Global trends

In this study, bibliometric methods were used to conduct a statistical analysis of 1,355 articles from the WoSCC database, and VOSviewer software was utilized for visualization. The analysis aimed to explore the research hotspots and development trends of AI in the field of anesthesia. Curve fitting revealed a rapid growth in the number of publications in this field over the past two decades, with a marked increase observed after 2019, reaching 326 articles by 2023. This trend indicates that AI in anesthesia is a hot research topic. The increase in publications may be attributed to the widespread application of AI in anesthesia, particularly in precision medicine and rapid postoperative recovery. The growing demands on anesthesiologists highlight the importance of comprehensive and in-depth research on AI in anesthesia, suggesting that such research will continue to expand.

Countries

The H-index and total citation count are fundamental indicators for measuring the academic impact and quality of publications (Bastian et al. 2017). According to the network visualization map, the top 10 countries with the most publications include 5 European countries (England, Italy, Germany, the Netherlands, and France), two Asian countries (China and South Korea), two North American countries (the USA and Canada), and one Oceanian country (Australia). The USA leads the field with the highest number of citations (n = 8764) and the highest H-index (44). Although China ranks second in total citations, its average annual citation count is significantly lower than that of other countries. Italy, while accounting for only 5.9% of the total publications, boasts the highest average annual citation count and an outstanding H-index of 20. Based on the publication trends of each country (Fig. 2B) and the overlay visualization of co-authorship analysis (Fig. 3B), China shows the fastest growth in the number of publications and has strengthened its academic collaborations with other countries. However, its relatively low average citation count per publication (n = 6.98) indicates that there is room for improvement in the quality of its research.

Institutions and authors

Stanford University, Harvard Medical School, and the University of Toronto are the three most prolific institutions in AI-related publications in anesthesia. Among the top ten institutions, 4 are from the USA and 4 from China. Based on co-authorship analysis, we observe that institutions from the USA are at the center of the collaboration network. Institutions in North America and Europe exhibit close cooperation; however, most institutions in Asian countries have predominantly internal collaborations (Fig. 4).

Lee, Hyung-Chul, Abbod, Maysam F, and Shieh, Jiann-Shing are the most prolific researchers in AI research in the field of anesthesia. 4 of the top 10 most prolific authors are from China. According to co-authorship analysis, authors from the same country often collaborate closely, but connections between authors from different countries remain low (Fig. 5). Therefore, it is recommended that scholars worldwide overcome academic barriers and strengthen collaboration to advance AI research in anesthesia.

Journals

Journals and co-citation analysis provide researchers with critical insights to guide them in selecting appropriate journals for publication. The co-citation network reveals that Anesthesiology (IF = 9.1, 2023) holds the highest recognition and authority in AI research in the field of anesthesia. Following closely are Anesthesia & Analgesia (IF = 4.6, 2023) and the British Journal of Anaesthesia (IF = 9.1, 2023), both of which rank among the top three in terms of the number of papers published and have notably high TLS in the co-citation visualization network. The top 10 most prolific journals account for 17.9% of the total publications, with only the British Journal of Anaesthesia having an impact factor exceeding 5. These findings suggest that most research in this domain still requires greater recognition by high-impact medical journals.

Current knowledge and hot topicsDOA monitoring and regulation

Numerous studies have found that electroencephalogram (EEG) signals can reflect the effects of drugs on the central nervous system of patients, thereby indicating the patient's state of consciousness. EEG has thus become a mainstream method for monitoring the DOA, with the bispectral index (BIS) being the most widely used. However, BIS has limitations such as lag and susceptibility to interference from electrocautery, which can affect its monitoring efficacy. Researchers have used machine learning techniques to analyze EEG data to more accurately assess DOA and improve real-time monitoring and accuracy. Nguyen-Ky T et al. (2011) proposed a method for constructing a wavelet-based DOA index, using adaptive thresholds to denoise low-frequency and spike noise. Gu Y et al. (2019) combined various EEG features, including frequency domain and entropy features, with neural networks to estimate DOA. Their experimental results showed that this method could effectively distinguish between awake and other anesthesia states. Afshar S et al. (2021) introduced a new hybrid deep learning structure, which includes convolutional neural networks (inspired by the inception module), bidirectional long short-term memory, and an attention layer. This model continuously predicts BIS using EEG signals. Madanu R et al. (2021) combined ensemble empirical mode decomposition with power spectral analysis and a convolutional neural network classifier to predict DOA. This method offers the potential for safer surgical procedures by establishing simpler devices for DOA prediction.

Closed-loop target-controlled delivery system

Closed-loop control of anesthesia delivery can be defined as set-point tracking, where the controller adjusts one or more inputs (manipulated variables) of the system based on feedback from one or more outputs (controlled variables) of the system (Ghita et al. 2020). In 1950, Mayo CW et al. (1950) used cortical activity to automatically titrate ether delivery during abdominal surgery. Since then, various signals have been used to guide the automatic titration of different anesthetic agents in various surgical settings (Mayo et al. 1950). In recent years, the rapid development of machine learning and AI has led to the creation of new anesthesia prediction models. Padmanabhan R et al. (2015) applied reinforcement learning to develop a closed-loop anesthesia controller. Lee HC et al. (Lee et al. 2018) demonstrated that deep learning models outperform traditional models in predicting BIS during target-controlled infusion of drugs in surgical patients. With the advancement of AI, automated feedback control is expected to mitigate the impact of variability among individual patients, optimize anesthesiologists'workload, increase the time spent in more optimal clinical states, and ultimately enhance the safety and quality of anesthesia care (LE Guen et al 2016).

Constructing prediction models

Based on machine learning methods and incorporating multimodal patient data, predictive models for critical adverse events such as hypotension, hypoxemia, postoperative acute kidney injury, and postoperative mortality can be developed. Hatib F et al. (2018) demonstrated that machine learning algorithms trained on high-fidelity arterial pressure waveform datasets can predict arterial hypotension events in surgical patients'physiological data. Lundberg SM et al. (2018) developed a machine learning model that more accurately predicts impending hypoxemia than anesthesiologists and identifies the causes of hypoxemia. Tseng P et al. (2020) utilized machine learning to predict acute kidney injury after cardiac surgery, aiding in postoperative risk assessment. Fritz BA et al. (2019) proposed a novel deep learning algorithm that incorporates preoperative and intraoperative dynamic data to predict 30-day mortality. Additionally, researchers have used AI to predict postoperative cognitive dysfunction (Xie et al. 2023), postoperative nausea and vomiting (Shim et al. 2022), intraoperative hypothermia (Dibiasi et al. 2023), perioperative blood transfusion requirements (Stehrer et al. 2019), and postoperative infections (Chen et al. 2023b). The efficient computational power of AI to process complex data allows for the early prediction of adverse events and timely intervention, ensuring perioperative safety for patients.

Image classification and recognition

Due to its low cost, portability, and real-time imaging capabilities, ultrasound has garnered significant attention from anesthesiologists. Ultrasound-guided nerve blocks, vascular access, and epidural analgesia are now widely used in clinical practice (Bowness et al. 2021; Pesteie et al. 2018; Tian et al. 2022; Viderman et al. 2022). AI-guided solutions can enhance the optimization and interpretation of ultrasound images, as well as visualize needle advancement and local anesthetic injection (Viderman et al. 2022). This assistive technology can be used to facilitate target recognition (e.g., peripheral nerves and fascial planes) and to identify optimal block sites by displaying relevant landmarks and guiding structures (e.g., bones and muscles) (Bowness et al. 2021). Additionally, AI can rapidly assess cardiac function (Chen et al. 2021), identify and classify difficult airways (Hayasaka et al. 2021; Tavolara et al. 2021), and recognize and classify pain (Bargshady et al. 2020). AI in medical image analysis is currently a hot research topic. AI technologies can successfully interpret anatomical structures in real-time ultrasound imaging and assist young anesthesiologists in practicing ultrasound-guided nerve blocks and puncture techniques.

Pain management

Postoperative pain is challenging to predict due to the multitude of influencing factors. Patient-related factors (such as age, gender, genetic characteristics, comorbid medical and psychological conditions) and surgical factors (such as surgeon, surgery type, surgical site, and anesthetic techniques) can all lead to severe acute or chronic postoperative pain (Rashidi et al. 2019). In acute pain research, big data is utilized to evaluate postoperative pain outcomes, opioid use, and the effectiveness of multimodal pain management strategies (Muller-Wirtz and Volk 2021). Shi G et al. (2023) developed a machine learning model to predict moderate to severe acute postoperative pain in orthopedic patients under general anesthesia by identifying risk factors. Nair AA et al. (2020) demonstrated that machine learning models could predict postoperative opioid needs in outpatient surgery patients, potentially improving the management of their acute postoperative pain. Another common application of AI in pain medicine is predicting the progression of chronic pain (Abd-Elsayed et al. 2024). Sun C et al. (2023) utilized machine learning methods to establish a predictive model for chronic postoperative pain in breast cancer patients, aiding in identifying high-risk individuals and improving postoperative management. Zhao Y et al. (2020) evaluated a deep learning framework for chronic pain assessment, using a large number of sensors on patients to evaluate chronic pain. The application of machine learning in telemedicine pain management enables physicians to make effective, real-time, data-driven decisions (Cascella et al. 2022).

Limitations

Through bibliometric analysis and literature visualization, this study provides insights into the development trends and research hotspots in the field, serving as a reference for researchers and clinicians in related fields. However, there are some limitations to this study. First, the bibliometric analysis is based on data retrieved from the WoSCC database, which only includes English publications. These factors result in selection bias by overlooking other databases, such as PubMed and Scopus, and potentially excluding important publications published in other languages. Therefore, future research should address this limitation by expanding the range of literature databases and language restrictions included in the search. Second, several biases may affect the results, such as publication bias. In addition, in the early years, AI manuscripts were often more easily accepted and published if the topic was novel. Lastly, the evaluation of an article's impact relies solely on citation counts. In reality, a more comprehensive evaluation should incorporate time factors due to differences in publication years. The earlier an article is published, the more citations it may receive. And more recent publications tend to be cited less frequently, so their impact may be underestimated.