See Article, p 896
In 2045, I hope we will have developed a planetary health infrastructure based on deep, longitudinal, multimodal human data, ideally collected from and accessible to as many as possible of the 9+ billion people projected to then inhabit the Earth.
–Eric Topol, MD, Executive Vice President, Scripps Research1
Over the 20th century, the global median life expectancy doubled. In general, this qualifies as a triumph of medicine. As a result, the nature and scope of the clinical practice of medicine have changed, with a tremendous increase in geriatric patients and medical complexity of disease presentation or before surgical management. Much of this epiphenomenon can be attributed to medical, biological, and technical advancements.
In the late 19th century, Dr William Osler mused, “Medicine is a science of uncertainty and an art of probability.” In the 21st century, today, and looking forward, with our sizeable armamentarium of advanced technologies, we are at an inflection point. Realistically, we can now aim for the highest level of granularity of disease mechanism at the individual level, rather than rely on statistical “averages” to develop “generalized” therapy for the population.
Precision medicine (PM) is a process rather than an event, whereby disease identification and therapy are designed for the individual rather than a group of people. The fundamentals of PM rely on the 4 Ps: predictive, preventive, personalized, and participatory.2 Recently, a fifth-P for pertinent was added, signifying an attempt to predict the best-individualized pathway per the patient’s preference—analogous to a route selection with automobile satellite navigation.
Given the current state of affairs, as a society, we stand at the confluence of novel scientific innovations where our day-to-day lives are being assisted by the predictive capabilities of devices, gizmos, and appliances. Indeed, this can be regarded as the second Industrial Revolution or the Technology Revolution, when science and technology have changed our lives dramatically. The only difference is that the basis of this current revolution is assisted intelligence (aka artificial intelligence [AI]) via advanced computing methods. In 2016, the National Institutes of Health (NIH) undertook a major initiative to enhance PM. The All of Us Research Program aims to build a “precision database” prospectively for at least 1 million people who agree to share their electronic health record (EHR) data, donate biospecimens, laboratory data, physical exams, and survey response.3 As of May 2019, about 200,000 individuals had enrolled.
In recent decades, it has been increasingly accepted that health is determined by multiple factors, with actual medical conditions accounting for just 10% of influence, genetics contributing 30%, and physiological, behavioral, psychological, and socioeconomic determinants accounting for 60%.4 In this era of Big Data, deep neural network analysis, cloud computing, data science, advances in clustered regularly interspaced short palindromic repeat (CRISPR) gene-editing technologies, and metabolomics, integration, and development of personalized pathways to health and wellbeing, incorporating all the above factors is truly conceivable. The “uberization” of health care—that is, patients managing their health with the help of smart devices like the continuous blood glucose monitor, Apple Watch for atrial fibrillation (AF) detection, and Fitbit for calorie intake—is finally happening!
IMMINENT PM: THE AMALGAMATION OF PRECISION SCIENCE, PRECISION INTEGRATION, AND PRECISION DELIVERY
We believe that perioperative PM is best conceived as a feedback loop, rather than the stratified-layer approach to PM (4-Ps) (Figure 1). Precision science involves magnifying granularity to identify disease mechanics, reduce adverse drug reactions (ADRs), and avoid drug-drug interactions. Precision integration is the rapid analysis and translation of data. Precision delivery ensures delivery and healthy homeostasis through titration of therapy, with live, real-time feedback based on patient-reposted outcomes and side effects as recorded by personal health data recorders.5 While PM works functionally as a feedback loop, for a more straightforward conceptual understanding, we may group it into the following aspects: genomics, n-of-1 trials, advanced biosensors, data sharing, and implementation science and societal acceptance.
The “one-gene-one disease” theory has not achieved much traction in clinical settings because of its variability in penetrance and expressivity. The onus lies in a multitude of reasons like gene polymorphisms, gene modifications, and epigenetics. Traditionally, genome-wide association studies (GWASs) are designed to identify genetic variants leading to a phenotype (eg, acute kidney injury and myocardial injury). In contrast, a phenome-wide association study (PheWAS) attempts to identify multiple phenotypes associated with a single genetic predictor (Figure 2). Cloud computing, data integration, and machine learning (ML) have added a new dimension. While the concepts remain the same, data integration and accrual can be done at a much faster pace using the EHR, either in real time or retrospectively. These sources can then be used to derive individualized predictive pathways. Such initiatives are already underway at institutions, including Vanderbilt University, Harvard University, and the University of South Carolina.6
Deriving data from the EHR to identify a clinical phenotype is helpful. These data could be hypothesis-generating and cover a wide breadth in a relatively short timeframe; however, at the same time, these data lack the specificity required to become a mainstay for clinical use. Randomized clinical trials (RCTs) still do not lose their utility but can be designed based on these preliminary hypotheses generated by integrated EHR and PheWAS studies. PM-based RCTs can be designed to be n-of-1 in nature.
Methodology for n-of-1 Trials
Clinical research principles currently stand at a crossroads. Evidence-based medicine relying on traditional RCTs generates an “average” dose or therapy. However, no actual patient exactly matches the average. This means the clinical decision ultimately lies with the clinician and therefore, by default, is subject to several biases. A new type of trial design, often referred to as n-of-1 trial, aims to advance PM by focusing on customized therapy for the individual patient. Essentially, the individual patient serves as their own control. An n-of-1 trial presents challenges, including: (1) it is challenging to design and conduct and (2) it may be difficult to sustain the distinction between care goals and research, as they can get intertwined.7
Broadly, these individualized trials may be designed as multiple crossover designs or as prepost intervention trials, depending on the nature of the study. A significant advantage over traditional RCTs is that n-of-1 studies do not require large numbers of patients. However, a disadvantage can be a required extended trial length, as a single patient must undergo multiple phases of various therapies with intermittent washout periods. Such study designs will by default have increased diversity, equity, and inclusion in the population enrolled and studied (Figure 3). Details of n-of-1 trials are beyond the scope of this article; they can be found in the review by Kane et al.7
The past few decades have seen exponential progress in monitoring devices, from point-of-care in vitro monitoring to self-monitored, wearable biosensor devices. For example, the Apple Watch (Apple Inc) is the most commonly used device that first detects AF, as up to 18% of all AF is diagnosed when patients present with a stroke.8 Photoplethysmography, the technology used to detect subclinical AF, could prove extremely valuable.9 Miniaturized biosensor devices are also expanding, including implantable or wearable types that are capable of analyzing electrolytes, metabolites (lactate, cortisol, tyrosine, and glucose), enzyme, nucleic acid, or antibodies from sweat, tear, or saliva. These biosensor devices are integrated with a handheld mobile technology platform.5 Related proof-of-concept studies exist, but further larger trials are needed.10 In addition, concerns about data security and privacy will need to be addressed.
An exponential increase in high-throughput technology to capture Big Data is one thing; synthesis to yield results is another requirement. Unless Big Data are coalesced and shared among investigators, unified algorithms may be difficult to decipher. Prior data sharing will likely help spare published findings about PM from the ongoing “reproducibility crisis” affecting traditional methodologies, enabling the transformation of data to usable information. As a step forward, the US NIH has recently endorsed this approach, and the NIH is beginning to require integrated data sharing in the routine conduct of any biomedical project.11 There have been 2 trends: (1) standardized data collection and definition of variables and outcomes and (2) establishing more nationwide registries such as the Society of Thoracic Surgery Adult Cardiac Surgery Database12 and the Multi-institutional Multicenter Perioperative Outcomes Group.13 Future perioperative PM-based studies with unified data are expected to increase.
Implementation Science and Societal Acceptance
In a recent editorial in the journal Science, Proctor and Geng14 astutely state that “knowing what to do” does not ensure “doing what we know.” As an example, they cite realities seen during the coronavirus disease 2019 (COVID-19) pandemic: not only did only 55% of Americans accept the known beneficial clinical intervention (vaccine), but also many met wearing a mask and social distancing with indifference or resistance. This illustrates the powerful role social acceptance can play, helping or hindering the health of a population, regardless of how good or precise the translation of science from bench to bedside.
Research in coming decades will need to emphasize understanding these social, psychological, and behavioral barriers that prevent such translation to the population writ large. Again, to achieve inclusion and equity, new discoveries in PM and societal acceptance and penetrance will need to be better balanced. The upcoming era of implementation, diversity, inclusion, and convergence in science will determine how best to make the bench-to-bedside translation occur (Figure 4).14,15
While PM in internal medicine is essentially the maintenance of health and prevention of disease, PM in the cardiology setting deals with the mitigation of disease and lifestyle modification to change the occurrence and progression of course of disease. For example, in cardiology, genetic predisposition to heart disease is informative, but disease modification and persistent sustained behavioral modification can change disease progression and promote health.
Some notable examples of PM in cardiology include: (1) P2Y12 inhibitor therapy optimization anticoagulation after drug-eluting stent management, as activity is influenced by loss-of-function alleles (CYP2C19*2 and CYP2C19*3) manifesting poor drug response, while gain-of-function allele (CYP2C19*17) is associated with increased risk of bleeding16; (2) avoiding risk of rhabdomyolysis with statins for hyperlipidemia by identifying at-risk patients with rs4263657 gene single nucleotide polymorphism (SNP)17,18; and (3) anticipating the variable dose requirement of warfarin in patients with SNP CYP2C9*2, CYP2C9*3, and VKORC1 genotypes.19,20
In oncology, PM can be applied to understand the etiology of cancer in specific individuals and to identify how to modify the course in terms of progress, survival, or recurrence. PM can also assess and address biologic milieu to understand how to avoid recurrence. Notable examples of PM in oncology include: (1) the treatment of chronic myeloid leukemia (Philadelphia chromosome positive) producing enzymatic aberrant Bcr-Abl with imatinib mesylate; (2) identification of RNA transcripts (transcriptomics) for solid tumor expressions to match with selected therapy (WINTHER-trial); and (3) immunochemical assays (proteomics) for identification of prognostic or predictive biomarkers for target therapy for breast cancer (HER2 overexpression).21
Therefore, contrasting PM for cardiology versus oncology, the former deals with more continuous modifications, while the latter deals with more absolutes. Overall, the real deal is the virtue of PM to convert nonmodifiable factors to modifiable factors for a disease state, as illustrated in the above examples.
Perioperative PM is unique from both these medical specialties. Generally, in the setting of perioperative care, exposure to anesthesia is short lasting. Nevertheless, in the midst of surgical stress, several subclinical changes occur at the cellular level during the perioperative period—referred to as the “perioptome.”22 The perioptome is the distinct phase, where the genome, surgery, anesthesia, postoperative recovery, and intensive care intersect. Major clinical concerns remain in the perioptome, such as postoperative cognitive dysfunction (POCD), myocardial injury after noncardiac surgery (MINS), and postoperative acute kidney injury (PO-AKI), all of which may be considered persistent perioperative events with a bearing on a patient’s recovery and long-term survival after surgery. It is likely that in the next decade, perioperative medicine research into these clinical conundrums will be shaped by the volume, velocity, variety, veracity, and value (5-V) of Big Data (Figure 5).4
In the perioperative setting, we foresee moving away from a reactionary “detection-and-deflection” pathway toward adopting a “logistic-prediction” pathway model. More than ever, owing to advances in innovative science, evolution of predictive preoperative pathways for intraoperative and postoperative management is the likely direction. Current major opportunities are postoperative patient-centered outcomes (personalized outcomes of relevance), ever-increasing complexity due to multiple comorbidities at the time of surgical management, a growing population of geriatric patients, and combating human error.
The future of anesthesiology is likely an expanded perioperative surgical home model in which, conceptually, other medical specialties and input from other experts, including data scientists, genetic engineers, and biochemists, will be incorporated to individualize and formulate precision perioperative management for each patient. The integration of vast data from omics, genetics, epigenetics, and gene dynamism to emanate predictive preoperative pathways is imminent.
Data science-based integration of real-time intraoperative hemodynamics for developing early anesthesia information management systems (AIMS) to guide clinical decision-making as well as identification of clinical subphenotypes (endotypes) are potential directions. For example, integrating clinical data, genetics, biomarkers, and multiomics has yielded several endotypes for acute respiratory distress syndrome (ARDS), namely, hypoxia severity phenotype (clinical), inflammatory versus noninflammatory phenotype (clinical), precipitating risk factor phenotype (clinical), direct versus indirect lung injury (etiology phenotype), timing of onset (temporal phenotype), radiographic (imaging phenotype), genetic (genomic phenotype), and biomarker phenotype.23 Similarly, occurrences of intraoperative adverse events such as hypotension, arrhythmia, and anesthetic drug sensitivity may yield certain yet unrecognized endotypes, as are seen with ARDS. Expectedly, these will yield an alternative pathway for each endotype, which can potentially prevent postoperative complications, such as perioperative neurocognitive disorders, postoperative AF, PONV, and PO-AKI.
PM can be applied to mitigate ADRs, a continued burden for patients and health care systems. ADRs may lead to increased harm and also systemic inefficiency from unnecessary risk-averse policies. For example, while malignant hyperthermia is a life-threatening ADR with certain anesthetics, overtesting and risk aversion in every family member may not be the most efficient approach.
Identifying true ADRs by integrating GWAS, PheWAS, and the EHR can potentially result in a high-throughput framework that efficiently identifies eligible patients, thus creating maximum clinical utility. As a proof of concept, the Vanderbilt University Medical Center BioVU DNA Biobank conducted a GWAS for ADR to 14 common drug/drug groups from 81,739 patient records for documented “drug allergies.” These investigators identified 7 genetic loci with genome-wide significance (P < 5 × 10−8) associated with ADRs associated with CYP2D6 and OPRM1 for CYP2D6-metabolized opioid ADR. These are the 2 most common gene loci influencing majority drugs used in perioperative settings.24 Such integrative studies, deriving phenotypes from the EHR and matching up with genomic testing, are increasing and gaining traction as an efficient means to identify selected patients, reduce costs, speed up diagnosis, and improve care.25
Futurists and advocates of PM opine that the best way to assimilate the vastness of the field into clinical utility is by open-science and open-resource tools. The Center for Precision Medicine at Vanderbilt University Medical Center has developed “phecodes” by integrating EHR-based phenotyping from International Classification of Diseases (ICD) codes and integrating it with genotypes. To date, they have successfully identified and cataloged 1867 phecodes for different diagnoses, symptoms, and findings.26,27
Another, similar web-based pharmacogenomic tool for perioperative care has been developed by identifying drug-gene interaction as described by the US Food and Drug Administration (FDA) and in published literature.6,28 The FDA has identified 258 drugs and their corresponding 461 gene variants/mutations leading to drug efficacy and ADRs. Of these 258 drugs, 20 anesthetic medications are composed of 7.75% pharmacogenes. The inhaled agents and opioids can initiate drug-gene interaction or cause ADR events intraoperatively. The FDA found that the Cytochrome P450 (CYP2D6 and CYP3A4), ryanodine receptor (RYR1), butyrylcholinesterase (BChE), and glucose 6-phosphate dehydrogenase (G6PD) genes to be the most interactive with anesthetic drugs. Although the FDA encourages inclusion of pharmacogenetic biomarker labeling for approved drugs, only 12% of drugs licensed between 1998 and 2012 carried these labels.29
Development, awareness, and utilization of such resources can improve selective genetic testing, thereby reducing the variations from uncertain significance, and dictate implementing preemptive clinical pharmacogenetics. While the development of these individualized or center-specific web-based tools is highly encouraging, the next step will be integrating these data into a common language for universal applicability. AI-based Natural Language Processing could be a major driver. This approach was used successfully in developing electrocardiogram ontology via the EHR across several nongovernmental organizations and agencies, using American Heart Association, American College of Cardiology, and Heart Rhythm Society standards.30
In the clinical practice of perioperative medicine, a vexing hurdle has been improving postoperative morbidity and mortality. Almost 234 million people undergo surgery globally every year; of these, 7 million suffer a postoperative adverse event and 1 million die.31 As highlighted by Sessler,32 despite the marked reduction in intraoperative mortality over the last 150 years, the struggle to reduce postoperative morbidity and mortality continues.
Development of predictive science integrating GWAS, PheWAS, and the EHR through AI could potentially address how best to identify and to monitor at-risk patients and to improve their outcomes. For example, Shin et al33 developed an ML-based model for predicting 30-day mortality from MINS using clinical data from 7629 patients. They found that C-reactive protein level, insulin prescription, antiplatelet prescription, calcium channel blocker prescription at discharge, and peak troponin level were all predictive of MINS (extreme gradient boosting model outperformed; area under the receiver operating curve, 0.923; 95% CI, 0.916–0.930). Prescriptions for antiplatelet drugs and beta-blockers at discharge were also associated with reduced mortality in the cohort. Another example of how genetic predisposition could help identify patients at risk for postoperative complications was highlighted by a recent study of patients who underwent cardiac surgery in whom the prediction of postoperative atrial fibrillation (POAF) was much improved by the addition of polygenic risk score (odds ratio, 1.63 per SD increase in PRS [95% CI, 1.41–1.90]).34
PO-AKI is another frequent complication. Tseng et al35 enrolled 671 patients prospectively undergoing cardiac surgery patients to design ML-based algorithms logistic regression, support vector machine, random forest, extreme gradient boosting, and ensemble. They found that 24% of 163 patients developed AKI by Kidney Disease Improving Global Outcomes (KDIGO) criteria in the first week after surgery. The ensemble model had the strongest predictability, and the most influential factors were intraoperative urine output, units of transfused packed red cells, and preoperative hemoglobin.
Several ML-based prediction models for postoperative delirium have been developed and shown to outperform traditional stepwise logistic regression.36–38 Such throughput information, integrating genetics with phenomics as extracted from electronically captured data, could significantly aid clinical decision-making, identifying at-risk patients who could benefit from targeted interventions, thereby mitigating a complication.
In the near future, it is inevitable that information technology and AI will be the significant drivers for perioperative PM. The framework for defining perioperative PM has been developed, but the next task is to find ways to disseminate and to integrate successfully across all clinicians caring for surgical patients. Although the process has been initiated with some ventures, significant challenges remain. Real-time unification and integration of electronic data, challenges in designing customized research studies with a focus on individual patient-specific characteristics, and implementation and convergence science to translate from bench to the perioperative period and the operating table are factors that need to be addressed for perioperative PM to fulfill its potential promise.
Once the key elements of perioperative PM have been integrated into the routine clinical care of surgical patients, further research opportunities will arise to help with testing whether individualized perioperative management, including anesthetic techniques and medications, could play a significant role in the risk of perioperative outcomes, such as AKI, MINS, POCD, and POAF. These future research opportunities will provide further evidence on how to prevent, mitigate, or treat perioperative complications and to avoid harmful drug interactions and ADRs.
Thus, the time has come to identify and focus our future efforts beyond protocolized recovery pathways and to work toward individualized recovery pathways. To promote that ongoing, iterative process, future perioperative team models must incorporate data scientists, biomedical engineers, pharmacoepidemiologists, and pharmacogeneticists in an integrated way. In the quest for minimizing preventable harm and maximizing patient safety, PM offers a great way to eliminate human error, reduce ADRs, and offer personalized patient management, with improvements in postoperative outcomes and patient satisfaction.
The authors thank Yvonne Poindexter, MA (Editor, Department of Anesthesiology, Vanderbilt University Medical Center, Nashville, Tennessee, USA) for editorial contributions to the article.
Name: Pal Nirvik, MBBS, MD, FASE.
Contribution: This author helped design the work, draft the manuscript, and approve the final manuscript.
Name: Miklos D. Kertai, MD, PhD.
Contribution: This author helped design the work, provide content expertise, critically revise the manuscript, and approve the final manuscript.
This manuscript was handled by: Thomas R. Vetter, MD, MPH.
1. Looking forward 25 years: the future of medicine. Nat Med. 2019;25:18041807.
2. Pal N, Kertai MD. Perioperative precision medicine: where are we in 2020? Curr Opin Anaesthesiol. 2020;33:463–474.
3. Denny JC, Rutter JL, et al.; All of Us Research Program I. The “All of Us” research program. N Engl J Med. 2019;381:668676.
4. Johnson KB, Wei WQ, Weeraratne D, et al. Precision medicine, AI, and the future of personalized health care. Clin Transl Sci. 2021;14:86–93.
5. Kim J, Campbell AS, de Ávila BE, Wang J. Wearable biosensors for healthcare monitoring. Nat Biotechnol. 2019;37:389–406.
6. Zarei S, Costas Y, Orozco G, et al. A web-based pharmacogenomics search tool for precision medicine in perioperative care. J Pers Med. 2020;10:E65.
7. Kane PB, Bittlinger M, Kimmelman J. Individualized therapy trials: navigating patient care, research goals and ethics. Nat Med. 2021;27:1679–1686.
8. Raja JM, Elsakr C, Roman S, et al. Apple watch, wearables, and heart rhythm: where do we stand? Ann Transl Med. 2019;7:417.
9. Karmen CL, Reisfeld MA, McIntyre MK, Timmermans R, Frishman W. The clinical value of heart rate monitoring using an Apple watch. Cardiol Rev. 2019;27:60–62.
10. Seshadri DR, Bittel B, Browsky D, et al. Accuracy of Apple watch for detection of atrial fibrillation. Circulation. 2020;141:702–703.
11. Jorgenson LA, Wolinetz CD, Collins FS. Incentivizing a new culture of data stewardship: the NIH Policy for data management and sharing. JAMA. 2021;326:2259–2260.
12. The Society of Thoracic Surgeons. Society of Thoracic Surgery Adult Cardiac Surgery Database. Accessed December 12, 2021. https://www.sts.org/registries/sts-national-database
13. Multi-institutional Multicenter Perioperative Outcomes Group. Accessed December 12, 2021. https://mpog.org
14. Proctor EK, Geng E. A new lane for science. Science. 2021;374:659.
15. Dzau VJ, Balatbat CA, Ellaissi WF. Revisiting academic health sciences systems a decade later: discovery to health to population to society. Lancet. 2021;398:2300–2304.
16. Ahmad T, Voora D, Becker RC. The pharmacogenetics of antiplatelet agents: towards personalized therapy? Nat Rev Cardiol. 2011;8:560–571.
17. Wei MY, Ito MK, Cohen JD, Brinton EA, Jacobson TA. Predictors of statin adherence, switching, and discontinuation in the USAGE survey: understanding the use of statins in America and gaps in patient education. J Clin Lipidol. 2013;7:472–483.
18. Link E, Parish S, Armitage J, et al.; SEARCH Collaborative Group. SLCO1B1 variants and statin-induced myopathy–a genomewide study. N Engl J Med. 2008;359:789–799.
19. Finkelman BS, Gage BF, Johnson JA, Brensinger CM, Kimmel SE. Genetic warfarin dosing: tables versus algorithms. J Am Coll Cardiol. 2011;57:612–618.
20. Gage BF, Eby C, Johnson JA, et al. Use of pharmacogenetic and clinical factors to predict the therapeutic dose of warfarin. Clin Pharmacol Ther. 2008;84:326–331.
21. Tsimberidou AM, Fountzilas E, Nikanjam M, Kurzrock R. Review of precision cancer medicine: evolution of the treatment paradigm. Cancer Treat Rev. 2020;86:102019.
22. Shaw A, VandeVen T. Secrets of the perioptome: new tools for a new concept. Curr Opin Crit Care. 2010;16:365–370.
23. Reilly JP, Calfee CS, Christie JD. Acute respiratory distress syndrome phenotypes. Semin Respir Crit Care Med. 2019;40:19–30.
24. Zheng NS, Stone CA, Jiang L, et al. High-throughput framework for genetic analyses of adverse drug reactions using electronic health records. PLoS Genet. 2021;17:e1009593.
25. Morley TJ, Han L, Castro VM, et al. Phenotypic signatures in clinical data enable systematic identification of patients for genetic testing. Nat Med. 2021;27:1097–1104.
26. Bastarache L. Using phecodes for research with the electronic health record: from PheWAS to PheRS. Annu Rev Biomed Data Sci. 2021;4:1–19.
27. PheWAS Resources. Phecode Map 1.2 with ICD-9 Codes. Accessed December 12, 2021. https://www.phewascatalog.org/phecodes
28. Zarei S, Costas Y, Orozco G, et al. A Web-Based Pharmacogenomics Search Tool to Exploit Precision Medicine in Perioperative Care. Accessed December 12, 2021. https://pharmacogenomics.github.io/pharmacogenetics/
29. Tutton R. Pharmacogenomic biomarkers in drug labels: what do they tell us? Pharmacogenomics. 2014;15:297–304.
30. Epstein RH, Jean YK, Dudaryk R, et al. Natural language mapping of electrocardiogram interpretations to a standardized ontology. Methods Inf Med. 2021;60:104–109.
31. Behrooz A. Pharmacogenetics and anaesthetic drugs: implications for perioperative practice. Ann Med Surg (Lond). 2015;4:470–474.
32. Sessler DI; American Society of Anesthesiologists. FAER-Helrich Lecture: reversing postoperative mortality. 2021. Accessed December 12, 2021. https://www.asameetingnewscentral.com/home/article/21771506/faerhelrich-lecture-reversing-postoperative-mortality
33. Shin SJ, Park J, Lee SH, Yang K, Park RW. Predictability of mortality in patients with myocardial injury after noncardiac surgery based on perioperative factors via machine learning: retrospective study. JMIR Med Inform. 2021;9:e32771.
34. Kertai MD, Mosley JD, He J, et al. Predictive accuracy of a polygenic risk score for postoperative atrial fibrillation after cardiac surgery. Circ Genom Precis Med. 2021;14:e003269.
35. Tseng PY, Chen YT, Wang CH, et al. Prediction of the development of acute kidney injury following cardiac surgery by machine learning. Crit Care. 2020;24:478.
36. Racine AM, Tommet D, D’Aquila ML, et al.; the RISE Study Group. Machine learning to develop and internally validate a predictive model for post-operative delirium in a Prospective, Observational Clinical Cohort study of older surgical patients. J Gen Intern Med. 2021;36:265–273.
37. Datta S, Loftus TJ, Ruppert MM, et al. Added value of intraoperative data for predicting postoperative complications: the MySurgeryRisk PostOp extension. J Surg Res. 2020;254:350–363.
38. Xue B, Li D, Lu C, et al. Use of machine learning to develop and evaluate models using preoperative and intraoperative data to identify risks of postoperative complications. JAMA Netw Open. 2021;4:e212240.