Medicine is traditionally slower than other industries to develop and adopt new technologies, but A.I. is beginning to enter clinical practice to complement the expertise of health care practitioners in fields such as radiology.2 A.I. and machine learning bear similar promise in anesthesiology, but many challenges remain in incorporating machine learning into daily clinical practice.3 To date, most A.I. projects published in the medical literature have focused on the development and validation of machine-learning models and algorithms. A small minority of these reports describe a project that has been integrated successfully into daily clinical workflow, and few, if any A.I. efforts have demonstrated measurable improvements in patient outcomes.
On January 30, 2020, a PubMed search of peer-reviewed publications yielded 48 original investigation articles when using combinations of the Medical Subject Headings (MeSH) and free-text terms “anesthesia,” “artificial intelligence,” and “machine learning” for 2015 onwards, sifted for relevance. Only 13 of these articles were published in clinical journals, while the remaining 35 articles were found in engineering and basic science publications. In contrast, there were 18 editorials and narrative reviews in anesthesiology journals alone (Supplemental Digital Content, Table 1, http://links.lww.com/AA/D53, lists the details of the literature search).
This imbalance of original and secondary articles would suggest that A.I. in anesthesia may be approaching the Peak of Inflated Expectations described by the Gartner Hype Cycle (Figure).4 The cycle describes the phases of public interest recognized during the development of new technology. A breakthrough Innovation Trigger leads to proofs-of-concept and significant publicity, but initially, no usable or commercially viable products exist. As publicity continues, development does not yet meet the expectations of the wider field, attitudes begin to harden, and general interest wanes in a Trough of Disillusionment. Slowly, the technology evolves along the Slope of Enlightenment and a greater number of examples of benefit emerge and are understood. Eventually, mainstream adoption takes off until the Plateau of Productivity is reached. At this point, broad applicability and relevance are evident and guidelines for assessing viability are more clearly defined. Eventually, ideas graduate from the hype cycle as mature technologies.
Why is there a dearth of original investigations in anesthesiology journals compared to engineering journal reports in the anesthesia domain? The answer is undoubtedly multifaceted. A.I. and machine learning are complex concepts rooted in computer science. Close collaboration between data science specialists and anesthesiologists is required to yield a high-quality match of innovative techniques with clinically relevant outcomes. In any Innovation Triggerphase, there are typically more failures than successes, which can contribute to publication bias. Because no clinical applications of A.I. have established a Slope of Enlightenment, clinicians’ exposure to real-world applications of A.I. may be limited to popular media accounts of autonomous vehicles and the supercomputer Watson.5
This special issue of Anesthesia & Analgesia moves to redress the imbalance of published hype versus progress in the field of A.I. for anesthesiology. (The authors of this editorial recognize the irony of their contribution to this imbalance.) A majority of the included articles are research reports that represent significant progress in the application of machine learning and A.I. in anesthesiology. Mathis et al6 report a novel use of machine learning with Multicenter Perioperative Outcomes Group data to predict future heart failure with reduced ejection fraction. Nagaraj et al7 developed a machine-learning model to predict hypnotic levels using repurposed electroencephalogram data, and then externally validated the model using data collected for a second study.7 Other articles describe the development of advanced machine-learning models that use continuous vital signs data to predict clinically significant deteriorations including hemorrhage, hypotension, bradycardia, and opioid-induced ataxic breathing.8–12 Wang et al13 report a proof-of-concept in manifold learning, an important new technique in the field. These articles represent Innovation Trigger proofs-of-concept and developed their new techniques using preexisting data sets, retrospectively collected electronic medical record data, animal models, healthy volunteers, and small observational clinical trials. While none of the investigators report integration into clinical practice, their efforts form a foundation of innovation on which clinical implementations might be based.
Perioperative and critical care generate data in both high volume and high veracity, making these environments ideal for the integration of machine learning in clinical care. However, much of contemporary anesthesia practice has become so safe overall that many adverse outcomes have become rare. While this development is admirable, this also means that few records are available to study events where patient safety was threatened. Large data sets are often used to power statistically and clinically meaningful studies, thereby necessitating data sharing across institutions. These large data sets have the potential to become unwieldy without initiatives to improve collection, storage, and validation, such as the efforts of the Multicenter Perioperative Outcomes Group.14
As the field of A.I. in perioperative care advances, A.I. will be implemented into the clinical workflow. It will be incumbent on anesthesiologists, data scientists, and informaticians to study the performance of these A.I. implementations to understand their true impact on perioperative care and patient outcomes. The health care industry must develop standards for the conduct and reporting of A.I. research to guide researchers, reviewers, and editors. To responsibly advance the field, researchers must take into account the full ethical, implicit bias, and safety considerations of deploying the technology.15 As with the introduction of any new treatment, the need for careful appraisal, validation, and monitoring of tools using A.I. does not stop with implementation—a framework for a type of “Phase IV” postmarketing surveillance will be essential to assess performance and monitor behavior. The US Food and Drug Administration has incorporated this need into its proposed regulatory framework for A.I.16 Successful integration of technology, just like successful patient care, requires a multidisciplinary team approach and the hearts and minds of clinicians. The engaged stakeholders for this A.I. care team may include data scientists, ethicists, medical subject matter experts, project managers, analysts, and end users.
Because A.I. has already begun to change our lives, so will it doubtlessly in time change the field of anesthesiology. Academics and clinicians should consider the possibilities of incorporating A.I. to enhance patient care, else we risk pushing development solely to commercial organizations who may be less inclined toward transparency and peer-reviewed research during the development process. With this special issue on A.I., Anesthesia & Analgesia has demonstrated remarkable foresight, opening the way for the rest of the specialty to follow.
1. Goasduff L. Top trends on the Gartner hype cycle for artificial intelligence. Available at: https://www.gartner.com/smarterwithgartner/top-trends-on-the-gartner-hype-cycle-for-artificial-intelligence-2019. Accessed January 31, 2020.
2. El Naqa I, Haider MA, Giger ML, Ten Haken RK. Artificial Intelligence: reshaping the practice of radiological sciences in the 21st century. Br J Radiol. 2020;93:20190855.
3. Hofer IS, Burns M, Kendale S, Wanderer JP. Realistically integrating machine learning into clinical practice: a road map of opportunities, challenges, and a potential future. Anesth Analg. 2020;130:1115–1118.
4. Gartner, Inc. Gartner hype cycle. Accessed January 31, 2020.
5. Herper M. MD Anderson benches IBM Watson in setback for artificial intelligence in medicine. Accessed January 31, 2020.
6. Mathis MR, Engoren MC, Joo H, et al. Early detection of heart failure with reduced ejection fraction using perioperative data among noncardiac surgical patients: a machine-learning approach. Anesth Analg. 2020;130:1188–1200.
14. Colquhoun DA, Shanks AM, Kapeles SR, et al. Considerations for integration of perioperative electronic health records across institutions for research and quality improvement: the approach taken by the Multicenter Perioperative Outcomes Group. Anesth Analg. 2020;130:1133–1146.
15. Wiens J, Saria S, Sendak M, et al. Do no harm: a roadmap for responsible machine learning for health care. Nat Med. 2019;25:1337–1340.
16. US Food & Drug Administration. Proposed Regulatory framework for modifications to Artificial Intelligence/Machine Learning (AI/ML)-based Software as a Medical Device (SaMD).