“Precision Medicine in Anesthesiology”

INTRODUCTION:

As one of the founding fathers of medicine, Hippocrates was prescient in his discernment that patients exhibit differences in the severity of disease symptoms and that some individuals can better cope with their disease compared to others. He believed that it was important to know the person who has the disease in order to tailor the treatment ¹. Hippocrates was one of the earliest physicians to practice precision medicine, where medical decisions and therapies are tailored to the individual patient.

In the modern era, precision medicine accounts for individual variability in genes, lifestyle, and environment that may cause patients to manifest disease and potentially respond to treatments differently². In the 1990s, scientific and technological advances brought the genetics revolution, where studies of the human genome suggested that disease may have a genetic basis³. Since then, advances in genomics and proteomics have provided great insight into the nature and evolution of disease. Simultaneous to the genetics revolution, shifts in health care reform led to the implementation of the 2009 Health Information Technology for Economic and Clinical Health Act, which incentivized the adoption of the electronic health record (EHR). The widespread deployment of EHRs led to the accrual of massive amounts of digitized clinical data, populated from biospecimens, health care visits, administrative claims, and many other sources.

From its inception as a field of acute care medicine, anesthesiology directly practiced precision medicine when the first anesthesiologists administered ether to their patients in unique amounts based on direct observation of clinical effects. Today, clinicians must interpret large amounts of clinical data for critical and timely decisions in the delivery of anesthetic and critical care. As a field, anesthesiology has a history of pioneering patient safety through the use of informatics. The data-rich environment allows anesthesiology to shape the developing field of acute care informatics in a way that may provide advances in patient safety and quality and in reducing healthcare expenditures^4,5. Anesthesiology clinical and research enterprises will require new thinking, training, tools, and a vision for how to utilize and interpret acute care data. This chapter explores how anesthesiology can become the acute care arm of precision medicine by utilizing informatics to address the increasingly complex needs of patients, populations, and organizations.

Computational Anesthesiology: A Paradigm Shift in Anesthesiology

The decade of the EHR led to the development of computational techniques that combined anesthesiology research, informatics, and computer science, leading to the growing field of computational anesthesiology. Computational anesthesiology uses data science approaches to inform measurement-driven high-quality care by assessing adherence to practice guidelines, care utilization, cost, and disparities. Given the increasingly complex nature of data obtained from each patient, deriving insights to guide clinical decisions may be a challenging task for any one person. Artificial intelligence, which is a subfield of computer science that studies algorithms to perform tasks typically associated with cognition, has the potential in assisting care teams in making sense of healthcare data (Table 1). Anesthesiologists need to understand what artificial intelligence (AI) technologies are and how they can be leveraged to derive meaning from EHR data.

Table 1.

Examples of Artificial Intelligence Applications

User Group	Category	Application Examples	Technology
Patients	Health monitoring	Devices and wearables	Machine learning
Healthcare teams	Surgical pre-habilitation	Smartphone app-based programs	Telemedicine Digital health
	Early detection, prediction, and diagnostic tools	Closed loop anesthesia Sepsis detection	Machine learning
Health system	Quality	Need for social “wraparound” service	Machine learning
	Efficiency	Discharge planning assistance	Machine learning

Artificial Intelligence and Machine Learning

AI’s ability to analyze large and disparate types of data and uncover patterns that may not be evident to humans makes it a valuable tool for delivering safer and higher quality care. As one of the major subfields of AI, machine learning learns from data without explicit programming. Data that can be analyzed through machine learning include structured and unstructured forms of data⁶. Most EHR records conform to structured data, where data is organized such that columns represent different variables and rows represent different patients or observations. Unstructured data, such as waveforms, videos, images, physiological time series data, and clinical text documents, convey information that does not have discretely labeled variables that define the information. Machine learning can analyze both structured and unstructured data, performing both classification (dividing data into discrete groups) and regression (modeling data to understand the relationship between two or more continuous variables with the potential for prediction). Supervised, unsupervised, and reinforcement learning approaches can each be used to address problems of classification and regression depending on the type of question and the type of data available⁷.

Algorithms that utilize supervised learning are trained to predict a predefined output, such as recognizing a person in a photograph or video. To develop an ML algorithm with supervised learning, the algorithm needs to be developed on a training dataset to learn associations between input and specified output, and the performance of the algorithm will need to be tested on a test dataset. Machine learning algorithms that use unsupervised learning identify patterns or structures within a dataset. Unsupervised learning will discover relationships in data by grouping similar data points together into more homogenous subgroups. An immediate use for unsupervised learning is in the development of electronic phenotyping, where unsupervised learning algorithms sift through EHR data and identify patient phenotypes that may require unique care. Machine learning algorithms that use reinforcement learning attempt a certain task and learn if it is successful at doing that task through trial and error. Reinforcement learning models map a sequence of situations into a sequence of actions by maximizing a reward signal⁸ and infers the correct actions based on the reward signal since the correct actions are not labeled. Reinforcement learning can be useful in healthcare whenever learning requires physician demonstration, such as in titrating propofol for sedation⁹.

Applications of Artificial Intelligence in Anesthesiology:

With the recent advancement in machine learning algorithms, AI has shown that it can perform many tasks at or above the level of accuracy in human experts¹⁰, and as a data-intensive field, anesthesiology has an opportunity to improve the quality of perioperative care through the implementation of AI systems. In anesthesiology, critical events for patients may occur acutely and lead to poor outcomes very quickly. Before decompensation, there may be a brief window where physiological data can determine whether a patient is about to decompensate, and if intervention would avoid decompensation. Predictive analytics can take this data, use machine learning to recognize patterns and make predictions that could benefit clinical teams in anticipating and preventing significant events.

Clinical Decision Support: Time-series Estimation of Physiological States and Prediction of Decompensation:

Predictive analytics form the foundation for real-time clinical decision support tools. Time series clinical data, where the same data points are taken continuously or intermittently, may allow for an estimation of a physiological state based on an observed trend. If this knowledge can be captured in a model, predictions regarding disease progression or health status over time can be used to assist in acute care management, such as early intervention or therapy adjustments. Understanding the evolution of a disease is important in the management of any acute disease and is particularly important to be able to detect when the disease becomes worse, so that providers may potentially intervene on a significant pathophysiological state before it develops. The intraoperative and ICU environments are rich with data from continuously monitored physiological variables and laboratory variables^11,12, and with the digitization of healthcare, large datasets^13,14 are available for use in observational studies and predictive model development. Machine learning algorithms have been tested on these datasets to provide real-time decision support, estimating risk from patient and environmental data.

Direct clinical applications of machine learning have largely focused on sepsis¹⁵, because of its high mortality rate and the need for early clinical intervention. Early identification of sepsis in hospitalized patients remains a challenging problem, and while laboratory and EHR-based sepsis identification phenotypes are needed, they are limited due to the time delay between sepsis onset and laboratory value changes and data availability^16,17. The addition of real-time vital sign monitoring to machine learning algorithms has improved prognostic accuracy of early identification of sepsis. Recently, machine learning algorithms incorporating vital signs and EHR data demonstrated the ability to accurately predict the onset of sepsis 4–12 hours before clinical identification^18,19. A small randomized controlled trial showed that the addition of an ML algorithm to existing EHR-based sepsis identification systems was associated with decreased hospital length of stay and mortality²⁰. In the future, the EHR will have background machine learning algorithms continuously scanning clinical data in real-time²¹, assessing who may have sepsis, and contacting the responsible care teams rapidly to enable them to take action.

In the perioperative environment, management of physiological instability is vital for high-quality anesthetic care, and predictive models that may anticipate instability could allow care teams to provide early interventions. Kang et al²² demonstrated that machine-learning algorithms using data from anesthesia machines between the start of anesthesia induction and immediately before tracheal intubation can predict hypotension during the period between tracheal intubation and incision. Hatib et al developed a machine learning algorithm that can predict intraoperative hypotension, based on high-fidelity analysis of arterial waveforms²³. Their ML algorithm detects early changes in arterial waveforms, which may signal weakness in cardiovascular compensatory mechanisms affecting preload, afterload, and contractility. Using large arterial waveform datasets, they demonstrate that a machine-learning algorithm can be trained to predict hypotension in surgical patients’ records. Lundberg et al developed an AI-based warning system called Prescience that predicts hypoxemia during surgical procedures up to 5 minutes before it occurs²⁴. When provided with decision support from Prescience, anesthesiologists predicted hypoxemia with greater accuracy than by clinical judgment alone. Using a high-resolution ICU database, Hyland et al used machine learning to develop an early warning system that integrates measurements from multiple organ systems, predicting over 80% of circulatory failure events more than 2 hours in advance²⁵.

Closed-Loop Systems:

The closed-loop anesthesia delivery system relies on a “closed” feedback loop, where an automated device uses sensors, controllers, and actuators to keep systems automatically close to prespecified targets²⁶. Closed-loop systems provide individualized anesthetic drug titration through more frequent adjustment of controls, which could provide better regulation in the intraoperative and ICU environments. In the OR, closed-loop control has demonstrated that autonomous systems provide better hemodynamic management compared to manual control. Joosten et al randomized 90 patients for non-cardiac surgery to manual anesthesia management versus automated using three closed-loop controls²⁷ and demonstrated that automated anesthetic management outperformed manual control. Reinforcement learning has been used to develop an anesthesia controller that used feedback from a patient’s bispectral index (BIS) and MAP to control the infusion rates of propofol (in a simulated patient model)⁹. Outside of hemodynamic management, closed-loop control has great potential for addressing glycemic management in critically ill patients, who develop stress-induced insulin resistance causing hyperglycemia, large glucose variability, and episodes of hypoglycemia, which all have been associated with increased rates of morbidity and mortality. A recent study in pigs investigated the safety and performance of an autonomous closed-loop glucose control system, comprised of a continuous glucose monitor and an AI-based glucose controller, and compared it with an ICU physician using a clinical conventional protocol when the animals received an overdose of insulin and a subsequent unphysiological bolus of 20% dextrose²⁸. The combination of an accurate continuous blood glucose monitoring system with AI-based glucose control software abolished all severe hypoglycemic events despite a grave hypoglycemic provocation and demonstrated its ability to minimize moderate hypoglycemia despite having to deal with an unannounced intravenous insulin injection.

Health Care Processes:

Machine learning algorithms able to process electronic hospital operational data may offer an opportunity to improve efficiency, quality, and cost of healthcare services. Safavi et al developed and validated a neural network model that could predict daily inpatient surgical care discharges and their barriers²⁹. The model identified systemic causes of discharge delays, suggesting that such models could be utilized for their ability to increase the timeliness of patient discharges and optimize efficiency in existing healthcare processes. Predictive models have been developed and validated to predict which patients have higher risk of hospital readmission³⁰. Hopkins et al analyzed the National Surgical Quality Improvement Program (NSQIP) data on posterior lumbar fusions from 2011 to 2016 and used machine learning techniques to create a model predictive of hospital readmissions³¹. The model evaluated 177 unique input variables, was able to predict patients who would not require readmission and predicted up to 60% of readmissions.

The increasing availability of patient and population-level datasets offers the potential for health systems to use analytical approaches to identify and mitigate social determinants of health. These issues may not be satisfactorily addressed in typical care encounters, and thus opportunities to improve health outcomes, reduce costs, and improve care coordination may not be realized. Kasthurirathne et al³² used patient and population-level data that measured clinical, behavioral, and social determinants of health to develop decision models that could identify patients in need of various “wraparound” social services. AI technologies that can reduce care inefficiencies, cost, and administrative burden are of great interest to a healthcare system transitioning from fee-for-service to value-based care. If successful, AI technologies could reduce between ~ 760 to ~ 935 billion in annual waste, estimated to account for 25% of total care spending³³.

Computer Vision and the Development of Novel Monitoring Devices

Computer vision uses AI algorithms to perform machine-assisted image and video interpretation. Computer vision utilizes deep learning, a subset of machine learning algorithms that autonomously learn and improve. Computer vision’s most direct applications have been in the fields of pathology³⁴ and radiology³⁵, where AI algorithms automate image analysis and identify areas on slides and x-rays that have a high probability of demonstrating pathology. For anesthesiology, novel monitoring devices using computer vision, outside of the traditional contact-based monitors, are under investigation for development. Computer vision also has the potential to improve clinical care processes for in-hospital and intra-operative procedures.

Non-contact physiological vital sign sensing:

Computer vision is the foundation for non-contact sensing for vital sign assessment, where AI algorithms utilize real-time video to assess heart rate, respiratory rate, and pulse oximetry³⁶. Photoplethysmography (PPG) provides a non-invasive and non-contact means of sensing the cardiovascular pulse through detecting variations in transmitted or reflected light. Heart rate can be estimated because each cardiovascular pulse wave induces pulsatile skin color changes, which are imperceptible to the human eye but measurable with a camera³⁷. This technology has been validated in infants during their stay in a neonatal ICU, demonstrating that a heartbeat induced photoplethysmographic signal is strong enough to be measured³⁸. Computer vision also is effective for measuring respiration. Taati et al developed the first vision-based method that differentiates apnea types (obstructive sleep apnea vs central apnea)³⁹. AI algorithms analyzed chest and abdominal movements captured via an IR video camera without any physical monitors on the patient, opening up the possibility for new diagnostic modalities for evaluation of obstructive sleep apnea. The remote assessment of pulse oximetry is particularly relevant for high-quality triage during the COVID-19 pandemic, allowing for remote surveillance of high-risk patients. Wieringa et al showed that camera-based oximetry and remote recordings at different wavelengths are feasible⁴⁰. Bilenca et al developed a non-contact method based on the temporal sampling of time-integrated speckle using a camera-phone for noninvasive measurements of physiological parameters across the human fingertip, opening the possibility for using consumer-based smartphones as remote monitors of heart rate and pulse⁴¹. In the future, computer vision may run as background software during a telemedicine visit, providing accurate physiological vital sign data to healthcare providers to allow for more effective triage based on both patient history and quantitative physiological data.

Clinical care processes- facial recognition of pain and ultrasound guidance:

One of the most exciting applications for computer vision lies in its possibility to provide autonomous and continuous assessments of pain for hospitalized patients. Pain assessment is required for effective treatment of pain but currently relies on self-reporting or clinical nursing assessments. However, neonates, children, vulnerable adult populations, and the critically ill are often unable to self-report their pain intensity, leaving clinical pain assessment susceptible to underestimation bias and under-recognition. Computer vision has demonstrated the capability to quantify pain from facial expressions captured through video^42,43, opening up a future of high-quality in-hospital pain assessment and management. Rashidi et al demonstrated that AI algorithms can autonomously and non-invasively assess pain facial expressions from image data⁴⁴. Sikka et al used computer vision to evaluate pediatric post-operative pain, with detection of pain versus no pain model accuracy of AUC 0.84–0.94⁴⁵. The assessment of neonatal and pediatric pain has continued to be a challenge clinically, but computer vision may have utility in neonatal pain assessment during a NICU stay^46,47.

Computer vision can assist in the spatial analysis of ultrasound images and the performance of ultrasound-based procedures. In anesthesiology, computer vision has largely been applied to the automated analysis of ultrasound images to assist with the identification of structures during procedures⁴⁸. Cobey et al demonstrated that periprocedural aortic valve imaging by an AI echocardiographic software can result in an annular measurement that correlates with multidetector computed tomography in the pre-TAVR population⁴⁹. Smistad et al.⁵⁰ used AI to identify the femoral artery or vein and distinguish it from other potentially similar appearing structures such as muscle, bone, or acoustic shadow, demonstrating the possibility for future assistance in procedures in the OR and ICU.

Telemedicine, Digital Health, and Remote Surveillance

Telemedicine, digital health, and remote surveillance are extensions of each other and use telecommunications technology to deliver health care outside the traditional in-person-based paradigm. Before the COVID-19 pandemic, clinical workflows and financial reimbursement supported an in-person model of care. However, public health measures designed to curb coronavirus’ spread have driven the rapid adoption of telehealth, facilitating the transition of healthcare delivery to a hybrid model of in-person and remotely delivered care that utilizes digital technologies. Telemedicine has been tested in different clinical settings for appropriate use cases, demonstrating high patient and health professional satisfaction^51,52, and has demonstrated reduced healthcare costs and improved access to care^53,54.

In addition to the rapid adoption of telemedicine, the COVID-19 pandemic forced advancement of real-time, remote surveillance and has accelerated the “hospital-at-home” paradigm. Safavi et al. describe the prerequisites and limitations for proper remote surveillance⁵⁵ and describe how EMR databases must interact with tools across the inpatient and outpatient worlds. Advances in sensor networks research allow for health monitoring systems to become integrated within the home environment, providing healthcare professionals with a window into a patient’s lifestyle and remote quantitative data for informed clinical decision making. Wearable digital technologies can also provide intermittent or continuous monitoring of vital signs. For the perioperative anesthesiologist, telemedicine and digital health technologies will play an increasingly important role in pre-operative clinics⁵⁶, pre-habilitation programs⁵⁷, and post-discharge follow-up⁵⁸ in select patient populations. Portable, home-based, and mobile technologies are being studied for post-discharge follow-up that may include mobility monitoring and post-operative pain management⁵⁹. While post-discharge remote follow-up is still nascent for anesthesiology, remote monitoring of high-risk, high-cost patients may allow care teams to anticipate who will be readmitted within 90 days following discharge, allowing them to tailor early intervention ⁶⁰. Ultimately, telemedicine and digital technologies open up a future for new integrated, comprehensive, decentralized models of care that may help to address persisting challenges in the management of high-risk, high-cost surgical patients.

Challenges and Limitations:

For healthcare systems to meaningfully use AI technologies, multidisciplinary team efforts (Table 2) will be needed to convert AI-derived predictions or recommendations into meaningful clinical action. Clinicians will need to work with data scientists and engineers to ensure that data used for ML training is representative of a wide population of patients and to ensure that ML interpretations of the data are clinically sound. To best work in these teams, medical students⁶¹, residents, and faculty will need knowledge of AI and some knowledge of computational and statistical methods⁶². Innovations in medical and graduate education may allow early exposure to computational medicine and the uses and possibilities of AI technologies, which will be key for clinicians and researchers’ adoption and acceptance. User-centered design principles will be needed to present vast amounts of patient-generated data in a context that allows for actionable insights for care teams. Knowledge of clinical informatics and health information technology will allow for AI to be appropriately integrated with health IT infrastructure. Clinician investigators will need knowledge of implementation science to identify the facilitators of and barriers to the use of AI technologies and to ensure that AI implementation augments rather than complicates clinical workflow. To ensure sustainability and adoption, strong relationships with operational partners, including health care system leaders, frontline staff, and health IT, will be needed.

Table 2.

Examples of Knowledge Expertise for Successful Development and Implementation of AI Technologies with Applications in Anesthesiology

Stakeholders	Examples
Users	○ Physicians  ○ Nurses  ○ Hospital staff  ○ Patients and families
Knowledge Experts	○ Physicians and nurses  ○ Artificial intelligence experts  ○ Health IT experts  ○ Design experts
Decision Makers	○ Health system leadership and business administration  ○ Governmental regulatory agencies  ○ Industry  ○ Research labs

While the future of healthcare using AI is promising, it is important to acknowledge AI’s limitations as a technology and the importance of defining the correct use cases for its application to ensure rigorous quality control. While data are abundant, it is unclear if the data represents a diverse population. Evidence indicates that AI algorithms have the potential to worsen pre-existing demographic disparities, including racial biases⁶³. Obermeyer et al identified racial bias in an algorithm used by health systems to identify patients for high-risk care management, noting that black patients were less likely to be identified by the algorithm as candidates for potentially beneficial care programs than were white patients who had the same number of chronic illnesses⁶⁴. To avoid these and other biased results, mechanisms will need to be developed to account for the under-inclusion of minority health care data^65,66 before AI solutions are created. Efforts should be dedicated to validating that these algorithms perform well in different populations at different times⁶⁷.

Conclusion: A Vision of Big Data for the Learning Healthcare System

The availability of large EHR databases and advances in artificial intelligence bring us closer to a future of precision anesthesiology. AI technologies give healthcare systems the possibility to continuously learn from their data to update medical knowledge, guide practice, and modify internal processes to improve outcomes and quality. These real-time EHR-based analyses form the basis of a learning health system^68,69, which the Institute of Medicine defines as one “designed to generate and apply the best evidence for the collaborative healthcare choices of each patient and provider; to drive the process of discovery as a natural outgrowth of patient care; and to ensure innovation, quality, safety, and value in health care⁷⁰.”

As a field, anesthesiology has pioneered the implementation of patient safety initiatives, and AI will enable continued innovations in the delivery of safe perioperative and remote care. As healthcare advances towards a more technologically oriented future, AI will supplement the performance of clinicians and care teams, allowing them to focus on higher-level tasks, rather than to replace them. It is important to acknowledge that even with advancements in technology, the trust between an anesthesiologist and patient remains at the heart of field, ensuring that our clinical will remain a human endeavor.