Artificial Intelligence in Perioperative Care: Opportunities and Challenges

Before anesthetic care, anesthesia providers must integrate multiple data streams to formulate a plan to be discussed with the patient or family. Currently, the specific perioperative risks and benefits may be grounded in large population studies risk score calculators or theoretical frameworks of similar patients with the presenting comorbidities and undergoing similar operations or anesthetics. Artificial intelligence algorithms can enhance the preoperative risk stratification process by automating data collection to assess a patient’s risk factors and predict individualized surgical outcomes, such as complications recovery trajectory mortality and length of hospital stay.  These models can help anesthesia providers and surgeons weigh the risks associated with a particular procedure and make informed decisions about patient selection and surgical planning.

After reaching an informed decision to proceed with surgery, navigating the logistics and experience of surgery can be daunting for patients. Artificial intelligence–powered chatbots and mobile applications can help guide patients through the preoperative process, providing appointment scheduling, reminders, preoperative health assessments, and answers to common questions. Immersive artificial intelligence–based virtual reality/augmented reality technology could additionally be leveraged to build applications that enable patients to virtually explore the surgical process, become acquainted with anesthesia procedures, and learn about pain management techniques.  This approach has the potential to alleviate anxiety, improve patient engagement, and enhance informed decision-making.

Artificial intelligence methods are particularly adept at producing prediction models by leveraging the underlying correlations from multiple sources of data such as clinical observation data (e.g., notes, physical exam, nursing assessments), monitoring data (e.g., hemodynamic waveforms), diagnostic data (e.g., labs, molecular studies, imaging studies), and patient-reported data. These data streams can be combined to produce prediction models to improve anesthetic care and patient outcomes (fig. 1).  The successful implementation of models has the potential to reduce the risk of complications and assist the anesthesia and surgical teams in making informed decisions. 

A unique area of event prediction modeling in anesthesiology is in real-time intraoperative monitoring. Due to the dynamic nature of the operating room, these models must incorporate constantly changing time-series data from hemodynamic monitoring, the ventilator, intraoperative events, and the clinical record to make accurate and rapid predictions to facilitate timely intervention. Commercially available medical devices are increasingly incorporating features that utilize artificial intelligence algorithms for real-time monitoring.  Additional ongoing research in this area has used anesthetic records to predict events such as hemodynamic instability depth of sedation and need for intraoperative transfusion for the current surgery.  There remains great potential for implementation of these models in addition to further research in predicting intraoperative events from hemodynamic, neurophysiological, and respiratory data.

Clinical decision support systems have been a major area of application for artificial intelligence tools to aid physicians in navigating complex decisions to provide personalized care to patients in an expedited manner. For example, these systems can aid in surgical decision-making dosage calculations treatment planning, and literature review.  In terms of surgical decision-making and treatment planning, artificial intelligence models have been explored for predicting surgical candidacy tracking the phases of surgical procedures and predicting complications.  Many existing clinical decision support systems focused on specific applications such as error reduction and patient safety have also been successful, for example for drug adverse reactions and medication dosages. 

In the operating room, clinical decision support systems can integrate multimodal data to further assist anesthesia providers by prompting tasks, analyzing images titrating medication and predicting critical events. Advancements in artificial intelligence have also enabled the emergence of automated closed-loop and semiautomated delivery systems for control of parameters such as the level of sedation blood pressure and ventilation. However, synthesizing these data from the electronic health record in an automated fashion can be challenging due to data fidelity and heterogeneity, and such systems raise new regulatory concerns such as safety, liability, and accountability. 

The advent of generative artificial intelligence including large language models such as ChatGPT (OpenAI, Inc., USA)  and those trained on clinical expertise such as Med-PaLM (Med-Pathways Language Model, Alphabet Inc., USA) have the potential to greatly augment clinical practice. Large language models are trained on vast quantities of natural language to learn statistical relationships between words to mimic human understanding of language, and these language representations can then be used to construct generative models to produce novel output. Such models would enable anesthesia providers to rapidly access and synthesize the wealth of available literature to guide decision-making.  They could also be used to increase the efficiency of chart review and charting, improve practice guidelines, and provide real-time predictions or recommendations to guide care delivery.  Although these models are powerful, they are still subject to bias, are prone to producing fictitious information (also known as “hallucinations”) and to unintentional memorization of training data, and are not infallible. Furthermore, results provided by generative artificial intelligence models are unvalidated, and usage of publicly available models carries the risk of disclosure of proprietary information, such as protected health information or trade secrets. Thus, expert human clinical judgment will continue to be necessary to navigate the intricacies of patient care. It is important that healthcare providers understand the fundamentals and limitations of these models to successfully leverage their use in the clinical setting.

Anesthesia providers are instrumental in numerous resource-intensive environments that require coordination and optimization to ensure efficient use of hospital resources. Perioperative care is a complex endeavor involving an interacting network of people, technology, and processes. Coordination of this complex system for optimal performance could be enhanced with artificial intelligence, for example to more accurately predict case duration hospital length of stay and case scheduling.  There exist multiple ongoing efforts in private industry and in academic research in collaboration with hospitals to test new artificial intelligence–based systems to improve operating room scheduling to increase output and efficiency, which hold future promise for implementation in the community and academic setting.

The practice of anesthesia is dependent on the availability of safe, effective, and reliable drugs and devices. Developing novel drugs is traditionally a time-intensive and costly process, involving screening large libraries for potential targets, pairing targets with compounds with desirable profiles, development of novel assays, in vitro and in vivo testing, and clinical trials. Multiple novel artificial intelligence–driven methods have been developed recently to expedite these steps. For example, methods have been developed in other domains for drug design and drug target screening pharmacokinetic and pharmacodynamic modeling and drug repurposing.  These methods can discover novel drug targets and indications through innovative approaches, for example using natural language processing on a combination of clinical data, published literature, and chemical repository data.  In addition to drug discovery and drug repurposing, artificial intelligence can provide greater insight into characteristics of currently used drugs to improve personalized delivery of anesthetic care.  One such example would be incorporating pharmacogenetics in the usage of anesthetic and analgesic agents. 

Devices permeate the practice of anesthesia, aiding in drug delivery, monitoring, airway management, pain management, and simulation. Many of the devices in current practice already utilize machine learning algorithms, such as arrhythmia detection, automatic ejection fraction calculation using point-of-care ultrasound, and anatomical structure identification for peripheral regional blocks. Building upon these examples, there is ample opportunity for the development of novel artificial intelligence–driven devices to enhance anesthetic delivery, minimize adverse events, and improve patient outcomes.

The incorporation of new artificial intelligence tools into clinical practice requires specific skills and knowledge to help construct, validate, test, deploy, and iterate new technologies and innovations (fig. 2). Successful development and implementation of any artificial intelligence approach will require an interdisciplinary team with a diverse range of expertise and the integration of varied perspectives from the multiple stakeholders affecting or affected by the artificial intelligence tools.

For anesthesia groups that are interested in developing new technologies, professionals specifically trained in data science will be essential to establish the systems and networks that allow for data collection, storage, and analysis. They will also be critical in designing, developing, and refining the algorithms that drive artificial intelligence tools, and they ensure the algorithms are accurate, efficient, validated, and able to learn and adapt over time. Professionals with experience in software integration are needed to link artificial intelligence tools with existing systems, ensuring seamless interoperability. Human factors engineers will also be vital to help inform developers on how to optimize systems for efficiency, effectiveness, trustworthiness, satisfaction, and usability by anesthesia providers, not all of whom may be familiar with the technical aspects of artificial intelligence. Additionally, data science specialists with expertise in both artificial intelligence and perioperative care can assist in upskilling general data literacy and act as a bridge between research, development, and clinical teams, accelerating the cycle of knowledge sharing.

Current anesthesia providers in all practice settings will be faced with choosing and implementing these novel technologies as they become more prevalent in our clinical practice. The adoption of artificial intelligence tools by the spectrum of anesthesia practices, from community to academic, is essential for continued acceptance and advancement of artificial intelligence technologies.

Although it is still unclear exactly how artificial intelligence will filter into bedside clinical practice, there appear to be several thought models emerging. One approach is that artificial intelligence applications will be developed by private industry for specific clinical applications, in collaboration with or commissioned by anesthesia practices. After initial development, additional anesthesia practices could license and customize such software to meet the needs of their individual practice, similar to the different versions of common electronic health record systems. Although beneficial for the specific target application, such artificial intelligence may not move forward agilely, because this approach has the limitation of potentially being tailored to multiple different proprietary systems.

Another approach is the development of clinical artificial intelligence tools for on-site use at vanguard centers (largely academic institutions with integrated clinical and informatics services). These tools would be vetted in these centers and then adopted by healthcare systems interested in implementation rather than development. These systems would need to be designed with a focus on applicability and usability in multiple independent practice settings to ensure safe and effective adoption in different populations.

Regardless of the development approach, continuous feedback from users in all practice settings will be important for the advancement of adoption of artificial intelligence in the clinical setting. Early implementers and adopters are crucial for prioritizing targets for artificial intelligence use, identifying value areas, and shaping the roadmap for technology-enabled workflows. The perspectives of stakeholders who are resistant to the adoption of artificial intelligence are also vitally important to the process. By providing constructive criticism and inspiring transparent dialogue, such voices can help to ensure that artificial intelligence initiatives maintain focused on improvements in patient-centered care. Development of strategies to educate and establish trust regarding new artificial intelligence technologies will be necessary to encourage adoption.

The development and implementation of new artificial intelligence–based algorithms in anesthesiology generates several ethical issues that must be carefully considered by an objective governance body. An objective artificial intelligence governance body is vital to ensure that these tools do not inadvertently perpetuate or exacerbate health disparities, are accessible to all patients, and are continuously assessed for changes in risk versus benefit.

Successful oversight will require a multidisciplinary approach that includes clinicians, data scientists, information technology specialists, human factors engineers, implementation scientists, ethicists, and regulatory experts to develop artificial intelligence initiatives that meet clinical need; can be seamlessly integrated into clinical workflows; comply with regulatory requirements; and are thoroughly tested to ensure their accuracy, generalizability, and safety. The governance body should objectively evaluate how a model was constructed, assess whether active steps were taken to mitigate inherent biases, monitor model validation and testing to avoid phenomena such as calibration or prediction drift and evaluate whether the risk–benefit landscape of a model needs to meet a standard of transparency and explainability.  Collaborations with other healthcare organizations and academic institutions to share knowledge and best practices related to artificial intelligence in healthcare may help develop strategies to achieve these goals.

In summary, the stakeholder landscape and expertise required for the implementation of an artificial intelligence initiative in perioperative care is broad and diverse. Ensuring active engagement and collaboration among all these stakeholders is essential for the successful integration and optimization of artificial intelligence in clinical practice.

There are several challenges that must be addressed to facilitate the implementation of artificial intelligence in perioperative care.

Introducing artificial intelligence into the clinical practice of anesthesiology, perioperative, and pain medicine requires specific infrastructure to support its development and implementation, which may need to be developed by anesthesiology departments. For anesthesia practices interested in developing novel artificial intelligence methodology, the development of efficient systems that support data collection, management, processing, storage, integration, retrieval, and validation will be essential for extracting knowledge from large quantities of unharmonized data.  The infrastructure should also support standardized data collection protocols, facilitate data annotation, and ensure data quality control measures.  Additional source data validation and auditing may be necessary to ensure high-quality and well-annotated data for training and updating artificial intelligence algorithms.  After implementation, infrastructure should support continuous assessment of model performance and its impact on clinical practice to avoid unintended adverse effects, also referred to as “algorithmovigilance.”  Deployed systems should also enable collaboration and interoperability among different stakeholders, including clinicians, researchers, data scientists, and information technology professionals.

The adoption of artificial intelligence algorithms by end users will also require computational infrastructure to support implementation and monitoring. For many community-based practices, systems for data storage and data processing that are compatible with developed artificial intelligence models may need to be built, purchased from commercial vendors, or obtained from industry partnerships. Although less computationally expensive than developing new algorithms, systems for ensuring data integrity and usability are paramount to ensuring model validity during implementation.

To facilitate integration of artificial intelligence into anesthesia and maximize potential benefits while minimizing potential harms, close clinician involvement with development and use are necessary. However, currently very few clinicians have a practical understanding of artificial intelligence and its limitations or of the related fields of data science, computer science, informatics, and human factors engineering. Positioning clinicians to be able to support artificial intelligence integration into clinical care will require focused educational initiatives to provide clinicians with sufficient understanding of artificial intelligence to propose and evaluate clinical applications to solve clinical needs. 

For those in settings with a formal education mission, such as large academic institutions, there are multiple existing structured educational opportunities, such as grand rounds, lecture series, and workshops, that could be repurposed into training time on artificial intelligence for anesthesia providers—all of which have been effective for clinical teaching of recent anesthesia subareas like regional anesthesia.  Encouraging existing research and quality improvement projects, the development of new research initiatives, and scientific collaborations can accelerate the education and implementation of artificial intelligence.  In terms of formalized education, developing a clinical artificial intelligence fellowship or residency track in the way that clinical informatics has become a fellowship may be beneficial to training new leaders in this domain.

In the private practice or community setting, where anesthesia providers may not have access to expertise in artificial intelligence, publicly available resources such as webinars, video tutorials, and online courses can make education on artificial intelligence in anesthesia accessible to a wider audience.  Similar to other technologies that have been adopted into routine anesthesia practice, such as ultrasonography, an end user should understand the utility, applicability, and limitations of these novel artificial intelligence technologies. Misinterpretation of artificial intelligence predictions or misuse of these technologies could cause inadvertent harm and should be carefully monitored when implementing new artificial intelligence tools. Collaborating with human factors engineers during artificial intelligence development can help ensure that these systems will be deployed smoothly in diverse practice settings, which will foster further independent usage and establish trust in using artificial intelligence.

As artificial intelligence technologies continue to advance and become more prevalent, several ethical concerns have been raised, such as considerations of equity and biases, job displacement, de-skilling, and accountability in the context of autonomous decision-making.  Artificial intelligence systems in healthcare must adhere to ethical and regulatory standards. As an example, the Food and Drug Administration has released a set of guiding principles for good machine learning practice and device development. Artificial intelligence methods must also comply with data privacy and security regulations, such as Health Insurance Portability and Accountability Act (HIPAA) in the United States or General Data Protection Regulation (GDPR) in the European Union. The usage of personal health information for model development and in artificial intelligence applications poses unique challenges, as users may unintentionally use protected health information as input in unprotected artificial intelligence models, and as models based on large quantities of data may reveal correlations that inadvertently generate identifiable data from deidentified data (e.g., genomic data, phenotypic data).  Considering these risks, artificial intelligence models will need to enable transparent, explainable and interpretable models to address ethical concerns and facilitate regulatory approval. Thus, close collaboration between clinicians, data scientists, human factors engineers, ethicists, and regulatory bodies is crucial to ensure the safe and effective implementation of artificial intelligence in anesthesiology, perioperative, and pain medicine.

Anesthesiology is a data-rich field that holds great potential for the application of artificial intelligence to significantly improve clinical care and patient outcomes. However, several challenges must be addressed to facilitate its implementation, including standardizing data systems and workflows, assembling necessary expertise, motivating integration and adoption, instituting surveillance and validation, and establishing regulatory and ethical governance. Overcoming these challenges will require a dedicated multidisciplinary and collaborative effort across departments and institutions. With success, artificial intelligence tools have the opportunity to transform clinical practice, enhancing the delivery of safe, effective, and personalized perioperative care.