Artificial Intelligence for Perioperative Medicine: Perioperative Intelligence 2023

Since the first public demonstration of anesthesia for surgery in 1846, we have made tremendous strides in improving anesthesia delivery and safety. Globally, millions of simple or complex surgeries are safely performed to optimize human health and function. Currently, postoperative mortality is far more common than intraoperative mortality, and if considered a disease category, it will be the third leading cause of death just after heart disease and cancer, presenting a huge opportunity for improvement. The key question is: can anesthesiologists help reduce perioperative morbidity and mortality? The anesthesiologist’s role has expanded beyond the operating room (OR), and anesthesiologist-led care teams can deliver coordinated care that spans the entire surgical experience, from the decision to have surgery to discharge and long-term recovery. Therefore, perioperative medicine is an evolving specialty focused on delivering coordinated and effective care for the surgical patients.

With aging populations, the number of patients who qualify for surgical treatment has increased, as well as the complexity of surgical procedures. But limited health care resources necessitate the use of innovative solutions in perioperative care focused on improving patient outcomes and reducing cost. According to the Institute of Medicine, we should strive to deliver care that is safe, effective, patient centered, timely, efficient, and equitable. Nonetheless, huge variability exists in many domains among anesthesia care providers affecting quality of care and risking efficiency, cost, and patient outcomes. Some of the variability is due to patient-specific needs, but most are due to knowledge gaps or subjective variation among clinicians, and can be categorized as low-quality care.

Data-driven approaches and artificial intelligence (AI) can help deliver high-quality care and reduce both patient-specific and clinician-specific variability. Using one of the first principles devised by Aristotle which is to think like a scientist until a basic assumption cannot be deduced any further, we need to critically evaluate the elementary components of anesthesia care, which is to achieve analgesia, hypnosis, and muscle relaxation while maintaining oxygen delivery, hemodynamics, and vital organ function. We need to accurately assess patient risk and tailor evidence-based care for an individual patient throughout the spectrum of perioperative medicine. Also, accurate assessment and maintenance of measurable parameters, such as anesthetic depth, blood glucose, hemoglobin, electrolytes, and body temperature, are crucial for our success, and specific sensors need to be built (eg, tissue oxygen monitoring or anesthetic drug level, which can accurately describe patient physiology and guide optimal interventions). Consequently, the amount and complexity of perioperative data will increase, necessitating near real-time processing for critical clinical decision-making. AI will be instrumental to make sense of this information and build decision support tools.

AI: HOW DOES IT WORK?

AI is the science of building computer systems that can mimic human intelligence. It is also a group of diverse computational techniques such as machine learning (ML), deep learning (DL), and natural language processing (NLP; Figure 1). ML techniques, unsupervised or supervised, teach patterns in a large amount of data, and can be used to build classification and predictive models. Supervised learning models (eg, logistic regression and decision trees) utilize labeled data to learn, whereas unsupervised models (eg, principal component analysis, k-means) draw inferences from patterns and associations in the data. However, in reinforcement learning, the model learns from the actions of the agent in its environment through a reward signal. Using ML on arterial pressure waveforms, Hatib et al predicted intraoperative hypotension up to 15 minutes in advance with a sensitivity and specificity of 88% (85%–90%) and 87% (85%–90%; area under the curve [AUC], 0.95 [0.94–0.95]). DL is a subset of ML that utilizes multiple layers of connected neural networks, like the human brain, to progressively extract higher-level features from the raw input. The multiple layers of networks are connected by backward and forward propagation of the data based on weights and biases to complete identification or predictive task. Ghorbani et al used image recognition and deep learning to accurately identify cardiac structures in echocardiography images; specifically, they were able to identify the presence of pacemaker leads, enlarged left atrium, left ventricular hypertrophy, left ventricular end systolic and diastolic volumes, and ejection fraction with good accuracy. Another AI technique, NLP, is used specifically to understand spoken or written content. For example, Xu et al developed a multimodal ML model using clinician notes and associated structured data to accurately predict International Classification of Diseases, 10th edition (ICD-10) diagnosis codes related to patients, a complex task commonly done by experienced coders. Large amounts of learning data sets are key for development of all AI models. Modern clinical practice is ripe for AI applications because of availability of complex structured or unstructured data from multiple sources, such as numerical data from monitors, text from electronic health records, and imaging data.

Using vast information, and with expert input, advanced AI application can help develop autonomous systems or robots that help in drug delivery, precision mechanical tasks, and decision support systems. Autonomous systems are ever more important for patient safety, especially with an aging workforce. For example, drug delivery closed-loop systems comprising sensors to monitor safe drug level, algorithms to assess needed change, and drug delivery systems to deliver drugs to patients can provide consistent anesthetic drug delivery, and finally sensors to monitor drug effect (Table). AI can also help change subjective assessment, a huge source of clinical practice variability, to an objective assessment. Pain assessment is one such clinical examination that is notoriously difficult and subjective, especially under anesthesia. For example, objective nociception assessment is now possible using electrocardiogram, plethysmogram, and skin conductance, and it provides nociception level index, ranging from 0 (absence of noxious stimulation) to 100 (severe noxious stimulation). If the level is high above the pain threshold, opioid drug dose will be requested. In fact, multiple closed-loop systems, each for a specific task, can be made to work together for consistent anesthesia delivery. Airway examination is another highly operator-dependent assessment that varies in terms of precise replication. Continuous progress is now being made using computerized analysis of facial structure to determine the degree of airway difficulty, in which complex neural networks will be used across thousands of facial features. Also, we know that sicker patients are more likely to experience worse outcomes and risk stratification can help tailor anesthesia care to an individual patient, reducing patient-specific variability. Thus, AI-based point-of-care decision support systems can support evidence-based clinical practice and improve patient care.

PERIOPERATIVE INTELLIGENCE: AI IN PERIOPERATIVE MEDICINE

Perioperative intelligence provides a framework for developing useful AI application for perioperative medicine (Figure 2). Our focus should be on 3 key areas: (1) identifying at-risk patients, (2) early detection of complications, and (3) timely and effective treatment.

Identification of High-Risk Patients: Predictive Analytics

Predictive analytics are the most common AI applications in health care. Various supervised and unsupervised ML models are used to predict binary events, such as readmission and mortality. This is logical because administrative data used for these models are widely available, and because hospitals are incentivized to reduce readmission and mortality, for example, risk stratification index (RSI) American College of Surgeons National Surgical Quality Program (ACS NSQIP) revised cardiac risk index (RCRI) and PreOperative Score to predict PostOperative Mortality (POSPUM) These models typically use regression analysis techniques and may not be regarded as real-time or dynamic risk prediction techniques. Other predictive models have used ML techniques for estimation of postoperative discharge destination to the floor within 24 hours of surgery. Real-time and complete patient information is necessary to improve predictive accuracy of these models.

Early Detection of Complications: Role of Sensors and Continuous Monitoring

Postoperative complications, such as myocardial injury after surgery, acute kidney injury (AKI), and postoperative opioid-induced respiratory depression and infections, are a leading cause of morbidity and also increased costs. Also, early-stage treatment may decrease the probability of bad outcomes, emphasizing the role of early detection. Novel sensors and continuous monitoring can help collect large amounts of physiological or electronic health record data, and AI can help build predictive algorithms and decision support systems. For example, perioperative hypotension is strongly associated with organ system injury. With continuous noninvasive arterial pressure monitoring enabled by novel noninvasive sensors, we can build an algorithm that can predict hypotensive events, alerting clinicians to intervene, thus reducing the incidence of hypotension. Similarly, up to one-third of patients experience AKI after cardiac surgery. A gradient-boosted tree classifier-based ML model using continuous high-fidelity monitoring of intra-abdominal pressure, urine output, and core temperature accurately predicted stage 2 AKI 24 minutes before the first appearance of Kidney Disease: Improving Global Outcomes (KDIGO) threshold criteria. Here again, early correction of perfusion pressure using a combination of vasopressors, volume, and diuresis may reduce adverse outcomes. Opioid-induced respiratory depression is common on the postoperative general care floor. Better, portable, continuous monitoring using wearable technology has now enabled early detection of respiratory depression episodes. Because these episodes happen ahead of actual code-blue events by a stretch of time, early detection and correction may offer an opportunity to avoid catastrophic outcomes. Scores such as Prediction of Opioid-Induced Respiratory Depression in Patients Monitored by capnoGraphY (PRODIGY) are the first step in prediction of the risk of respiratory depression using multivariable regression modeling on continuous oximetry and capnography data, and the next step is pattern detection with DL techniques.

Timely and Effective Treatment: Decision Support Systems

Frequently, even intervention based on high-quality evidence is not delivered in routine clinical care. A decision support system can assimilate patient information and high-quality evidence to generate point-of-care guidance. For example, Joosten et al³³ used 3 closed-loop systems for precise titration of anesthesia, analgesia, and fluid, and showed favorable impact on neurocognitive recovery. In the absence of high-quality evidence, not only can AI help build guidelines and recommendations of various professional societies, it can also help build decision support systems based on guidelines.

APPLICATIONS OF AI IN SURGERY

The growth opportunities of AI in surgery study predict that by 2024, the AI market for surgery will reach $225.4 million, up from $69.1 million in 2019. The use of AI and predictive analytics in conventional ORs will help hospitals address inefficiencies and clinical challenges physicians face when performing surgery with decision support and image-based navigational tools. Patients would be end-beneficiaries of these solutions that better help the surgeons perform their job. Some of the AI solutions can help determine the risk of complications even before a patient is wheeled into the OR so that doctors can preempt them and ensure smoother surgeries and faster recovery. Fewer complications, readmissions, or need for corrective surgeries and earlier recoveries will ultimately drive down the cost of health care. The use of AI and predictive analytics will have a multitude of downstream effects, including enhanced patient experience, increased provider satisfaction and engagement, improved outcomes, and reduced cost.

AI also is already helping to identify areas to target for quality improvement. One example is the OR “black box” platform system, which records and analyzes everything that occurs in surgery, which can reveal potential problems. For example, 1 hospital using the OR black box learned that the OR doors were being opened too often during surgery. Subsequent discussion with OR leadership revealed that the suture cart had been relocated to outside the room, so it was returned to its original location within the OR. ORs are very complex environments in which digitalization of the OR environment, the digital information coming from the different information systems, electronic equipment, and sensors can be used to develop an AI system that can understand the surgical processes. For example, the Triton system uses AI and infrared camera technology to analyze photos of sponges taken by an iPad in an OR or delivery room to quantify blood loss.

APPLICATIONS OF AI IN RESEARCH

Evidence-based medicine is enabled by high-quality evidence generated through well-powered randomized trials. However, trials are expensive and time-consuming, and only a fraction (16%) of perioperative medicine is guided by high-quality randomized clinical trial evidence. AI can help in all areas of research, ranging from novel trial design and analytics to patient recruitment strategies. For example, recruitment optimization and alternating intervention trials with the use of automated electronic health records information, help with efficient recruitment of large number of patients. Similarly, ML techniques can help develop quick insights from large amounts of complex physiological data, which can help with evidence generation. Several automated patient-screening systems using AI-based techniques have been used in the emergency room to assess patient eligibility for clinical trial recruitment. The choice of surgery for patients with epilepsy using trained physician notes to appropriately identify candidates has been examined. AI-based chatbots have been used for cancer trial screening and reporting by Google.

APPLICATIONS OF AI IN QUALITY IMPROVEMENT

The problem of low-quality care achieved prominence with the release of “Crossing the Quality Chasm” in 2001 by the Institute of Medicine; however, progress has been slow. Fifteen years later, a seminal paper suggested that medical error was the third leading cause of death in the United States, indicating that over time, little progress was made in improving health care quality. During the perioperative period, patients are exposed to a variety of therapeutic interventions of different complexities, but despite the advancement in medical care, complications are common and, at times, deadly. During the perioperative period, clinicians gather information to assess baseline patient condition and associated risk by obtaining a detailed history, physical examination, and investigation. Clinicians then generate a specific treatment and monitoring plan to achieve the best outcome. However, most of the decision-making is based on static knowledge acquired through previous experience. Enormous amounts of readily available health care data from electronic health records have made AI techniques a very attractive proposition to help with clinical decision-making and improve quality and safety. Not only can AI help with routine decision-making, it can also proactively identify potential harm. For example, it is common to see duplication of tests, ordering of unnecessary tests or prescriptions, and delivery of a less-than-optimal treatment plan. Therefore, AI can help deliver better quality of care by: (1) delivering pertinent information about surgical patients to providers at critical decision points, (2) assisting with development of a personalized care path based on patients’ medical condition and needs, and (3) monitoring compliance to evidence-based practice of medicine.

Locating information in a large volume of health care data is like finding a needle in a haystack, and even current AI technologies are not perfect. Techniques such as NLP can potentially help extract pertinent information from medical health records during the perioperative period and present it in a concise, explainable, and actionable format. Mathis et al showed that an ML-based algorithm developed from preoperative and intraoperative features can detect heart failure in early stages, possibly allowing initiation of confirmatory testing and treatment.

AI can help discover subtle differences in surgical populations and uncover practices associated with the best and worst outcomes, giving clinicians the ability to design the most optimal perioperative pathways for patients based on patient baseline characteristics, surgical procedure, and trajectory of recovery. Maheshwari et al demonstrated that AI-powered applications can very easily identify differences in colorectal surgical patients and identify desirable interventions, which are associated with better patient outcomes and lower hospital cost.

During the perioperative period, patients often develop complications, the majority of which can potentially be avoided if interventions were implemented in a timely fashion. Identification of the signal in the electronic health record, which forewarns about impending complications, can allow clinicians to take action to mitigate undesirable outcomes. Lundberg et al⁵³ demonstrated that an AI algorithm could predict intraoperative hypoxia 5 minutes before it occurred. The algorithm also identified important predictors, thereby helping physicians make appropriate management plans.

Providing feedback to individual perioperative clinicians can result in measurable improvements in patient care and guideline compliance. It has been suggested that feedback to clinicians is most effective when it comes from a credible and validated source, is ongoing and close to real time with clear targets, and when there is a scope to improve. ML techniques provide individualized patient-level feedback to providers, pointing out actionable drivers of performance. Using a neural network model, Schulz et al predicted postanesthesia care unit (PACU) and length of stay (LOS) based on variables outside of the anesthetist’s control to better appreciate the LOS variation that may be under the individual anesthetist’s control and potentially modifiable.

APPLICATIONS OF AI IN EDUCATION

AI applications can enhance education content development, improve interaction between teachers and learners, and help with grading and evaluation. For example, AI is revolutionizing education by empowering students with targeted courses based on student needs and skills. Another example is Hellothinkster a math tutoring program that uses AI to track the steps a student takes to solve a math problem and guide them with alternate approaches to solve it. The approach to identify the student’s knowledge gap and tailor subsequent lessons in the deficient area can be used in anesthesia training. Similarly, Content Technologies develops AI that creates customized educational content. DL analyzes existing course materials, and the technology creates custom learning materials, chapter summaries, and student tests. Even grading can be automated; for example, Gradescope helps grade all assessments whether online or in-class, and provides a clear picture of how students are doing. Anesthesia training prompts high-quality feedback on resident performance that can help improve resident training; however, not all faculty feedback is high quality. Neves et al screened faculty feedback using an ML model to ascertain high-quality versus low-quality feedback, which, in turn, can improve feedback provision. Finally, we need to introduce AI competency in anesthesia trainees. Radiology training programs are taking a lead in designing and implementing focused data science pathways for radiology residents.

LIMITATION AND BARRIERS FOR THE ADOPTION OF AI

Despite enormous potential of AI to enhance health care delivery, there are substantial barriers to its universal adoption. There is significant anxiety in the health care community to implement AI systems without proper validation and explainability. It is hard to imagine that major treatment decisions will ever be based on black box AI systems, which lack reasonable clinical explanations and accountability. Lack of generalizability of AI solutions is another concern because many algorithms are trained and tested on a specific, narrow data set and may not perform well on different populations. Therefore, there is an urgent need for development of large and robust clinical data sets, which will allow development and testing of AI algorithms to ensure generalizability and validity. Not only do we need to focus on the quantity of data, but also on the quality of data captured. For example, perioperative physiological waveform data quality is difficult to maintain, threatening algorithm outputs. Unfortunately, due to regulatory barriers, it will be difficult to develop publicly available data sets that are big enough and contain a wide array of clinical data. Despite some successes of AI applications in health care, especially image recognition, there is a great deal of skepticism for AI implementation in medicine. Ethical and legal ramifications of decisions based on AI algorithms are just getting attention and need appropriate regulations. Furthermore, underrepresentation of certain populations in the data sets used for AI model training can lead to inherent biases, adversely affecting health care delivery and outcomes. For example, with a limited data set, AI may deliver superior predictive performance, but it could also compound inequities (“algorithmic bias”).

Even if a generalizable AI solution becomes available, implementation in the clinical workflow could present a huge challenge. Clinicians at the bedside need to be asked “What should be done differently with the knowledge derived from these models?” Implementation science should guide and evaluate the impact of novel AI solutions on clinical workflow. Some of the decision-making will be automatic, and there is concern about skill degradation. We need to answer the question of whether this skill degradation presents opportunity to acquire new skills, or whether it is a real risk for patient care.

Current institutional structure of learning, data sharing, and collaboration may hamper AI development and deployment. Data ownership, patient versus institutional, further limits free access to much-needed data. We believe that patients should have full ownership of health data, and this change is inevitable. Success of AI in health care is dependent on collaborative work among institutions caring for diverse communities. Departmental structures should change to promote acquisition, retention, and partnership with skilled AI data scientists to build useful clinical applications.

Finally, a strict regulatory environment is a key barrier for adoption of AI, which differs by country and institution. For example, the United States is a laggard in physiological closed-loop technology adoption due to strict regulatory or political reasons. However, there are signs of change. Recently, the US Food and Drug Administration ran workshops and provided framework for regulatory considerations for physiological closed-loop medical devices used in anesthesia and critical care. Also, many of the algorithms are now considered software as medical device (SaMD) to promote safe innovation and to protect patient safety.

RECOMMENDATIONS

A systematic approach is required to realize the benefits of AI in perioperative medicine. The opportunity is huge, but the barriers to adoption are also not trivial. We recommend the following changes in specific areas of perioperative medicine to realize the full potential of AI and improve patient care.

1. Education: Introduce AI and data science education for medical students and residents. At a minimum, clinicians need to have a better understanding of the terminologies and techniques used for AI applications and understand how to evaluate the growing volume of scientific literature. Liu et al described the basic knowledge required to understand an ML paper. The American College of Graduate Medical Education (ACGME)-accredited organizations such as the American Board of Artificial Intelligence in Medicine are providing pathways toward training and certification for clinicians of all backgrounds (https://abaim.org/).
2. Collaboration: Fostering collaboration among clinicians and data scientists is extremely important to develop solutions that will support the needs of clinicians to serve their patients. Aboab et al proposed a “datathon” or “hackathon” model in which participants with disparate but potentially synergistic and complementary knowledge and skills effectively combine to address questions faced by clinicians. Given limited resources and a paucity of skilled AI engineers in health care, organizational structure needs to change with focus on problem-solving, collaboration, and innovation.
3. Data: Despite increasing terabytes of data in health care being generated, access to data continues to be a major hurdle for development, validation, and generalization of AI tools. Professional societies and organizations need to foster open-source availability of health care data to promote future research and development of AI tools as demonstrated through the use of Medical Information Mart for Intensive Care (MIMIC) and Radiological Society of North America (RSNA) data sets.
4. Transparent algorithm development and validation: External multicenter validation is currently lacking for most AI tools, limiting generalizability and acceptance.
5. Implementation: Considering lessons learned from physician burnout from electronic health records, it is important to consider translational research to be a fundamental part of implementation of AI in workflows. In a pilot trial, Maheshwari et al failed to show hypotension reduction using a validated AI decision support solution, mostly because clinicians ignored half of the alerts.
6. Regulatory changes: Government and nongovernment organizations need to reform policies to promote safe data sharing and the development of effective AI tools.