Patients commonly use substances that pose risks to their health such as tobacco, alcohol, and other drugs in the perioperative period, with important implications for their anesthetic care. In recent surveys of patients presenting for surgery, more than two in five patients evaluated in the preoperative clinical setting endorsed some level of substance use that poses surgical risk, including use of cannabis products.  Clinically, it can be challenging to consistently assess use of different types of substances in an efficient yet reliable manner. Cannabis use presents a complex yet increasingly prevalent practice reported by patients. Most Americans now live in states that allow medical use of cannabis as a treatment for conditions including chronic pain, and many states have legalized recreational cannabis use. Perhaps unsurprisingly, more Americans report daily use of cannabis than daily alcohol consumption.  This change in population-level behavior is likely to translate to individual patients who present for surgery. The recent surge in cannabis use has exposed significant gaps in our understanding of how presurgical cannabis consumption affects anesthesia management and patient care.

In this issue of Anesthesiology, Sajdeya et al. examine the association between cannabis use in the 60 days before surgery and the dose of inhalational anesthesia administered in the operating room.  Cannabis use was identified using classifications derived from natural language processing methods along with other structured data points such as diagnostic codes. They hypothesized that anesthesia doses would be higher among persons with current cannabis use compared to those with no use. The study focused on persons aged 65 yr and older who received care at one health system during a 3-yr period. After using propensity score matching methods to balance confounding variables, they examined the minimum alveolar concentration (MAC) values for isoflurane and sevoflurane for 1,340 patients, converting MAC values into an equivalent unit that was then averaged over every minute of the case. They found that the difference in the dose of anesthesia was higher for patients with cannabis use (mean MAC, 0.58) when compared to those without (mean MAC, 0.54), a result that was statistically significant although likely not clinically meaningful (mean difference, 0.04; 95% CI, 0.01 to 0.06).

Among strategies to classify cannabis exposure for included participants, natural language processing methods offer one way to glean insights from data within electronic health records. Unstructured free-form text in electronic health record notes often contains crucial clinical information. Analyzing unstructured text from these notes may identify a particular individual as someone who possesses a relevant characteristic or not (e.g., whether they use specific substances). Manual review of unstructured text in a limited number of notes is feasible for researchers. However, analyzing thousands or millions of notes requires automated processing to efficiently scale this size of project. In this study, the approach to determine cannabis use incorporated data from unstructured clinical notes from a patient’s chart documented in the preoperative period up to 60 days before surgery.

Natural language processing builds on a linguistics approach using rule-based models of human language along with statistical, machine learning, and other analytical approaches. Before the analytical approaches may be applied, consensus must exist about the target concept. Defining a comprehensive and exhaustive list of terms and their meaning in a lexicon is a fundamental task. To generate a list of key words, phrases, and concepts that contribute to a relevant lexicon, researchers may review previous literature, query the opinion of experts, and draw from databases built for this purpose. For example, terms for cannabis include not only words like tetrahydrocannabinol, marijuana, and cannabidiol, but also their abbreviations, brand names, and potential misspellings. One concern for accurately identifying use of cannabis, as well as substances of other types, is the common use of different names, nicknames, or slang words that refer to specific drugs, which may evolve over time and introduce challenges in misclassification. Whether pot indicates cannabis or a container for cooking clearly depends on context.

Extracting phrases from charts and rating their status manually allows for researchers to refine their approach to classification, such as whether active cannabis use is present or not. An existing validated assessment or consensus among two or more raters with clinical domain expertise is often used as a benchmark and thus serves as the basis for data on which to train machine learning and deep learning models. In testing these models, researchers often evaluate the precision, sensitivity, specificity, and other scores that indicate performance. Similar to other validation studies, replicating the analysis using a separate database highlights the degree to which findings from classification algorithms may apply more broadly. Externally validating findings from natural language processing algorithms takes time and resources, although without this step, model results may be inaccurate, biased, and not generalizable.

Classifying a patient as having active substance use or not using natural language processing presents a set of challenges compared to doing so in routine clinical settings. While clinicians may tailor questions and follow-up to discern whether a patient is currently using substances, the same is not true when querying data within unstructured notes in the electronic health record. Certainly, some patients will clearly have documentation of active use of cannabis in their chart, given 16% of adults endorse using cannabis in the past month.  On the other hand, many patients will also have documentation that confirms that they are not currently using cannabis, an important concept that may also be referred to as negation of cannabis use. Another group of persons will fall into a third category: review of their notes leads to an inability to either confirm or deny current cannabis use. Because of the uncertainty regarding their use, including persons with unknown cannabis use would lead to biased estimates in an analysis, and this group was appropriately excluded by the authors. Similar approaches may be used to classify patients and identify those with no use of any substances, which the authors also performed.

The need for natural language processing methods to identify persons with active cannabis use would be less of a priority if our health systems widely implemented the use of validated screening tools to ascertain whether and to what degree a patient uses risky substances. Alternatively, natural language processing could supplement validated screening tools. Tools such as the Brief Screener for Alcohol, Tobacco, and other Drugs (BSTAD) for adolescents and the Tobacco, Alcohol, Prescription medication, and other Substance use (TAPS) for adults provide efficient means to screen patients in both patient- and clinician-administered formats.  Answers to three or four screening questions unveil insights about the use of cannabis, tobacco, and alcohol that are typically more informative than content on substance use found in many “social history” sections. Both patients and surgeons appear receptive to the use of screening tools in the preoperative period.  Given gaps in the use of screening tools for surgical patients, natural language processing methods have the potential to unlock meaningful information within unstructured clinical notes in the perioperative setting, permitting a deeper phenotyping of patients beyond the more commonly used approaches based on structured data.

What are some considerations for next steps in natural language processing for substances in the operating room?

1. Identifying persons with active, recent, or past use of risky substances. Broad lines of inquiry that examine the associations and implications of active and past substance use on patient-centered and health-related outcomes represent one of the clearest possible benefits of deploying natural language processing methods. A more complete understanding of substance use includes the degree, timing, and frequency with which such exposures may create risk. This has the potential to drive additional investigations, increase awareness among patient and professional communities about how substance use may impact surgical outcomes, and foster proactive change in habits among patients, all of which may advance care for patients in the operating room. This can be particularly challenging with cannabis giving the lack of standardization across different products and formulations. For cannabis, complexity and variability arises from diverse products and individual use patterns, including varying cannabinoid composition. Products contain unique concentrations and ratios of delta-9-tetrahydrocannabinol and cannabidiol, among other active compounds, each of which is expected to have different mechanisms of action. Even among the same compounds, different routes of administration lead to important changes in the bioavailability and resulting side-effect profiles. Altogether, the varying physiologic effects are of high interest within the perioperative period, based on dose, frequency, and timing of exposure relative to surgery.
2. Mitigating biased and stigmatizing language in clinical documentation. Narratives and unstructured notes reflect the perspectives of clinicians who author them, including potentially biased or stigmatizing language. Among hypothetical cases, one could imagine interventions that recognize biased or stigmatizing language, bring this to the attention of a note writer, and suggest alternative language. Nudges to change the language that we use may advance efforts to reduce stigma encountered by persons who use risky substances, as well as other groups, and lead to improvements in their health along the way. In addition to stigma, confidentiality and patient privacy considerations arise given patients may not disclose their use status due to fear of legal repercussions, especially when substances may be illegal in their state.
3. Identifying persons in need of connection to care for risky substance use. Looking back to identify persons with high risk or disordered use of cannabis, opioids, alcohol, and tobacco products using reliable and validated natural language processing methods tailored to the perioperative setting is an appropriate starting place. Classifying patients for a retrospective study presents one set of challenges, including misclassification and bias. The consequences of misclassification increase significantly when developing risk prediction, where similar methods are applied to classify risky substance use for purposes not of research but of clinical care and real-time clinical decision-making. Using natural language processing tools to assist the screening process will require clinicians to verify and comprehensively assess the output of tools before making decisions and recommendations that directly impact the patient.
4. Addressing a broader range of risky substance use before surgery. The development and validation of natural language processing methods enable the examination of a greater breadth and depth of substances. For example, these methods boosted rates for correctly identifying risky alcohol use before surgery from 29% when using diagnosis codes alone to 87% when using natural language processing.⁸  Analyzing certain types of exposures within a substance use category may be possible to a degree that was previously not feasible. For example, distinguishing among the various routes of cannabis use may reveal differences in edible versus topical forms. Studying substances less commonly used than cannabis, alcohol, and tobacco products is also of interest, such as emerging drugs including xylazine or synthetic cathinones.

Natural language processing methods offer a timely and important tool to identify and quantify aspects of risky substance use. Applying robust methods to the perioperative setting will improve our understanding of how patterns of substance use impact the anesthetic management and perioperative outcomes for millions of patients who undergo surgery every year.

Every day in hospitals around the world, anesthesiologists face important questions about what treatments provide the best outcomes for patients. Advancements in methodology and data collection facilitate evaluation of different treatment strategies in real-world situations using observational routinely collected data. Unfortunately, even in large data sets with detailed patient and system characteristics, confounding and indication bias leave interpretation of estimates derived from observational comparative effectiveness studies uncertain. In fact, observational studies routinely estimate treatment effects that are directionally incongruent with subsequent randomized trials. Even when treatment effects are directionally congruent, observational studies typically overestimate the size of treatment effects identified in randomized trials.  These findings act as a clear reminder that randomized controlled trials remain the most important tool for understanding the true causal effects of novel and routine treatments, which highlights the clinical and methodologic implications of the Dexamethasone for Cardiac Surgery (DECS)–II trial reported this month in Anesthesiology. 

While in an ideal world we would have randomized trial results to inform all treatment decisions, conducting randomized trials is hard. Many barriers to successful trial conduct exist. Key barriers relevant to evaluating a routine treatment include (1) convincing centers to participate, which requires departures from their usual care; (2) having the resources required to enroll and randomize thousands of patients, the majority of whom may ultimately decide not to participate after staff have spent hours screening and contacting them; and (3) having the resources required to collect complete, valid, and meaningful outcome data.

Fortunately, there has been an explosion of innovative clinical trial methodologies that can help to overcome barriers to efficient conduct of randomized trials. In a seven-center, international randomized trial, Myles et al.  used several innovative techniques to conduct an efficient trial aimed at answering an important question in routine care of cardiac surgical patients, namely, does administering dexamethasone (1 mg/kg) lead to patients spending more days at home in the 30 days after surgery (primary outcome), having lower incidence of organ failure, less time in critical care, or better biomarker profiles?

First, DECS-II was a pragmatic trial meaning that the care provided to participants reflected what was done daily by providers and hospitals participating in the trial. In contrast to an explanatory design, where all aspects of care other than the treatment (e.g., induction drugs, bypass settings, ventilatory and hemodynamic management) are tightly controlled and protocolized patients in DECS-II received typical care from their providers, other than random assignment of dexamethasone. This approach allows the results of the trial to generalize more directly to real-world care.

Second, the primary outcome, days at home, as well as organ complications, were captured from routinely collected data. Use of registry-linked clinical trial methodology means that bias-reducing power of randomization can be used to efficiently generate causal treatment effect estimates by leveraging the scale and efficiency of data that are already being collected. While trialists (and readers) must always ensure that outcomes captured from routine data sources are accurate and valid, days at home is a validated, routinely collected metric, while national surgical registries can provide valid assessment of complications if standardized methods are used to prospectively capture events using trained assessors.

Third, to overcome the challenge of disrupting routine care with implementation of a randomized trial, randomization was designed to favor local practice. This means that even though allocation to corticosteroid treatment was random, centers in the Netherlands, where corticosteroids are almost always given, had two patients randomized to corticosteroids for each randomized to no corticosteroids, with the inverse approach carried out in Australian centers, where corticosteroids were rarely used. This approach supported acceptance of the study protocol with local providers, as without participation of study centers, multicenter trials cannot succeed.

Last, the investigators should be congratulated for having the adaptability required to conduct an international trial, where jurisdictional standards and processes can vary enormously. In Australia, where patients were already being routinely enrolled in a cardiac surgery quality assurance database, prerandomized consent (also called a Zelen design) was used. This meant that only patients randomized to dexamethasone (the local departure from routine care) were required to provide informed consent. In the Netherlands, all patients were required to provide preliminary consent. For anyone randomized to not receive dexamethasone (their local departure from usual care), further written consent was required. While these altered consent approaches provide efficiency in enrolling participants, they are not without risk. If patients choose to withdraw after randomization, serious bias can emerge.  Fortunately, the study team achieved low levels of attrition (less than 5% in each arm) and conducted multiple required sensitivity analyses that were consistent with primary results, substantially reducing concerns about bias.

Utilizing these innovative design strategies, what conclusions have emerged? Based on the combined results of three large multicenter trials of suprapharmacologic doses of corticosteroids in cardiac surgery to mitigate the inflammatory response to cardiopulmonary bypass some trends are evident (table 1). First, it appears unlikely that, in general, corticosteroids produce clinically meaningful improvements in outcomes and resource use. Second, any positive signals appear to emerge from use of dexamethasone whereas methylprednisolone was the only drug implicated in increased release of cardiac biomarkers and provided no signals of benefit.  For clinicians strongly in favor of dexamethasone administration, focusing use in subgroups of patients most likely to benefit (higher degree of critical illness, younger age, greater risk of postoperative pulmonary complications, female patients) may be an acceptable strategy.

For those who remain uncertain, but hopeful, that dexamethasone’s mechanistic promise will lead to improved outcomes, drawing on further examples of innovative clinical trial designs will be required. Clinicians intrinsically recognize that different patients often respond differently to the same treatment, a phenomenon known methodologically as heterogeneity of treatment effect. For steroids in acute inflammatory states, biologic mechanisms and some epidemiologic evidence suggest that heterogeneity in treatment effects is plausible. While typical approaches to heterogeneous treatment effects have involved exploratory approaches like testing for subgroup effects based on simple categorization of baseline variables such as age or sex (via tests of statistical interaction that are always underpowered), emerging evidence suggests that identifying optimal target populations for clinical trials likely requires a more granular approach. Specifically, advances in predictive modeling applied to large, high-quality clinical trial datasets, like those available from DECS-I and -II, can be used to identify the most likely group of responders based on combinations of baseline variables, including complex interactions.  With these predictions in hand, an enriched subpopulation where larger treatment effects were expected could be identified and randomized, providing robust evidence in a more precisely targeted population. Furthermore, with advancements in electronic health records and enterprise-based prediction models, such responder scores may be operationalizable via real-time decision support to facilitate translation of promising results into clinical care.

In our opinion, despite rapid advances in observational methods, randomized trials will remain our most important tool for evaluating treatments. However, an innovative environment where best practices in trial design are integrated with advancements in predictive modeling, and then leveraged together using high-quality routinely collected data and pragmatic trial design, holds enormous promise. For perioperative clinicians and patients, DECS-II provides crucial clinical and methodologic insights and should inspire further innovation in clinical trials that are likely to deliver increasingly robust and efficient answers to the treatment questions that we most highly prioritize.

For many Americans, the words “conscientious objector” conjure memories of an unpopular war and those who declined to fight, at the cost of freedom, careers, fractured family relationships, and hatred and threats. Yet the most famous conscientious objector, Muhammad Ali, explained that, in refusing to go to war in violation of his Muslim faith, “I have gained a lot…I have gained peace of mind. I have gained a peace of heart.”  Today, conscientious objection causes significant consternation in medicine. In this issue of Anesthesiology, Koganti et al.  discuss medical conscientious objections, and why anesthesiologists can and should accommodate them.

Federal rules and various state laws explicitly provide legal protection to physicians who refuse to provide care that violates their religious or personal moral beliefs regarding abortion, sterilization, contraception, physician-assisted suicide, and transgender care.  Many medical ethicist experts are concerned that vulnerable patient groups are particularly impacted by these issues: the poor, elderly, women, children, ethnic and faith-based minorities, and the LGBTQIA+ community.  Understanding why deserves our serious attention and study, but we set that aside for another time.

Understanding conscientious objection starts with defining what it is, and what it is not. Conscientious objection is more than a disagreement over medical treatments, discomfort with patient healthcare desires, or pragmatic difficulties in providing care. By definition, conscientious objections—whether due to religious beliefs or personal moral codes—are conflicts that threaten the core of a person’s moral identity. Conscientious objections, in other words, are about the soul. Being forced to provide care that creates deep personal moral dissonance presents real and serious dangers to the well-being of healthcare providers. 

Differences of opinion in health care are not uncommon. Physician refusals to provide patient treatments because of deep moral conflicts are fortunately more rare. And many conflicts termed conscientious objections are not.  Refusing to provide care that a provider believes is irrational, futile, or below professional standards is not necessarily a conscientious objection, for example, but a professional conflict that can be resolved through review of professional standards and treatment indications. Leaving individual providers to define “futility” or “lack of medical benefit” is highly subjective, and has been shown to invite discrimination, particularly of disabled patients.  Most conflicts in medicine do not rise to the level of conscientious objections the way that withdrawal of life-sustaining mechanical ventilation raises moral questions about killing versus letting die; that contraception and abortion raise moral questions about when human life begins or whether the life of a mother carries more or less weight than that of her potential child; or that transgender care raises moral questions about whether gender is a choice. Such questions are not medical ones, but moral ones, and present many physicians with deep conflicts over the answers, regardless of whether their ideologies arise out of religious tenets or personal moral codes, and regardless of whether they are “conservative” or “liberal” in nature. Sadly, the true nature of conscientious objection is becoming further obscured as it is increasingly co-opted by groups pursuing political ends rather than mutually respectful understandings.

Conscientious objections test us to respect our colleagues, and to manage their serious refusals with justice and compassion—but they do not absolve us from medicine’s foundational purpose and values: to set aside self-interest and put the patient’s needs first and at the center in our sphere of care. If a conscientious objection is impossible to accommodate without compromising patient well-being, precedence must be given to patient well-being over the provider.

Koganti et al. appear to take a position that only a medical emergency would justify not accommodating a conscientious objection. But vulnerable patients often face insurmountable barriers outside of emergencies in finding alternative caregivers, let alone negotiating significant limitations that will arise from broader healthcare workforce challenges created by institutional conscientious objections, even in major urban areas.

This latter concern is no longer hypothetical. In Washington state, for example, due to mergers of faith-based healthcare institutions, greater than 41% of hospital beds are held by organizations that conscientiously object to abortion, contraception, sterilization, many forms of transgender care, and some forms of end-of-life care. Even in large urban areas, such as Seattle and Bellingham, significant shortages and sometimes entire losses of these services have resulted.  As faith-based healthcare conglomerates grow, institutional conscientious objection will take bigger and bigger bites out of the availability of such care for patients who do not necessarily share their providers’ beliefs but have no place left to go.

The objector, too, has responsibilities: to avoid when possible situations in which they will predictably feel morally compromised, and also to provide alternative caregivers when they recuse themselves so that patient well-being is not compromised. Conscientious objectors should consider how their practice choices may impact their own well-being, just as such considerations have been examined by providers with strong convictions to provide abortion care who find themselves in settings with restrictions on reproductive choice, for example. These environments impact the provider’s ability to enjoy their work and have caused many to reconsider their location and practice setting. The Jehovah’s Witness Church calls upon believers to “carrying your own load of responsibility,” and seek alternative employment when they have moral objections related to their work.  In anesthesiology, we have broad flexibility in the focus and type of practice we decide to pursue, making many conscientious objections potentially avoidable. Whenever possible, conscientious objectors should provide known objections in advance.

The American Society of Anesthesiologists (Schaumburg, Illinois) Committee on Ethics has provided some guidance. The Guidelines for Ethical Practice of Anesthesiology adopts the American Medical Association (Chicago, Illinois) principles of ethics, which in turn were developed into multiple guidances, including conscientious objection.  The “Statement on Ethical Guidelines for the Anesthesia Care of Patients with Do-not-resuscitate Orders” requires conscientious objectors to provide a suitable alternative provider.  Granular details of how practices manage conscientious objections will necessarily depend on specific conditions of the practice: additional barriers to patient care should in no case be endorsed by the American Society of Anesthesiologists.

While threats to patient autonomy and reproductive justice are pushed in political and legal arenas, it is important to distinguish those conflicts from the more personal battlefields that physicians face, both in providing care to patients and addressing the very real moral concerns of their colleagues. Our specialty has been significantly impacted by burnout and issues concerning well-being. We would be remiss not to consider this when addressing the impact of care we provide in the context of conscientious objections.

We posit that diversity and inclusion within the physician anesthesia workforce will be critical in improving the impact of conscientious objection. Healthcare disparities are already well documented, and those who are members of vulnerable populations, whether they be racial groups, LGBTQIA+ communities, migrant populations, or those to whom English is a second language, would benefit from care provided by individuals with similar lived experiences, cultural humility, and shared values, and who therefore may have fewer conscientious objections to their care. Broader diversity and inclusion also create spaces for providers of true belonging and collegiality by helping us to experience and better understand our ideological differences. 

Our specialty is at a critical juncture as more is being asked of us despite workforce shortages. When leaders across the specialty consider short-term and long-term goals, we call upon them to prioritize the training of a diverse and inclusive anesthesia workforce to promote well-being among anesthesiologists who have conscientious objections. As our specialists occupy larger leadership roles within health systems, we must also advocate broadly for a diverse and inclusive physician workforce across the field of medicine.

The Special Article by Kanjia et al.  in this issue of Anesthesiology provides an interesting analysis of the role of various theories of safety, drawn mostly from those primarily created for nonmedical arenas. Some of these theories have been adapted in part by various medical organizations and institutions, mostly addressing experienced personnel and their work. This article extends notions of other authors about ordering the theories, promoting a climb from Safety 0, which is aimed at clinicians’ capabilities; to Safety 1, in part using “high reliability organization theory”; Safety 2, which entails largely concentrating not on “errors” but on how things work well; and Safety 3, which concentrates more on the entire “system.” The notion of a strict hierarchy of theories is relatively new, although the hierarchy presented here acknowledges that beyond presenting novel issues, safety does require merging the levels.

Overall, the authors’ analysis is useful, especially to alert the anesthesia community to this specific collection of theories. Some components of the theories, especially for Safety 0, are already known to anesthesiologists, but many aspects of Safety 1, 2, and 3 are not. In particular, they call special attention to the 2020 publication about the new Safety 3 by Massachusetts Institute of Technology (Cambridge, Massachusetts) Professor Nancy Leveson in a very long (110 page) internal article, “Safety III: A Systems Approach to Safety and Resilience” (https://sunnyday.mit.edu/safety-3.pdf). This author, known best for earlier work on software safety, provides a conception of Safety 3 as the need to simultaneously address and unify safety approaches of complex hardware, software complexity, and a variety of human factors issues.

Scholars inside and outside health care have for a long time investigated and written about patient safety, especially in anesthesiology, and have analyzed or created a variety of “organizational safety theories” (some included in the article by Kanjia et al.¹  but also many others)  There are too many to describe, with most originally applied in domains like aviation, human spaceflight, aircraft carrier flight deck, nuclear power, open-water oil drilling, chemical manufacturing, and so forth.

Health care overall seeks to provide excellent care for all patients, but anesthesiology is unusual in many ways. It is highly dynamic (things happen in seconds, minutes, hours) while also being of “high intrinsic hazard,” with risks arising by its very nature, even without errors. For example, evolution did not intend for human beings to be rendered (1) deeply unconscious, (2) insensitive to pain, (3) substantially amnestic, or (4) having deep neuromuscular blockade.

Other unique aspects of anesthesiology are that (1) many anesthetic cases require special and careful attention to complex surgeries or to serious pre-existing medical condition(s) or both; (2) anesthesia care is often conducted by a single anesthesiologist, sometimes without a guarantee of help if needed; (3) anesthesiologists are typically their own pharmacist, technician, cardiologist, and so forth; (4) many medications used in anesthesia can be lethal; and (5) the work often has intense production pressure for speed. Yes, anesthesia care has become remarkably safe, but it is never simple.

These aspects apply to the “sharp end” of clinical care in anesthesiology, but additional system issues arise from the surgery–anesthesia–institution organization. In fact, every level of an organization has its many sharp ends—things that need to be accomplished—under the control of its own “blunt end” that oversees/drives the work. This goes all the way from the clinicians caring for patients through various levels of institutional systems, even to those imposed by national or state politics. Moreover, healthcare systems themselves differ in a number of ways from systems outside of health care, including the following:

• I.Knowledge about the system’s primary target(s). In all but health care, the most relevant objects and devices are designed and produced by human beings and are well understood. In health care, we do not “design” or “build” human beings, nor do we ever receive a complete “instruction manual.”
• II.Underlying characteristics of organizations
  • a.Centralization versus decentralization. Many nonhealthcare endeavors have substantial centralization. Taking nuclear power for example, there are only 94 reactors in the United States, operated by 54 plants, in 28 states (with approximately 20,000 nuclear technicians and nuclear engineers in the United States or ~250 at each plant). As another example, each day U.S. commercial aviation flies many thousands of aircraft (each with two cockpit crew members) between hundreds of major airports; yet, only 10 U.S. airlines account for nearly all of the passengers carried annually; the top five alone carried 84%. In contrast, health care, despite a growing amalgamation of institutions, is extremely decentralized. On any given day in the United States, there are millions of patients cared for by millions of clinicians and others in approximately 6,000 hospitals and a far larger set of clinics and independent sites.
  • b.Control by and of the firms. Most industries use commercial “firms,” each of which exercises substantial control over many aspects of how the work is done, while staying within the oversight rules from regulatory bodies (some federal, some state, some nongovernmental). In decentralized health care, the exact number of “firms” is unknown but certainly very high. The different firms and their subordinate units, if any, create policies and procedures, but primarily physicians control the diagnosis and treatment of patients. Some rules are imposed by accrediting bodies, or state and federal agencies, but here too, clinicians retain the bulk of the authority.
  • c.Harm to the means of production. All arenas of high intrinsic risk fear harm to life safety for their participants and workers and consider the financial burdens of litigation and “bad press.” However, for example, no airplane is ever “supposed” to crash. Most industries also fear that an accident might seriously harm or destroy one or more critical parts of its “means of production.,” e.g., no more passengers can ever be carried on an airplane that has crashed. Collectively, the multiple concerns provide substantial, albeit imperfect, financial incentives to pursue safety.

On the other hand, all humans will die, and in high-resource places, most will die in connection to health care. It can be difficult to tease out which events were avoidable errors versus inevitable due to clinical condition. However, drawing conclusions of unavoidability may mask situations in which better attention to safety might have saved someone, or at least reduced the risk for others in the future. In addition, somewhat cynically, I posit that health care may see fewer incentives to pursue maximal safety efforts in part because we rarely, if ever, harm the institution’s means of production.

In summary, the article by Kanjia et al.  provides one entry into the complex arenas of patient safety. All of the various theories they discuss are meaningful, whether created for healthcare settings or adapted from nonhealthcare settings. I urge everyone involved significantly in patient safety to read this article but also to dig deeper into the many other safety theories that have been developed or adapted by a wider cohort of scholars from outside and inside health care. A detailed discussion of applications to anesthesiology of human factors and many patient safety theories can be found in Miller’s Anesthesia.  Additional relevant work includes Gaba et al., Tamuz and Harrison and Leveson et al. 

The more each of us can learn about why, what, where, and how to deploy the various patient safety approaches—singly or in conjunction with others—the more likely we will be able to better serve our patients and our institutions. In doing so, we pursue the vision statement of the Anesthesia Patient Safety Foundation: That no one shall be harmed by anesthesia care.