjective experiences and established practices rather than standardised, predictive approaches, even with major advancements in medical technology. In this regard, moving from reactive to proactive decision-making models that can foresee crucial events before they materialise clinically is made possible by artificial intelligence and its subfields, such as machine learning (ML), deep learning (DL), and artificial neural networks [1-4].

Predicting intraoperative and postoperative complications, including bradycardia, hypotension, hypoxia, postoperative delirium, postoperative nausea and vomiting (PONV), intensive care admission, and mortality, is a primary objective of artificial intelligence in anaesthesiology. Many of these incidents happen unexpectedly, necessitating quick action to avoid negative outcomes. AI systems analyse preoperative clinical and anamnestic data, as well as real-time multi-parameter signals from automated anaesthetic delivery systems, electroencephalograms (EEGs), and physiological monitors. Finding patterns that are hidden from view, estimating the likelihood of future occurrences, and proposing prompt intervention techniques are the objectives.

Numerous observational and randomised clinical studies have assessed the efficacy of these technologies at different points during the perioperative period during the past ten years. For instance, the Hypotension Prediction Index (HPI) software was created with the help of artificial intelligence (AI) and has been thoroughly investigated for its capacity to anticipate intraoperative hypotension episodes up to several minutes beforehand [5-9]. This enables medical professionals to use fluidic interventions or pharmacological modulation to stop the event from happening. The automated administration of intravenous anaesthetics, including propofol and remifentanil, has also been optimised through the use of AI algorithms. This has improved the depth of anaesthesia and haemodynamic stability while allowing for more customisation than conventional target-controlled infusion (TCI) systems.

More effective control of arterial pressure, depth of anaesthesia, fluid administration, and mechanical ventilation has been made possible by intelligent controllers built on fuzzy logic or neuro- adaptive predictive models [10]. By minimising inter-individual variability in anaesthetic response and maximising the balance between efficacy and safety, these tools have been demonstrated to increase the precision and consistency of therapeutic strategies. Additionally, AI-powered clinical information management tools, like predictive electronic anaesthesia records, have enhanced perioperative data collection and analysis, opening up new avenues for quality improvement and clinical audit.

The current body of scientific literature is disjointed and frequently contradictory, despite the excitement these innovations have generated. Few studies have shown a significant impact on critical clinical outcomes like length of hospital stay, major complication rates, or mortality, despite the fact that many have reported encouraging results in terms of predictive accuracy and improved intraoperative parameters. Moreover, it is challenging to reach firm conclusions regarding the applicability of these technologies in routine clinical practice due to variations in the AI models used, the clinical contexts examined, and the methodological calibre of publications.

In light of this, the goal of this systematic review and meta- analysis is to present a comprehensive summary of the data that is currently available regarding the application of AI in perioperative anaesthesia. Through a comparative analysis of data from observational and randomised clinical trials, the review will assess the safety and effectiveness of AI-based tools for intraoperative monitoring and complication prediction. The results will be stratified by the technology (deep learning (DL), fuzzy logic, neural networks, or machine learning (ML), the type of complication targeted, the operative setting (major versus minor surgery; elective versus urgent), and the methodological quality of the included studies.

The PICOT framework, which is advised for creating structured clinical questions, was used to define the inclusion criteria that were chosen for this review. Adult patients undergoing regional or general anaesthesia make up the target population. The use of artificial intelligence (AI)-based technologies for intraoperative monitoring or perioperative adverse event prediction is the intervention (I). Standard, non-AI-assisted monitoring techniques are used in comparison (C). Predictive accuracy in relation to particular complications (such as hypotension, hypoxia, post- operative vomiting syndrome, delirium, and death) is the primary outcome (O), and clinical variables like length of hospital stay, ICU admission rate, and intraoperative stability are the secondary outcomes. Since 2010 is when AI-based technologies were first introduced and used in anaesthesiology, both prospective observational studies and randomised controlled trials (RCTs) and relevant retrospective studies published since then were included.

This meta-analysis differs from earlier synthesis attempts in that it focusses on studies published after 2010 and has stricter inclusion criteria. Methodological quality, which is evaluated with particular instruments like ROBINS-I for observational studies and RoB 2 for randomised controlled trials (RCTs), is emphasised. Additionally, subgroup analyses will be carried out based on the kind of algorithm employed, the perioperative phase of application (pre-, intra-, or postoperative), and the kind of complication that is being targeted. There are also moral, legal, and financial concerns with the broad use of AI in anaesthesia. Thorough consideration must be given to issues like algorithm transparency, medico-legal liability for mistakes, sensitive data protection, and guaranteeing fair access to these advancements. Training medical staff to enable anaesthetists to consciously incorporate the new technologies into their daily clinical practice is another essential component. Therefore, educational institutions and healthcare facilities ought to support professional development courses that address fundamental AI concepts, data analysis, and human-machine interaction.

Establishing a strong body of evidence to direct the clinical application of AI tools in anaesthesiology and to guide future research aimed at creating dependable, secure, and genuinely practical predictive models for daily use is the ultimate goal of this work. Technical innovation, strong proof of clinical efficacy, safety, and acceptability, as well as long-term economic viability, are necessary for the adoption of such technologies.

Methods

To maintain openness, reproducibility, and methodological rigour, this systematic review with meta-analysis closely adhered to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines. Because the protocol was pre- registered with the PROSPERO (International Prospective Register of Systematic Reviews) database, the authors were bound to adhere to a predetermined plan, reducing the possibility of selective bias and boosting the accuracy of the findings.

Type of Studies Included

Randomised controlled trials (RCTs), retrospective studies and prospective observational studies on the application of AI-based technologies in anaesthesiology that were carried out in adult human populations were included in the review. The recent adoption of AI technologies in clinical practice, which has reduced the quantity of RCTs available, served as the impetus for the decision to incorporate observational studies. This does not, however, rule out the possibility of gathering pertinent data from carefully planned non-randomized research.

Inclusion Criteria

Every study that satisfied the following requirements was deemed qualified:

- To guarantee the inclusion of current research in line with recent advancements in AI technologies, the publication date should be between January 1, 2010, and December 31, 2024.

- Research design: observational studies with a control group and prospective RCTs.

- Adult subjects (≥18 years old) undergoing general or loco- regional anaesthesia in a surgical or procedural setting comprise the population.

- Intervention: the application of models, algorithms, or predictive tools based on artificial intelligence, such as expert systems, machine learning, deep learning, artificial neural networks, or fuzzy logic algorithms.

- The ability to predict or prevent perioperative complications, such as intraoperative hypotension, hypoxia, adverse cardiovascular events, bradycardia, postoperative delirium, PONV, cardiac arrest, and AKI, is the main outcome.

- Long hospital stays, the need to be admitted to an intensive care unit, haemodynamic stability, and the effectiveness of intraoperative physiological monitoring are examples of secondary outcomes.

Exclusion Criteria

Research that satisfied these requirements was not included: - research done on children (less than 18 years old) or in animal models.

- Research that lacked a clear and documented AI-based intervention.

- Simulation studies, case reports, editorials, letters to the editor, and conference abstracts without peer-reviewed full texts are all examples.

- To guarantee linguistic homogeneity in data extraction and interpretation, studies published in languages other than English were disqualified.

Table. Inclusion and Exclusion Criteria

Sources of Information and Search Methodology

Three primary biomedical databases—PubMed/MEDLINE, Embase, and the Cochrane Central Register of Controlled Trials (CENTRAL)—were thoroughly searched. The search approach was predetermined and comprised MeSH (Medical Subject Headings) terms and keywords related to the fields of artificial intelligence, perioperative complications, and anaesthesia. Terms like "artificial intelligence," "machine learning," "deep learning," "anaesthesia," "perioperative complications," "monitoring," "prediction," and others were examples of Boolean combinations. Only human subjects’ studies published between 2010 and May 2025 were included in the results.

Search Strategy

Study Selection Process

Two reviewers (A.A.C., M. R. E.) independently screened all identified abstracts and titles. Using the predetermined criteria, the full texts of potentially eligible articles were retrieved and evaluated for inclusion. A third reviewer (T. R. E.) was consulted or discussed in order to settle disagreements. (Table 1.) (Table 1.)

Data extraction

A standardised collection form that was tested in a pilot phase was used to extract pertinent data from each includedstudy.

This form was used to extract pertinent data from each included study. The variables that were gathered included the author's name, the year of publication, the study design, the sample that was analysed, the population's demographics, the type and characteristics of the AI algorithm that was used, the clinical setting, the type of surgery or anaesthesia that was performed, the primary and secondary outcomes, the main results, and the authors' main conclusions. Two authors independently collected the data in duplicate.

Risk of Bias Assessment

Validated instruments were used to evaluate the included studies' methodological quality:

- RoB 2.0 for RCTs, in accordance with the Cochrane Collaboration's guidelines. Randomisation, treatment deviation, attrition, outcome measurement, and the selection of reported data were among the biases evaluated.

Non-randomized observational studies were conducted using ROBINS-I, which takes into account biases in participant selection, confounding factors, intervention classification, missing data, and analysis techniques.

Two authors conducted each evaluation independently, and any disagreements over the results were discussed. Each study's overall quality was assigned a risk of bias rating of "low," "moderate," or "high."

Statistical Analysis and Data Synthesis

Meta-analysis was used to synthesise quantitative data with homogeneous populations, interventions, and outcomes. To account for variation amongst studies, a random effects model was used. With 95% confidence intervals, effect sizes were presented as odds ratios (OR), risk ratios (RR), or mean differences (MD). Cochran's Q statistic was used to evaluate study heterogeneity, and the I2 index was used to quantify it based on three criteria: 0–25% (low), 25–50% (moderate), and >50% (high).

To investigate the efficacy of AI, subgroup analyses were designed according to the following criteria: the algorithm type (e.g., HPI, fuzzy logic, deep learning, or generic predictive models); the type of complication (e.g., neurological, cardiovascular, or respiratory); and the surgical context (major vs. minor surgery; elective vs. emergency).

Studies with a high risk of bias were excluded through sensitivity analyses.

Affiliation and Peer Review

A multidisciplinary team unaffiliated with academic institutions carried out the work, and during the protocol phase, the review was not exposed to external peer review. However, after the manuscript is finished, it will be submitted for peer review and publication.

Results

There were 18 studies in all, and they were a mix of randomised controlled trials (RCTs), prospective observational studies, and retrospective analyses. The studies took place in many different parts of the world, including Europe (the Netherlands, Italy, Germany, and Spain), North America (the United States), and Asia (Korea and China). This shows that people all over the world are interested in using machine learning (ML) and biomarkers to predict and manage perioperative events. (Table 2.)

The sample sizes were very different. For example, Schenk et al. (2021) had only 60 patients (n=60), while Lee et al. (2022) had 454,404 patients in a very large retrospective dataset [11]. The ORACLE Trial (Fritz et al., 2024) was the biggest interventional study [12]. It randomly assigned 5,071 patients to either ML- assisted or standard perioperative care. (Table 2.)

The Main Results of the Studies Were Different and Included

- Most of the time, intraoperative hypotension (IOH) is defined as a mean arterial pressure (MAP) of less than 65 mmHg (e.g., Schenk et al., 2021; Wijnberge et al., 2020).

- Acute kidney injury (AKI) is often used as a secondary endpoint (for example, Fritz et al., 2024; Ripollés-Melchor et al., 2025).

- Mortality after surgery: both 30 days after surgery and while in the hospital (e.g., Fritz et al., 2019; Lee et al., 2022).

- PIHI (Te et al., 2024) is an example of a composite haemodynamic instability index.

Modelling Methods and Algorithmic Techniques

There were a number of different predictive modelling methods used in the studies. These can be divided into three main groups: traditional ML algorithms, deep learning architectures, and biomarker-based indices (especially HPI).

Some of the models that were tested the most were random forests (RF), logistic regression (LR), gradient boosting machines (like XGBoost), and support vector machines (SVM).

Lee et al. (2022) directly compared several algorithms for predicting 30-day mortality. They found that XGBoost did better than both RF (AUROC = 0.92) and DNN models, with AUROC = 0.942 and AUPRC = 0.175.

Luckscheiter et al. (2022) made models for bad things that happen during surgery. RF got an AUROC of 0.96 and a PRC area of 0.83 [13].

Deep Learning Structures

Deep learning methods took advantage of the changing nature of data streams during surgery.

Fritz et al. (2019) used a combination of convolutional neural networks (CNNs) and LSTM-based architectures (MPCNN) to predict 30-day mortality with an AUROC of 0.867 and an AUPRC of 0.097. Hu et al. (2024) used these ideas on high-frequency arterial waveform data and found that R² = 0.915 and MAPE = 12.25% were good for predicting changes in blood flow during surgery [14].

The Hypotension Prediction Index (HPI)

HPI, a proprietary algorithm trained to predict IOH minutes before onset, was evaluated in a number of RCTs and observational studies (Schenk et al., 2021; Wijnberge et al., 2020; Ripollés- Melchor et al., 2025). In addition to a 77% decrease in the severity and duration of hypotension in the intervention group, Schenk et al. (2021) reported 88% sensitivity and 87% specificity for HPI. When HPI guidance was used, Wijnberge et al. (2020) also showed a significant decrease in cumulative hypotension time.

Innovative and Hybrid Methods

Te et al. (2024) developed the Post-Intubation Hemodynamic Instability (PIHI) index using 14 clinical and ventilatory variables, achieving R² = 0.9047. [15,16] For AKI risk, Abin et al. (2024) used a dual regression approach, reporting accuracy = 80.6%, F1- score = 0.821, and recall = 84.8%.

Predictive Factors and the Significance of Features

MAP and derived waveform parameters were universally critical, according to a comparative analysis of input features (Table 3).

In HPI-based studies, significant attention was paid to variables such as systolic and diastolic blood pressure (SBP and DBP), heart rate (HR), pulse volume (PV), stroke volume variation (SVV), and dP/dtmax (Frassanito et al., 2023; Šribar et al., 2023[16]).

Larger retrospective models included preoperative variables like age, comorbidity indices, baseline renal function, and ASA status (Lee et al., 2022; Kang et al., 2020) [17].

Analysis of Comparative Performance

Among the best-performing models (AUROC > 0.90) were: '

- Lee et al. (2022): AUROC = 0.942, AUPRC = 0.175 (XGBoost).

- Luckscheiter et al. (2022) : PRC = 0.83 (RF), AUROC = 0.96.

- Te et al. (2024)]: R² = 0.9047 for continuous hemodynamic predictions.

Performers in the middle (AUROC 0.80–0.89):

- Fritz et al. (2019) : AUROC = 0.867, AUPRC = 0.097 (MPCNN).

- Fritz et al. (2024, ORACLE) : AUROC = 0.807 for 30-day mortality; 0.766 for AKI.

Models for experiments or lower performers:

- Tan et al. (2021): sensitivity ~90%, specificity ~52%, AUROC = 0.77 (95% CI: 0.62–0.92).

- Morisson et al. (2022) : Brier score = 0.194, AUC = 0.753, penalised logistic regression [18]. (Table 3.)