Chapter 25. Anesthesia Risk

1. Anesthesia risk estimates are influenced by the circumstances in which they are generated; estimates developed in one clinical setting or selected population may not be relevant to other settings or specific patients.

2. The risk of death related primarily to anesthesia is estimated to be as low as approximately 1 in 200,000 anesthetics in some large populations and reflects improvement of perhaps 2 orders of magnitude over more than 60 years.

3. Despite the very low anesthesia-attributable mortality rate, the very large and increasing number of anesthetics engenders a substantial burden of mortality and morbidity, much of which may be preventable.

4. The very low anesthesia-attributable overall mortality should not be a cause for complacency but rather an impetus toward a greater emphasis on improving anesthesia-related morbidity in selected subpopulations of high-risk patients or high-risk procedures.

5. Approximately 75% of risk actually relates directly to patient-specific characteristics, including age, gender, and comorbidity; 20% to surgical issues, such as experience and judgment; and the remaining 5% to anesthesia factors, including experience, board certification, pharmacologic issues, and overall management of care.

6. The American Society of Anesthesiologists physical status classification correlates with risk for mortality and morbidity, but it is somewhat subjective and, without additional clinical information, is alone not as strong a predictor of poor outcomes as other, morbidity-specific measures.

7. Although randomized clinical trials have rightly become the gold standard for establishing efficacy in clinical research, randomized clinical trials have a limited role in studying anesthesia risk, particularly because they cannot efficiently and at reasonable cost identify the confounding clinically relevant variables.

8. Well-conducted observational studies reveal that anesthesia risk is influenced much more by how the anesthesiologist provides care rather than specifically what methods are used. Thus careful attention to such factors as blood pressure, perfusion, oxygenation, body temperature, and depth of anesthesia, for example, are often more beneficial than the specific anesthetic drug or technique that is selected.

"What is the risk of anesthesia," patients often wonder. Their concern really is multipart. First, how likely is something to go wrong; then, what "bad things" could happen; and, finally, what can be done to reduce the risk to me and to others following me in the future?

Anesthesia care providers try to allay patients' concerns by citing statistics based on large populations or reassuring that common anesthesia side effects, however annoying, are fortunately transient. Serious, life-threatening complications now are uncommon, if not rare. Such comments may offer comfort, but they do little to specify the magnitude and types of risks that a given individual faces, nor do they indicate that we know how to reduce the risks. The current national focus on patient safety leads to expectations that we provide more to our patients, and this chapter will provide such information.

The concern for enhanced safety has special import for anesthesia care because anesthesia care usually confers no therapeutic benefit but rather facilitates other therapeutic or diagnostic interventions. The potential for harm is expressed quantitatively by a variety of risk metrics, most commonly as the incidence rate for an adverse event. Such metrics enable us to estimate the contributions of various aspects of anesthesia care (eg, anesthesia methods and drugs, clinical care sites, temporal periods) to overall anesthetic risk, allow clinicians to formulate care strategies that are evidence based, and form the basis for comparative effectiveness studies that evaluate new drugs, anesthetic methods, and specific procedures. These diverse efforts have included inquests and closed malpractice-liability claims, case reports and series, study commissions, registries, cohort studies (some with multi-institutional and geographically diverse data), clinical trials, clinical practice guidelines and standards, establishment of the multidisciplinary Anesthesia Patient Safety Foundation, and use of simulation for clinical training, among other initiatives (see Chapter 3).

The intensity of these myriad efforts documents the specialty's commitment to decreasing the occurrence of adverse events and thereby improving anesthesia safety. An almost 100-fold decrease in estimates of anesthesia-attributable mortality among studies over the past 6 decades1 demonstrates the specialty's high level of professionalism and leadership toward ever-greater patient safety, which has been cited by the Institute of Medicine as a model for the rest of health care.2

The risk of death related primarily to anesthesia has been estimated to be as low as approximately 1 in 200,000 anesthetics in large populations.1,3,4 (This chapter emphasizes that such estimates vary greatly depending on circumstances of the given study, and that no rate can be considered representative of all clinical settings.) Yet such figures should not engender complacency among providers; as we will show, the death rate is still 20 to 30 times greater than that of air transportation or working in the construction industry, for example. An estimated 39 to 48 million surgical procedures5-7 in the United States, plus an uncertain number of diagnostic procedures undertaken with anesthesia, results in several hundred anesthesia-associated deaths each year. Many more—perhaps 10 to 15 times as many—perioperative deaths occur in circumstances where perioperative anesthesia management, surgical care factors, and/or patient-specific characteristics jointly contribute to poor outcomes. Moreover, an even greater prevalence of anesthesia-associated morbidity poses a substantial burden of individual suffering and societal costs. Thus, despite substantial improvement in anesthesia safety, potentially preventable anesthesia-associated mortality and morbidity remain "of sufficient magnitude to constitute a public health problem," as Beecher and Todd8 noted almost 60 years ago.

This chapter surveys the factors that the practitioner needs to understand both to gain a greater understanding of anesthesia risk and to interpret the literature on anesthesia risk appropriately.

The concepts risk and hazard are sometimes confused: A hazard has the potential to cause harm; examples include stairs, automobiles, motorcycles, hospitals, hypertension, and anesthesia. In contrast, risk refers to the likelihood of harm resulting from a hazard: Risk is the probability of an event within a population during a specified period of time.9 Although risk is ubiquitous and a natural part of life, it is not uncontrollable; it can be mitigated by knowledge, attitudes, and behaviors. Typically, risk is influenced by a variety of "risk factors," including especially human behavior, equipment, colleagues, fatigue, location, and experience. In common usage, anesthesia risk connotes negative occurrence, such as death or injury.

Risk quantifies the occurrence of a specific outcome within a specified population that is exposed to a specific hazard (eg, anesthesia, air travel) under specific conditions. For example, the population receiving general anesthesia might share exposure to a hazard (eg, loss of protective airway reflexes) during the period of observation; however, only some among the population develop an adverse outcome (eg, aspiration pneumonitis) that is associated with this hazard. Although the definition of risk is precise, risk estimates can vary widely to the casual observer. Such variation usually results from differences in study design, not from differing definitions of risk. Careful reading of a report often reveals possible sources of the differences (Box 25-1).

Although the bulk of the anesthesia risk literature relates to estimates of risk (including identification of risk factors), an important subset of our risk literature relates to the other, less common usage of risk: descriptions of specific adverse outcomes. Examples include case reports, case series (eg, critical incident studies10), and registries such as the American Society of Anesthesiologists (ASA) Closed Claims Project.11 Because such reports do not include information on the "at-risk" or "exposed" population, incidence rates cannot be estimated and risk factors for the occurrence of the outcomes cannot be identified. Yet such reports are widely used to alert us to possible, but still unquantified, risks, and they have been especially useful in developing initiatives to improve safety (see Chapter 3 and this chapter).

How Does Anesthesia Risk Differ from Anesthesia Safety?

Risk and safety often are discussed colloquially as if they were closely related, yet they are not. Whereas risk is an objective, probabilistic term, safety connotes personal and/or social value judgment about the acceptability of a given level of risk.12,13 We generally regard as safe that which poses an acceptable level of risk, given that nothing is risk free.13,14 The acceptability of a given risk level is a function of each individual's risk tolerance, which, in turn, is influenced by factors characterizing the adverse outcome, exposure, and perceptions of the individual evaluating the risk (Box 25-2).12,15 The tolerance for risk varies considerably among individuals. Not surprisingly, parents are generally much more concerned about the safety of anesthesia for their children than for themselves, even though the risk of injury is generally similar.

Models for Studying Anesthesia Risk

Approaches to studying anesthesia risk have evolved over time. Most investigators who published before the late 1980s were guided by a restrictive model that focused almost exclusively on serious clinical outcomes (deaths, adverse events; see Fig. 25-1). In this approach, a population of patients is "exposed" to the dual "hazards" of anesthesia and surgery, and the model counts clinical outcomes that result from this experience.

A newer model (Fig. 25-2), consistent with studies of anesthesia risk undertaken since the latter 1980s, recognizes that anesthesia care and surgery occur in a complex matrix of administrative, cultural, and clinical processes, all of which can add hazards to the system. This model also includes a broader array of outcomes that include cost, functional status, and patient satisfaction (Box 25-3). These assume greater importance now with the growing call throughout medicine for greater patient-centered care (see Chapter 3).

Quantifying Anesthesia Risk

If exposure to a "hazard" increases the likelihood of an adverse outcome, then duration of exposure should be a better predictor of a negative outcome than just occurrence of exposure. Indeed, this intuitive notion is the basis for the investigative approach in toxicologic and epidemiologic studies of health hazards. Time-related exposure risks (eg, 1-year odds) of real-world health hazards have been tabulated.15 Runciman and Moller17 applied this notion to quantify the mortality risk of anesthesia and complex surgery in relation to a variety of real-world hazards (Box 25-4). Although the mortality rates are only approximations, the relative risks are informative for both providers and patients, who can more clearly understand that anesthesia is 20 times more hazardous than commercial flight but 10 times safer than parachute jumping!

Unfortunately, there are few anesthesia risk studies examining time-related hazards. Many studies note an association between greater duration of anesthesia (or surgery) and poorer outcome, but longer anesthesia also reflects longer and/or more complicated surgery, making it difficult to separate the independent role of anesthesia. However, recent research documents that anesthetic depth—or cumulative deep hypnotic time, estimated as the cumulative duration of low Bispectral Index Sensor (BIS <45)—is an independent predictor of postoperative18-22 and criticalcare mortality,22 particularly in combination with patient comorbidity and intraoperative hypotension.

In the absence of time-related risk data, anesthesia risk is commonly summarized as an incidence rate computed as the quotient of the number of cases and the overall population:

Four important features of this simple relationship are apparent. First, a risk cannot be estimated without knowing the size of the population from which the cases (ie, those with the outcome) arose. Thus the absence of the denominator precludes estimating risk in studies using case series and registry data, and imprecision in estimating the size of the population influences the computed risk.

Second, risk estimates can range from 0 (ie, no one experiences the outcome) to 1 (ie, everyone suffers the outcome). Depending on the outcome studied, anesthesia risk estimates span much of this range, from death primarily related to anesthesia care in large unselected populations (eg, as low as 0.000005) to postoperative nausea in especially susceptible individuals under specific circumstances (eg, as high as 0.8). Because decimal values can be unwieldy and awkward in conversation, we commonly refer to risks as percentages (eg, the risk of perioperative death in relation to cardiac surgery is "3.0%" rather than "0.03"). For smaller risk values, we often express the risk in either the scientific notation (eg, 5 × 10−6 rather than 0.000005) or, more commonly, the ratio of 1 to the smallest number of individuals whose exposure to the hazard would result in one case (eg, risk of death primarily related to anesthesia is "1 in 200,000" or "1:200,000" rather than "0.000005").

Third, by definition, the calculated risk is the risk faced by the population under study, whereas we are often interested in the risk faced by a particular individual or perhaps a specific group of individuals who were not part of the study population and may have special characteristics that might place them at greater or lesser risk. Thus risk estimates in the literature, especially some very low mortality rates widely touted in public media, may not relate well to the person(s) about whom we are concerned. Again, risk estimates are very much the products of the circumstances in which they are generated (Box 25-1).

Fourth, this simple and common risk calculation is but one approach to expressing the magnitude of risk. The magnitude of risk can be expressed in alternative ways, some of which are particularly useful in understanding clinical trial results and comparing treatment benefit or harm across studies (Table 25-1).23 Such risk metrics are easily computed from summary results presented in the literature (Box 25-5).

Risk Values Are Estimates

Risk values are estimates, and estimates from small samples can vary greatly. For this reason, randomized clinical trials usually involve thousands of patients and, even then, we may learn of new risks in subsequent postmarketing surveillance. Small trials that report "no complications" or similar are especially suspect, and unfortunately small trials are all too common in the medical literature, especially for new procedures (new drugs tend to be evaluated more fully). Thus our personal confidence in small trials or case series should be suspect,27 and this is expressed mathematically as the confidence interval (CI).

As the sample size increases, the CI for the estimated rate narrows, enhancing the precision of that estimate (Fig. 25-3 and Box 25-6).28 Thus, other things being equal, we regard rates calculated from larger cohorts as more reliable than those from smaller studies, and the statistics support our intuition.

Generic Challenges in Studying Anesthesia Risk

Identifying the Outcomes of Interest

Diagnostic Problems

Diagnosing or identifying an adverse event is critical to the assessment of risk and not a trivial problem. Identifying the occurrence of death might seem inordinately simple, but postanesthesia death may not be recognized if the defined surveillance interval is brief, such as the first 48 hours postoperatively or only during hospitalization. This ascertainment bias results in underestimation of the true mortality rate. Even greater challenges are posed by adverse events that present in a graded fashion, from barely detectable to life threatening. Thus differences in incidence rates can be influenced greatly by the prevalent diagnostic methods or definitions. For example, the observed rate of myocardial ischemia is greater if the diagnosis is made by continuous Holter monitoring than by clinical symptoms such as angina. Such imprecision constitutes an important source of systematic error (ascertainment bias, information bias) that can lead to failure to recognize outcome occurrences (misclassification bias) and inaccurate risk estimates.

Diagnostic challenges often are compounded by problems with the definitions and classifications inherent in clinical information systems and administrative databases. Institutional clinical information systems may use nonstandard definitions for data elements, such as adverse events, diagnoses, surgical procedures, and thresholds for abnormalities among laboratory values. Administrative databases, which classify the diagnoses, complications, and surgical procedures relating to care in a standard procedure coding system, are subject to institutional variation in coding practices and typically have few anesthesia-specific categories (eg, "E" codes in the International Classification of Diseases, Ninth Revision, Clinical Modification [ICD-9-CM] system), and some can be quite arcane (eg, ICD-9-CM code 968.3: poisoning by intravenous anesthetics).

Surrogate End Points

Even if the outcome of interest (eg, death, specific adverse event) can be identified accurately, it may occur so infrequently that a proposed study may not be feasible. Thus investigators have opted for alternative end points that are surrogates or intermediate variables—symptoms, signs, clinical events, laboratory values, and even process variables—which occur more often and are believed to be linked to that outcome. Elevated blood pressure is used as an end point in clinical trials because high blood pressure is a known risk factor for stroke and myocardial infarction (MI); similarly, myocardial ischemia is commonly used as a surrogate for MI.

Such approaches are appropriate only if the surrogate end point truly predicts the occurrence of the outcome. However, this requirement is often not met, possibly because the outcome may occur via some pathway(s) not mediated by the surrogate.31 Among recent examples is use of ventricular ectopy as the surrogate end point in the evaluation of drugs to prevent sudden death. Food and Drug Administration approval of a new class of antiarrhythmic drugs followed clinical trials demonstrating efficacy of such drugs in treating ventricular ectopy, but a subsequent trial revealed a greater risk for sudden death among those receiving the new drugs, and the drugs were withdrawn.32 Problematic surrogate end points in anesthesia have included the interval to eye opening after general anesthesia, which does not correlate well with earlier discharge after ambulatory surgery, lower rate of unplanned hospital admission after ambulatory surgery, greater patient satisfaction, or overall lower cost of care.33

Broadened Array of Outcomes

End points of interest have expanded from "clinical outcome" to a more comprehensive set of "outcomes" that reflect many facets of the care process, particularly factors that are important from the patient's perspective, or administrative end points such as the cost of services (Box 25-3). Studies of anesthesia risk already include some nonclinical outcomes, such as unplanned hospital admission after ambulatory surgery34,35 and poorer quality of immediate postoperative recovery.36 Patient-reported outcome measures37 include health-related quality-of-life38,39 scales (eg, Medical Outcome Study's SF-36 Health Survey40) and disease- and condition-specific specific scales (eg, Visual Analog Scale for Pain,41 McGill Pain Questionnaire42). Such instruments have been validated in large populations, typically with chronic disease, for use in longitudinal and cross-sectional studies,43 and they have been used in chronic pain clinics.

Patient-Specific Descriptive Data

Risks are not distributed randomly. Some individuals are more likely to experience health problems as a result of diverse factors, including genetics, coexisting disease, personal health behaviors (eg, smoking, poor diet, sedentary living), socioeconomic status, and environmental setting. Thus estimating anesthesia risk requires accurate information to characterize each patient, including nonclinical information. Such information enables categorizing patients by likelihood of poor outcome (risk stratification) and "risk adjusting" their data for outcome comparisons.44

The ASA physical status classification originally was advanced to standardize terminology and group patients45-47 but subsequently was found to correlate with clinical outcomes.34,48-51 In addition, the Charlson Comorbidity Score,52 a disease-specific classification system, is increasingly used in anesthesia risk studies. With growing interest in health care disparities, noting the patient's race and ethnicity accurately is important, typically using categories from the US Census Bureau (white, black or African American, Hispanic or Latino, Asian or Pacific Islander, or American Indian or Alaskan Native).53 Differential selection for surgical procedures is associated with race: Blacks are more likely to receive nonoperative care for hip fracture, and they experience increased survival compared with whites who receive operative care.54

Duration of the Surveillance Interval

Intuitively, we know that too short a surveillance interval may miss some occurrences, leading to systematically underestimating risk. Yet, as the time between the "exposure" (eg, administration of anesthesia) and event increases, the greater the chance that the occurrence may be associated with circumstances partially or totally independent of the exposure. These other circumstances might include even chance patient-specific factors (eg, MI that might have occurred 2 weeks preoperatively while walking the dog). This dualism of comprehensiveness versus precision plagues most cohort studies in the anesthesia risk literature.

Variation in the surveillance interval is especially apparent in the early studies of anesthetic-associated mortality. Some considered a death "anesthesia related" only if it occurred during or within 24 hours of administration of the anesthetic or after failure of the patient to regain consciousness after anesthesia.55-58 Others extended the surveillance interval progressively to 48 hours, 3 days, 6 days, 7 days, 30 days, 1 year, and 2 years.18,59,25,60-63 Beecher and Todd8 adopted an empirical approach, using the duration of the surgical hospitalization.

Warner et al64 found that 39% of major morbidity occurred more than 2 days but within 30 days after ambulatory surgery among 45,090 adults, including 2 of 5 cases of respiratory failure, 5 of 14 cases of MI, and 3 of 5 cases of pulmonary embolism. Although results could not be shown to be statistically different from those of the general population due to sample size, it is clear that substantial numbers of events would have been missed had the surveillance interval been limited to the immediate postoperative period.

There are no formal guidelines on duration of the surveillance interval, yet 30-day duration was adopted by the Veterans Affairs (VA) National Surgical Quality Improvement Project (NSQIP),65 and this duration is commonly used with Medicare data. For most procedures, this represents a reasonable compromise between too short an interval, which misses adverse events, and too long an interval, which introduces extraneous events.

Assessing the Relationship with Anesthesia Care

Problem of Attribution

Attributing causation is inherently problematic. Most studies of anesthesia-associated mortality have included a panel of blinded case reviewers who were asked to attribute the death to one of several possible causes: Was anesthesia care the primary cause of death? Or was anesthesia management a contributory cause (along with patient- and surgery-related factors)? Or was the poor outcome unrelated to anesthesia care? Panels were often composed mostly or wholly of anesthesiologists, but some have been more inclusive, including surgeons, obstetricians, referring physicians, and others.66 The latter seem more appropriate, especially given the current emphasis on team-based care.

Unmasking Biases

The mere occurrence of a poor outcome that may be related to anesthesia management can bias our judgments regarding quality and appropriateness of care.67 The belief that a poor outcome is presumptive evidence of error on the part of anesthesia providers is a bias that pervades much of the early anesthesia-risk literature.68 Notably, high interrater reliability in peer review does not ensure that judgments are accurate, for review panel determinations are necessarily products of their collective current knowledge of practice at the particular time, as well as numerous other biases that result from inadequate study designs, limited statistical capabilities, and the complexity of phenomena being studied.

Characterizing the Population

Our discussion thus far has focused largely on patients experiencing an outcome of interest (becoming "cases") after exposure to a "hazard." However, if we do not give due consideration to the larger "at-risk" population from which these cases emerge, we have only a set of cases and cannot quantify the risk.66

Ecology of Care

A large population is not homogeneous with regard to health status or the need for health care services.69 Rather, there is a hierarchy in the need for health services and, by implication, poorer health status or greater acuity of illness increases sequentially from the unselected general population to outpatient care settings to community hospital settings, and finally to the academic medical center (Fig. 25-4A).70 Surgical care is delivered in a similar spatial hierarchy (Fig. 25-4B). Thus risk estimates derived in one setting may differ from those in another environment (Berkson fallacy71). Recognizing these relationships can help reconcile seemingly disparate risk estimates from studies undertaken in widely differing clinical settings.

Statistical Problems in Studying Anesthesia Risk

Dealing with Low-Incidence Phenomena

Although some anesthesia-related adverse events may occur frequently (eg, postoperative nausea), most severe events do not. The incidence rates may be so low that estimating rates of occurrence with acceptable precision (ie, narrow CI) or demonstrating meaningful differences in occurrence (eg, comparisons of drug regimens or sites of care) can be problematic (Fig. 25-5 and Box 25-7).

The most obvious remedy is merely using a larger sample size. Yet, even when feasible, a larger sample usually connotes higher cost and a longer patient accrual period, during which other countervailing circumstances may arise (eg, deterioration in quality of data collection because of study team turnover or loss of enthusiasm, emergence of better technology that obviates study, information that alters or invalidates earlier results, changes in concurrent care processes that affect the event being studied). Other options include using a more sensitive end point, repeated measurements, matched cases and controls, multiple controls, lower statistical certainty ("power"), a higher-risk population, or concurrent multicenter trials, or some combination of such designs.

Multi-institutional studies, such as the 10-hospital Beecher-Todd study,8 34-hospital National Halothane Study,73 or the 460-hospital study by Tiret et al57 in France, offer the added benefit of enhancing the external validity (applicability of the results to other settings, termed generalizability in epidemiology) of the results because cooperating institutions may have different patient populations, surgical case mixes, clinical care processes, and/or administrative characteristics. Yet multi-institutional designs may introduce additional challenges, such as adherence to study design and definitions. Alternatively, investigators may opt to use a more frequent surrogate end point (eg, myocardial ischemia) that is associated with the outcome of interest (eg, MI), but, as noted earlier, the relationship between the occurrence of the surrogate and outcome may not be sufficiently close for valid results.31,33

One especially productive multi-institutional approach involves a consortium jointly collecting large amounts of data prospectively for a series of studies. An example was the Multicenter Study of Perioperative Ischemia Group that began in the late 1980s and comprised as many as almost 200 hospitals worldwide during its 20-year life, each contributing the same data elements to a database on patients having cardiac and noncardiac surgery. This group studied important perioperative risk topics, including the long-term risk of cardiovascular events,74 atrial fibrillation after cardiac surgery,75 mortality reduction with prophylactic use of β-adrenergic blocking agents,25 aspirin-related mortality from cardiac surgery,76 and risks related to use of aprotinin in cardiac surgery.77

Similarly, in the early 1990s, the dozen US Department of Veterans Affairs hospitals that performed cardiac surgery joined to create a common database78 that became the model for an expanded project covering all major surgery in more than 120 VA hospitals, the VA NSQIP.79 Using project-specific, trained medical-record abstractors in each hospital, NSQIP collected prospectively defined clinical data that have contributed greatly to measuring risk, outcomes, and quality of care.65,79,80 Recently, it was adopted by the American College of Surgeons (ACS), as the ACS-NSQIP, a collaboration of more than 200 academic health centers and large community hospitals. Although anesthesia-specific data elements comprise little more than the principal anesthesia method, the VA/ACS NSQIP database includes comorbidities and postoperative complications for thousands of patients having each surgical procedure and is beginning to enhance our understanding of preoperative risk assessment and choice of anesthesia method.81-84 Similar comprehensive clinical databases relating to cardiac surgery have been established by the Society of Thoracic Surgeons85 and the New England Cardiovascular Disease Study Group.86

As clinical information systems become ubiquitous, other clinical databases will arise from customary clinical care. In turn, vast institutional databases will result, creating innumerable opportunities for observational studies not even imagined by the early pioneers who first used computers to develop crude estimates of anesthesia risk.63 Examples include the various databases maintained by the University HealthSystem Consortium, a collaboration of 107 academic health centers and 232 affiliated hospitals; member institutions can benchmark their clinical results for performance improvement and enhanced patient safety. Challenges awaiting those exploiting such clinical treasure troves include ensuring that information systems use standard terminology for diagnoses, surgical procedures, complications, and other clinical information; adequate software to merge data from different databases in real time; availability of sufficiently robust risk-adjustment methods for meaningful and valid comparisons; user-friendly software to facilitate use of such databases without reliance on computer technicians; and uncertainty whether results generated in large and academic health centers can be generalized to smaller and rural hospitals that often lack resources to participate or invest as fully in required technology.

An alternative source of multi-institutional data has developed as a by-product of the payment system for health care: the Medicare Provider Analysis and Review inpatient ("Part A") file and the federal Healthcare Cost and Utilization Project's all-payer Nationwide Inpatient Sample. Although offering the benefit of millions of patient encounters at much lower cost than clinical databases, claims or administrative data have very limited clinical information. Moreover, challenges in merging this hospital file with other files (eg, physician or "Part B" file) and numerous coding issues, including inconsistency and imprecision (eg, is a given coded disorder a comorbidity or complication?), plague efforts to risk adjust outcomes data.44 Nonetheless, claims data have been used to explore risk in relation to medical direction of anesthesia care,87 board certification of the anesthesiologist,88 and location of surgical care.35,89 Such data have also enabled identification of patient and hospital characteristics that influence anesthesia time for common major operations.90 Recently, investigators have developed robust 30-day postoperative mortality and morbidity risk predictors based on readily available claims data—the surgical procedure code, patient age, and ASA Physical Status category—that rival the accuracy of NSQIP risk models.91 Similar risk indices based solely on procedure and diagnosis codes from claims data perform almost as well as those developed with added patient demographic information.92

Hybrid data collection combining claims and clinical data sources are also developing: ASA has recently established the Anesthesia Quality Institute to foster the development and use of a National Anesthesia Clinical Outcomes Registry based on claims and other data volunteered by anesthesia practitioners.

Finally, public health data provide another opportunity to study low-incidence events: The occurrence of death and the chain of events leading to death, including anesthesia-related complications and comorbidity, are entered on multiple cause-of-death certificates, whose information is coded, using the newer, more detailed ICD-10 classification system, and entered into the National Center for Health Statistics' National Vital Statistics System. Using its anesthesia-related ICD-10 codes, with estimates from NCHS's National Hospital Discharge Survey, Li et al in 2009 obtained estimates of anesthesia-related mortality risk similar to those derived in recent studies.1

Dealing with Covariates and Confounding

Clearly, anesthesia care occurs in a multifactorial setting. The outcome of each episode of anesthesia care—whether administration of anesthesia in the operating room, critical care in an intensive care unit, pain management, or other care—ultimately reflects the interaction of various patient characteristics ("covariates"), including age, gender, and health status (eg, ASA physical status, other metrics, presence of specific comorbidities), as well as factors relating to the particular situation (eg, surgery, obstetric delivery).

When complications occur among elderly men, are the poor outcomes related to the age or gender of the patients as well as the care? Might advanced age merely be common in the study sample? Or perhaps advanced age is a proxy for some underlying severe comorbidity (eg, coronary artery disease) that happens to be more common in older men? Only by disentangling these myriad relationships can we truly understand anesthesia risk and begin to target remedial efforts to the appropriate factor(s) (see Chapter 26).

Underlying this complexity is the reality that covariates may be confounding variables that share 2 properties: They may be associated with the exposure (the care) because an older or sicker person is more likely to have the care, and they also may be independent risk factors for poor outcome (ie, morbidity begets a poor outcome by itself unrelated to the care). Thus covariates may "confound" the relationship between the exposure and outcome, resulting in an anesthesia risk estimate that mixes the risk due to anesthesia care with that due to the confounding variable.

Such confounding can be controlled in the study design by randomization and inclusion criteria in a clinical trial or by matching (eg, ensuring that a specified covariate is present in the study sample in the same proportion as the population).

Stratification

If confounding has not been controlled in the study design, as would be the case in an observational design, stratification can be used. An anesthesia risk estimate can be calculated for men and women separately, then compared with the rate for the combined study sample, thereby revealing the influence, if any, of gender. Almost all cohort studies of anesthesia-associated death attempted stratification, even if they did not use the term.

Multivariable Modeling

When more than a few confounding variables are present, multivariable modeling is used to both control confounding and determine the independent contributions of different factors. (Strictly defined, multivariate modeling refers to simultaneously predicting multiple outcomes, whereas multivariable modeling involves predicting a single outcome at a time.)

In developing the "model," the outcome is viewed as a numerical result of a mathematical function composed of risk factors identified by prior research on the selected outcome, candidate variables of special interest to the investigators, and any variables that may potentially confound the relationships between risk factors and outcome. The particular mathematical form of the model is set by the type of outcome. Most commonly, binary outcomes (eg, death vs survival) are modeled using logistic regression analysis, yielding a set of odds ratios (ORs). As in dichotomous circumstances (eg, survival vs death), the OR is the ratio of the probability (odds) of the outcome in the "exposed" group to that in the nonexposed group (Table 25-1); however, the multivariable modeling yields a set of "adjusted" ORs that reflect the independent effect of each variable (covariate) with the influence of other variables held constant. If time to event is believed important in genesis of the outcome, the relationship is modeled with a proportional hazards model, yielding equivalent hazards ratios; the study by Monk et al18 identifying cumulative deep hypnotic time as a predictor of 1-year mortality is an example.

Although intuitive as one's "chance" of experiencing the outcome, ORs often are less useful in estimating the risk faced by a given individual because several predictors (eg, male gender and advanced age) may be relevant for many individuals. Alternatively, Sinclair et al,93 among others, demonstrated the usefulness of the logistic regression equation, a part of the analytic output not typically presented in study reports, in estimating patient-specific risk of postoperative nausea by substituting into the equation the patient's particular numerical values for the independent predictors.

Propensity–Score Matching

Generally, modeling controls confounding adequately; however, confounding and related bias is ever present and often undetected.94 One type of confounding that is an especially problematic source of bias is confounding by indication, where clinical features prompt use of a given intervention that is also related to the patient's outcome. An example is the historical predilection of anesthesia providers to administer spinal anesthesia preferentially to "poor-risk" patients, believing that this method poses a lesser physiologic trespass. Not surprisingly, many early outcome comparisons noted that spinal anesthesia, sick patients, and higher mortality were associated! Because conscientious physicians will naturally opt to "do what's right" for their patients, confounding by indication is a frequent problem with observational data (selection bias).

A special approach for dealing with selection bias is propensity–score matching. This method uses logistic regression first to identify from the observational data matched patient pairs that, based on their covariates, are equally likely to receive the intervention.95,96 Then, using the matched-pairs subgroup, logistic regression analysis identifies independent predictors for the outcome. This approach has been applied to diverse risk-related topics, including pulmonary artery catheter use in critical care,97 anesthesia choice for hip fracture repair,98 medical direction of anesthesia care,87 and anesthesiologist's board certification.88 The method is commonly used with observational data because a clinical trial is either logistically difficult or not feasible. As with modeling generally, the propensity–score approach cannot control for confounding by covariates that are unknown or overlooked in the analysis; such oversight itself is unlikely to be detected.99

Recursive Partitioning

An alternative approach for studying risk amid confounding avoids fitting data to a mathematical regression equation and instead uses the covariates to repetitively partition the study population into "high-risk" and "low-risk" subgroups with regard to the outcome. Recursive partitioning (also termed data mining) exploits lack of homogeneity among the study population to yield a treelike graphic that depicts how each subject fared with regard to the outcome. This classification approach has been used to identify MI risk among emergency department patients presenting with chest pain,100 poor credit risks for financial services, and genotypes in genomic analysis, but it is only beginning to find use in studying anesthesia risk.101 Recently, this method enabled exploration of the relationships between severity and duration of intraoperative hypotension and 1-year postoperative mortality in a sample population that was inadequate for multivariate modeling.102

Study Designs for Anesthesia Risk

Taxonomy of Study Designs

Interpreting anesthesia risk literature does not require extensive knowledge of statistics, but it does require an appreciation of the strengths and limitations of study designs. Box 25-8 presents a taxonomy of study designs used in anesthesia risk studies. It categorizes these studies as randomized controlled trials, cohort studies, or case-control studies and further describes their purpose and design as experimental (eg, clinical trial) or observational; in the latter, the investigator observes what is often a natural or unplanned experiment, hence the alternative name, observational design. It also distinguishes between prospective and retrospective studies, with the former denoting explicit definition of study variables and outcomes before data collection begins. Most anesthesia risk reports have been retrospective, although prospective designs have been more prevalent in recent studies.

Ranked in order of reliability of results, the clinical trial is first, followed by well-conducted prospective cohort studies, other cohort studies, case-control studies, and case aggregations.103 When similar results are obtained from studies using different designs, we gain confidence in those results. The salient features of the 3 principal designs used in anesthesia risk studies are compared in Table 25-2 as an aid to surveying examples that inform us about anesthesia risk.

Mortality as a Risk of Anesthesia

Early "Anesthesia Deaths"

In 1848, only 15 months after William Morton's successful demonstration of clinical anesthesia in Boston, an Edinburgh medical journal reported the death of Hannah Greener, a 15-year-old British girl who had a toenail removed under chloroform anesthesia, some 4 months after she had a similar procedure performed with diethyl ether anesthesia.104 This first report of an anesthesia-related death initiated a decades-long debate over the relative safety of ether versus chloroform and also indicated that surgical anesthesia could be fatal. Similar case reports followed.

Were these "anesthesia deaths" rare, occasional, or perhaps common? Although case reports and subsequent case series provided descriptions of the fatal events and even clues to possible etiologies, they could not quantitate the frequency of these catastrophes and allow an estimate of the risk in undergoing anesthesia. Anesthesia risk measurement benefited from the early epidemiologic work of anesthesiologist John Snow, who applied to an analysis of 50 deaths during chloroform anesthesia the epidemiologic logic that enabled him to identify the mode of transmission of cholera related to colonization of the Broad Street water pump.105

Cohort Studies of Anesthesia-Associated Mortality

Table 25-3 presents the principal results of international studies in which the mortality risk of surgical anesthesia was studied. These studies were major, often pioneering efforts, with most performed before the advent of modern information technology that facilitates aggregating and analyzing vast data and well before epidemiology and biostatistics had reached their current level of sophistication. Despite these limitations, their reports provide estimates for surgery-related, anesthesia-attributable mortality and, in many, morbidity.

In surveying the unadjusted or crude mortality estimates listed in Table 25-3, we are immediately impressed by a wide variation in rates. Undoubtedly much of the variation is due to different study designs and other factors listed in Box 25-1. Yet there is a suggestion of improved outcome over time, especially in well-developed countries.

Holland66,114 documented a 5-fold decrease in anesthesia-attributable mortality in 1 Australian region over 25 years, much of which he attributed to enhanced risk management education of practitioners. Ever greater progress in Australia was documented by Mackay129 and Gibbs and Borton.4 Similarly, Lienhart et al128 noted an 11-fold decrease in anesthesia mortality over 15 years in France when repeating (albeit with different methodology) the study by Tiret et al,57 which noted poor outcomes in relation to lack of postanesthesia care units.

Moreover, although we might regard the study results listed in Table 25-3 akin to a collection of mixed fruit, a plot of the primary-cause mortality estimates in those studies span almost 2 orders of magnitude (or a near 100-fold decrease) over 6 decades,1 and the putative improvement trend predates the vigorous US patient safety initiatives of the mid-1980s (Fig. 25-6), even if we might be uncomfortable concluding that any specific amount of "improvement" has occurred.

However, even if an apparent trend is wholly an artifact of differing methodologies, as some argue,126 the clinical terrain has changed dramatically over the past half century. Concerned in 1959 that mortality statistics were unchanged from those decades earlier, the National Academy of Sciences' Committee on Anesthesia invited Chauncey Starr from the National Academy of Engineering to comment. "Well, that is exactly the way it is in farming," he related as he began to discuss the tractor principle132: The rate of farming accidents remained high despite many safety improvements in tractors (eg, wider wheel base, roll bars, seat belts). Seeking an understanding, Starr made site visits during which he learned that improved tractor design enabled farmers to plow on steeper inclines! The same is true in medicine, where we now routinely perform major "high-risk" operations on patients who as recently as 35 years ago were "too sick" for surgery. Thus even stable mortality rates in the face of increasing severity of illness reflects improved clinical outcome.

Sources of Mortality and Morbidity Risk

Our literature abounds with attempts to allocate overall perioperative mortality and morbidity risk to specific characteristics of the patient, anesthetic, operation, or clinical setting. An important reference point is the perioperative mortality rate that was estimated to be 1.43% (or 1 in 70) for US hospital–based surgery in 2004133 and has been relatively stable over several decades. Comparable mortality estimates based on individual cohort studies have ranged from 1 in 256 to 1 in 53,8,59,51,46,63,134 without any apparent temporal trend.

The earliest attempt to identify specific sources of mortality risk appeared in Beecher and Todd's mid–20th-century study,8 which noted a perioperative mortality rate of 1 in 75 and ascribed 78% of deaths (1 in 95) to the patient's disease, 18% (1 in 420) to surgical management, and 3% (1 in 2680) solely to anesthesia management. They attributed approximately 5% (1 in 1560) to anesthesia when considering it as both a contributing and a primary cause. Subsequent studies attributed somewhat greater proportions of mortality to surgical and anesthesia causes, as understanding of the pathophysiology of disease and the physiologic effects of anesthesia and surgical intervention increased.3,59,51,46

However, substantial confounding underlies such simple categorizations. Being male, elderly, and in poorer condition (ie, ASA physical status ≥3), each individually augments the risk of a given surgical procedure. An early (and still valid) effort to model perioperative mortality is that of Cohen et al,59 who identified the relative importance of patient characteristics (advanced age and ASA physical status, male gender), surgery-related factors (invasiveness, whether emergent), and anesthesia management (method and drug choice; Table 25-4).

Although definitive allocation of perioperative risk, and specifically of anesthesia risk, is lacking, what follows is an attempt to dissect some of these relationships, with the goal of offering guidance to practitioners as well as identifying improvement opportunities.

The Patient

Importance of Comorbidity

Intuition tells us that sick patients tend to do poorly. Among the many studies8,34,46,48-50,52,57,59,62,63,65,74,75,78-81,89 documenting this truism is Pedersen's prospective cohort study51 of 6307 patients, in which the extent of comorbidity was associated with greater complication and mortality rates (Table 25-5).

Modeling Outcome with Patient Characteristics

Although the general principle is now well established, capturing the precise quantitative relationship between comorbidity and clinical outcome remains an active focus of research. The venerable ASA physical status classification, never designed as a risk metric, has long been known to correlate with perioperative outcome,49,59,63,126,134 with mortality rising sharply with advanced physical status (Table 25-4 and Table 25-6). ASA physical status interacts with age, becoming a much more potent predictor of major complications beyond middle age (Fig. 25-7).

Yet there is substantial subjectivity and variability in use of middle ASA physical status categories, even in settings where payment for services is unrelated to this classification.135-138 Prediction improves when this metric is used in combinations in multivariable modeling with other covariates, such as age and/or the presence of specific morbidities.18,25,34,48-51,59,65,74,80-83,89,122,139,140 Concerns about subjectivity of the ASA physical status metric may also be addressed in risk modeling studies by including the Charlson Comorbidity Score,52 a morbidity-specific metric that has been at least as good an outcome predictor in several studies18,98,141-143; the Acute Physiology Score (APS) of the Acute Physiology and Chronic Health Evaluation (APACHE) II score, a metric commonly used in critical care144; or perhaps the Sickness at Admission Scale Score.145

Alternatively, there have been embryonic efforts to reduce the subjectivity in ASA physical status scoring by enhancing its specificity, thereby enhancing its precision and predictive power: Barbeito et al138 suggested adding a "G" to the scoring of parturients (eg, "2-G" to indicate the gravid status), and Holt and Silverman146 proposed a superscripted notion to indicate specific morbidity (eg, "3RESP" to indicate the affected organ system). Emphasizing the confounding of physical status by surgical complexity, Pasternak147 suggested a preoperative assessment scale that includes both. These await further development and validation. The Charlson Comorbidity Score was developed to predict 1-year mortality among hospitalized medical patients,52 and, although it has taken several decades, similar morbidity-specific scoring systems for predicting surgical mortality are emerging. For example, Sessler's group has developed a risk metric using age and Medicare comorbidity and surgical procedure codes to predict duration of hospitalization 30-day postoperative mortality and morbidity.92

Several simple scoring systems for organ- or operation-specific complications are available. The Cardiac Risk Index of Goldman et al48provided an early model that was improved by Detsky et al,148 by adding angina to the risk factors. Lee et al149 developed and validated a Revised Cardiac Index that specified 6 independent predictors: history of ischemic heart disease, history of congestive heart failure, history of cerebrovascular disease, preoperative treatment with insulin, preoperative serum creatinine more than 2.0 mg/dL, and high-risk surgical procedure. Similar clinical prediction rules have been developed for early outcome after coronary artery bypass surgery.150 Glance et al151 have studied the substantially enhanced postoperative mortality risk associated with the metabolic syndrome (obesity, hypertension, and diabetes).

Even in the absence of simple, comprehensive risk scoring systems, multivariable modeling of several large cohorts has identified a common set of independent risk factors for severe complications and death (Tables 25-4 and 25-5, and Box 25-9). Differences in the risk associated with specific morbidities among studies probably are explained, in part, by differences in how disorder-specific acuity of illness is measured.152 Racial and socioeconomic characteristics, possibly proxies for unadjusted underlying comorbidity and/or problems with access to care, also influence outcome.89,153,154

Failure to Rescue

Building on the relationship between health status and clinical outcome, Silber et al155,156 developed a novel outcome metric, failure to rescue. Using multivariable modeling, they showed that patient characteristics (eg, age, gender, comorbidities) predict complications better than they predict death; whether a complication turns into a death reflects the capability of the facility to rescue the patient. The rescue capability itself depends on hospital characteristics that include proportion of board-certified anesthesiologists and ratios of nurses to patients.88,155,157 Subsequently validated in many studies, this metric is included among quality indicators used by the Agency for Healthcare Quality and Research, National Quality Forum, University HealthSystem Consortium, and independent health care researchers.

The Health Care Setting

Unexplained Variation in Outcome

Long before investigators had tools to identify the source of risk differences, they began to document variations in patient outcome for ostensibly similar surgery.16 Buried in the mid 20th-century Beecher-Todd study8 is an unexplained 3-fold difference in mortality across the 10 participating university hospitals. A decade later, the mortality variation was so great in the mid-1960s 34-hospital National Halothane Study (0.27%-6.40%) that its statisticians concluded, "Such variation in so important an outcome of surgery compels attention."158 Even after adjusting their data for age, ASA physical status, and surgical procedure, an unexplained, 3-fold variation remained.

Emergence of an Ill-Defined "Provider"

Intrigued by unexplained mortality variation in the National Halothane Study, one of its investigators joined with sociologists and statisticians to explore nonclinical factors, including structural characteristics of hospitals, as potential explanatory variables for the postoperative mortality variation among 1224 hospitals using a chart abstracting service in the early 1970s.159,160 Despite case-mix adjustments, 3- to 4-fold outcome variations remained for the 15 major surgical procedures studied in this Institutional Differences Study. Probing deeper in a 17-hospital subset with more detailed institutional data, they explored the influence of the individual surgeon's experience, anesthesia-provider type, and various measures of hospital structure and medical staff organization.161,162 Although such factors were weak predictors of outcome, the mere presence of such relationships heralded the beginning of our understanding of the multifaceted role of an ill-defined "provider" in surgical outcomes. (Provider here is a comprehensive term indicating not only the clinicians but also the facility, its associated staff, organizational structure, and policies.)

Role of Sociomanagerial Factors

The Institutional Differences Study160,162 identified diverse sociomanagerial factors that are associated with better surgical outcomes, including teaching hospital status, greater number of residency programs, higher hospital expense per day, stringent medical staff admission requirements, power over senior surgeons, higher proportion of surgeons' practices at the study hospital, and board certification of physicians. Subsequent studies have validated the influence of teaching hospital status and board certification,155,163-165 although research has not explored the full array of care.166,167 Silber et al155 specifically identified a low proportion of board-certified anesthesiologists on the anesthesia staff as an important hospital characteristic associated with "failure to rescue" the surgical patient who experiences a complication. Generally, hospital characteristics are stronger predictors of clinical outcome than are surgeon's characteristics.161,162

As a natural extension, investigators have explored whether mortality after high-risk surgery is influenced by provider experience. Studying procedures of varying complexity in the mid-1970s in 1498 hospitals using a chart abstracting service, Luft et al168 found an inverse relationship between mortality and procedure volume for high-risk procedures; a volume–outcome relationship existed for lesser risk surgery but was weaker at low volumes. What underlies this phenomenon, a team effect in the hospital, expertise of individual surgeons, or regional referral patterns? Research has been more equivocal, in part, because of technical issues, such as low statistical power in comparisons of lower total case volume and a statistical association (collinearity) affecting hospital and surgeon-specific volumes.169,170 Yet accruing literature documents better outcomes for high-risk surgery performed at high-volume sites and suggests that a more experienced team may underlie the phenomenon.171-173 As a result, performing high-risk surgery at high-volume sites has become a principle that guides contracting for services of some organizations (eg, the Leapfrog Group) and third-party payers.

Certain aspects of critical care medicine are also associated with better surgical outcomes. Pronovost et al174,175 showed that the presence of a dedicated intensive care physician making daily rounds, a nurse-to-patient ratio of at least 1.2, a monthly case conference, and tracheal extubation in the critical care unit rather than the operating room are associated with better outcomes after abdominal aortic surgery. The implications of not having a dedicated critical care physician are so grave (OR for in-hospital mortality: 3.0; 95% CI, 1.9-4.9) that this has also become a Leapfrog Group standard. The contribution of nurse staffing ratios for good outcomes echoes the studies of general hospital nurse staffing by Aiken et al.157

Delivering good outcomes in complex settings also requires effective team functioning, an essential part of an effective patient safety climate.2 Higher scores on the Safety Attitudes Questionnaire are associated with fewer medication errors, lower ventilator-associated pneumonia rates, and decreased risk-adjusted mortality.176 Makary et al177 developed a "teamwork culture" metric that is sensitive to operating room caregivers' perceptions and beginning to be used in improvement work. Awad et al178 showed that a preprocedural operating room briefing improves teamwork climate, including coordination among caregivers, and Haynes et al have demonstrated a 44% reduction in surgical mortality in association with use of an immediate presurgical checklist that fosters enhanced communication among team members.179 (See Chapter 3 for a more thorough discussion of team performance.)

Mortality risk appears to be influenced also by the facility in which the procedure is performed (ie, a "provider" effect). Fleisher et al89 studied outcomes of Medicare patients having surgery in a hospital outpatient unit, freestanding surgery center, or surgeon's office. Although death rates on the day of operation were not different, there were differences among the rates for 7-day mortality (1 in 2000, 1 in 4000, and 1 in 2856, respectively) and 7-day hospital admission (1 in 48, 1 in 119, and 1 in 110, respectively). The readmission rates resembled those in other studies, but the authors could not exclude the likelihood that the differences across sites in both readmission and mortality rates reflected patient referral patterns (selection bias), with sicker patients treated in the more intensive settings (Fig. 25-4B).

The Surgeon and the Operation

The Surgeon's Characteristics

The surgeon's board certification,155,162,164 experience,3,162,164 and case volume162,164,169,171,172 is each associated with clinical outcome when comparing different hospitals, although these relationships may be weak when comparing individual surgeons at the same site. However, outcome variation has been demonstrated among individual surgeons for complex procures.180 Lunn and Devlin3 called attention to the association of poor outcomes from surgery and anesthesia care provided by unsupervised and undersupervised trainees at night, particularly in emergent care and sicker patients, as have others more recently.4,127,129

Outcomes associated with individual surgeons are confounded by diverse hospital characteristics, such as teaching-hospital status,162,166,167 total hospital expenditures,162 total case volume,162,169,172,173 team culture,176,177 critical care organizational characteristics,174,175 and general nurse-to-patient ratio.157

The Operation

The surgeon's influence on outcome is confounded by important characteristics of the operation. In most cohort studies, invasiveness of the surgical procedure (ie, "major" vs "minor" surgery) and whether the operation is undertaken emergently are potent independent determinants of both mortality and morbidity (Tables 25-4 and 25-5).34,48,51,59,148 Underlying these relationships undoubtedly are the implications of the extent of physiologic derangement and the limited preparatory care before emergent surgery. Procedure invasiveness interacts with age, becoming a much more potent predictor of major complications beyond middle age, particularly for the more invasive procedures (Fig. 25-8).50 Invasiveness (or complexity) of operation is so potent a predictor of outcome that it has been included in several proposals for modifications of the ASA physical status classification or new approaches to preoperative patient assessment.82,147

An especially problematic procedure-related factor is duration of operation (or anesthesia). Consistent with the notion that increasing exposure to a potentially hazardous intervention risks a poorer outcome, duration of operation (or anesthesia) is associated with increased mortality and morbidity35,51,59 (Tables 25-4 and 25-5) and hospital admission after ambulatory surgery.34,35 Greater procedure duration may be a proxy for more extensive (perhaps unrecognized) surgical disease, lesser surgical skill, or the occurrence of intraoperative anesthesia and surgical complications, any of which may independently determine the postoperative outcome. Cohen et al59 specifically included occurrence of intraoperative complications among candidate predictor variables and found that such complications, rather than procedure duration, was an independent predictor of outcome (Table 25-4). Thus, unless the study has included occurrence of intraoperative complications in the outcome modeling, procedure duration alone should be regarded as a possibly "tainted" variable.

The Anesthesia Provider and the Anesthesia Care

The Anesthesiologist's Characteristics

Board certification seems an even more important predictor of outcome with the anesthesiologist than with the surgeon.88,155 Using sophisticated modeling, Silber et al155,156 showed that a lesser proportion of board-certified anesthesiologists on the anesthesia staff is associated with a greater likelihood of "failure to rescue." In a direct comparison of mortality among Medicare patients treated by midcareer anesthesiologists with and without board certification, Silber et al88 showed that absence of board certification is associated with greater likelihood of "failure to rescue" (OR: 1.13, 95% CI, 1.01-1.27) and higher mortality (OR: 1.13; 95% CI, 1.00-1.26). So confounded are the outcome predictors, however, that they noted that the poorer outcomes of noncertified practitioners may reflect hospital characteristics as well.

Apart from board certification, the effect of the anesthesiologist's experience on outcome is less clear. In a comparison of anesthesia providers at the Massachusetts General Hospital in the mid-1970s, Gilbert181 detected slightly better outcomes when anesthesia was administered directly by a senior anesthesiologist. However, Cohen et al59 were unable to detect an outcome difference among anesthesiologists with greater time in the specialty and greater case volumes (Table 25-4). Lunn and Devlin3 noted the association between poor outcomes of surgery and anesthesia care and unsupervised and undersupervised trainees at night, as have others.4,129 Exploring the linkage between myocardial ischemia and MI in anesthesia for coronary bypass, Slogoff and Keats182 famously identified Anesthesiologist 7, whose cases had a greater proportion of intraoperative ischemia and much greater postoperative infarction than did others in the group practice. They implied this was due to individual skill factors, but analysis was superficial and without risk adjustment of the clinical data.

Anesthesia Provider Type

More extensive efforts to explore outcome differences by type of anesthesia provider also have been indeterminant because of design flaws, principally failure to adequately address patient selection bias and/or unmeasured, potentially important clinical confounding variables.

Gilbert181 found no outcome differences when comparing anesthesiologists providing direct care, anesthesiology residents medically directed by faculty anesthesiologists, and nurse anesthetists similarly medically directed. Yet a fully trained anesthesiologist was involved in every case, the data were only partially risk adjusted, and patients were not randomly allocated to providers (Note: The physicians likely received the more difficult cases). Similar flaws are present in an analysis based on the Institutional Differences Study: In one portion of that study, clinical outcomes in "[9] hospitals in which anesthesiologists primarily were the providers" were compared with those in "[7] hospitals in which nurse anesthetists were primarily the providers."183 Outcomes were the same in both groups, yet anesthesiologists were involved in all care.

Multiple design flaws also plague an analysis of postoperative deaths in North Carolina during the period 1969-1976, in which half of the anesthetics were administered by nurse anesthetists medically directed by the surgeon and the other half by anesthesiologists working alone or a team of a nurse anesthetist and an anesthesiologist.184 Anesthesia-related death rates were similar across the 3 provider types; yet, unlike the studies by Gilbert181 and Forrest,183 there was no attempt to adjust the data for case-mix differences (eg, age, ASA physical status) and type of operation (eg, emergent vs elective, major vs minor). Because nurse anesthetists working without anesthesiologists (then and now) are located typically in smaller, often rural hospitals, the 2 other provider types (with anesthesiologists) likely were treating sicker patients having more complex procedures. Also, as with the other studies, there had been no random allocation of patients to provider type, resulting in likely selection bias.

A more recent provider-type comparison, although much more sophisticated, still suffers from serious design flaws: Pine et al185 compared mortality of 404,194 Medicare patients having 1 of 8 common surgical procedures, whose anesthesia was provided by nurse anesthetists working without anesthesiologists, anesthesiologists providing direct care, or an anesthesia care team of a nurse anesthetist medically directed by an anesthesiologist. After stratification by procedure and adjustment for patient, institutional, and geographic factors, the mortality rates were similar across the 3 provider types. However, 80% of the cases in which nurse anesthetists worked without anesthesiologists were performed in rural hospitals, most assuredly caring for a lower-risk population (severe selection bias). Direct comparison of outcomes in this circumstance is likely to be misleading because patients in the 3 groups would be expected to differ markedly in measured and unmeasured risk factors and thus their likelihood of poor outcome. Just as Silber et al88 noted that the poorer outcomes of noncertified anesthesiologists may also reflect residual confounding by hospital characteristics, there is likely to be similar, unresolved facility-related confounding in the study by Pine et al,185 particularly without a propensity–score analysis.

In another study, Silber et al87 demonstrated in a matched-pair, propensity–score analysis (addressing selection bias) that lack of medical direction by an anesthesiologist is associated with a higher failure-to-rescue rate (OR: 1.10; 95% CI, 1.01-1.18) and higher 30-day postoperative mortality (OR: 1.08; 95% CI, 1.00-1.15). They estimated that the higher mortality rate engenders 2.5 excess deaths per 1000 patients, which represents a number needed to treat (NNT) of 400 (Table 25-1). Although a NNT of 400 reflects a modest effect, anesthesiologist's medical direction gains enormous importance because of the tens of millions of anesthetics administered each year.

Several recent comparisons of clinical outcomes associated with different anesthesia provider types undertaken with less analytical sophistication—principally the absence of a propensity–score analysis—undoubtedly were flawed by selection bias and/or unmeasured confounding variables. For example, a comparison of maternal outcomes following obstetric anesthesia care by the 3 anesthesia provider types for 1,141,641 patients in 369 hospitals in 6 states during the period 1999-2001 found no statistically significant differences186: Among sources of selection bias was excluding tertiary hospitals expected to receive referrals of high-risk patients (eg, level 3 obstetric units, facilities having neonatal intensive care, Council of Teaching Hospital membership) and patient populations with low prevalence of important chronic illnesses (eg, diabetes).

Another recent example, with particularly subtle flaws, is a study of surgery-related complications and deaths associated with the 3 anesthesia-provider types during the period 1999-2005 in the 14 states that opted out of the Medicare requirement that a physician oversee delivery of anesthesia care by a nurse anesthetist compared with those in states that had not opted out187: They specifically chose to focus on surgery-related adverse outcomes, given that those specifically related to anesthesia care are generally believed to occur at very low rates. Present is the previously mentioned selection bias (particularly in the absence of a propensity–score analysis) related to comparing outcomes from small, often rural hospitals in which nurse anesthetists treat healthier patients having simpler procedures with those from tertiary care centers where anesthesiologists alone or working with nurse anesthetists in an anesthesia care team provide complex care for patients with the highest acuity. In addition, opting out of the physician-supervision requirement had a minimal impact on anesthesia delivery in the 14 affected states because the opt-out provision does not mandate any change but rather gives each hospital the discretion to permit nurse anesthetists to function without physician oversight. Under this circumstance, anesthesia provider change (to unsupervised nurse anesthetists) would probably occur predominantly in the small rural hospitals where anesthesiologists are often not present. Although the presentation of their analysis is not sufficiently detailed to identify personnel shifts by hospital size, case complexity (as measured by mean ASA Relative Value base units) was noted to be significantly lower for solo nurse anesthetists than for either solo anesthesiologists or the anesthesia care team in both opt-out and non–opt-out states. The multivariate analyses, however seemingly exhaustive, did not include a propensity–score analysis or even variables characterizing the site of care. Finally, anesthesia risk (unmeasured here) is but a very small component of overall surgery risk. Hence the inability to demonstrate statewise outcome differences in relation to states opting out should not be unexpected.

The Anesthesia Care

Outcome comparisons of specific anesthesia drugs and methods have not identified a single ideal approach to anesthesia but rather have emphasized characteristics, typically pharmacologic, that can be regarded as tradeoffs among different options. Keats68 suggested that anesthesia agents are "inherently toxic" and that, as our knowledge grows, anesthesia risk diminishes. Past controversies about specific drugs were resolved by either learning how to use them safely or discarding those that were problematic, usually on a pharmacologic basis (eg, methoxyflurane due to dose-related nephrotoxicity). Thus Beecher and Todd8 attributed "intrinsic toxicity" to curare (a generic term applied to muscle relaxants) before it was appreciated that postanesthesia residual paralysis could be both hazardous and avoided with pharmacologic antagonism. That drug class now is a mainstay of anesthesia practice (see Chapter 34).

Perhaps expectedly, well-conducted studies have failed to identify important risk differences among anesthesia options. Regional anesthesia (eg, epidural analgesia) may pose lower risk than general anesthesia for graft thrombosis and deep venous thrombosis, among other morbidity, in vascular, lower extremity, and pelvic surgery (see Chapter 55); however, the magnitude of such benefit is uncertain, and whether it results from neuraxial blockade or avoiding general anesthesia is also unknown.188 Cohen et al59 noted that anesthesia-related factors (eg, principal anesthesia method) added negligibly to the contributions of patient- and surgery-related factors in accounting for mortality risk (Table 25-4). However, they did find markedly greater mortality if a single-agent anesthetic rather than "balanced anesthesia" with multiple drugs was used (OR: 4.85; 95% CI, 1.97-11.96). Although their inability to identify benefits of specific anesthesia methods arguably could result from selection bias (confounding by indication), Forrest et al50 were unable to detect meaningful morbidity differences among inhalation agents in a clinical trial that would minimize biased selection.

However, in support of a long-held belief that it is more important how rather than specifically what one does, Arbous et al189 have shown that there are multiple opportunities to decrease anesthesia risk by adopting a set of good practices in the management of anesthesia care (Table 25-7). Although the efficacy of most of the practices identified by Arbous et al is well documented in the literature (eg, epidural opiates for postoperative pain management rather than intramuscular or intravenous narcotic administration; antagonism of nonmetabolized muscle relaxants at conclusion of anesthesia), these practices continue to require support and encouragement toward full adoption.

More recently, investigators have begun to probe deeper into specific aspects of anesthetic management to gain insights into additional ways to enhance patient outcome and safety: The identification of a relationship between cumulative deep hypnotic time (BIS <45) and greater 1-year mortality18-21 has been followed recently by recognition that outcomes worsen when low mean arterial pressure (MAP <75 mm Hg) and low anesthetic concentration (minimal anesthetic concentration <0.7) are also present.190 Moreover, the duration of such a "triple low" predicts worsening outcomes,190 and the promptness with which a vasopressor is administered attenuates the potential harm of the "triple low."191 These findings from Sessler's group, based on observational data from one large academic medical center, require validation in a prospective study; however, other evidence demonstrates that intraoperative parameters influence patient outcome: a surgical Apgar score—a composite reflecting intraoperative blood loss, lowest pulse rate, and lowest MAP—predicts 30-day complications and mortality rates.192,193

These results emphasize the inadequacy of viewing anesthesia care as merely the presence of certain drugs, anesthetic methods, and devices. We may infer that anesthesia risk studies that consider patient characteristics and their clinical outcomes—but omit intraoperative detail relating to the anesthesiologist's clinical practices—provide an incomplete and possibly misleading perspective. We also might speculate that the beneficial practices identified are likely to be among those that underlie the anesthesiologist's beneficial influence detected by the failure-to-rescue metric.

As if the evolving story of anesthesia risk were not sufficiently complex, communicating effectively with patients about anesthesia risk is even more challenging. Basic concepts of risk are confusing, if not foreign, to patients, and they are likely to be wary of risk-related statements, given the many prominent public reversals about the benefits and safety of common drugs (eg, hormone replacement therapy, cyclo-oxygenase-2 analgesics).

Creating an Effective Message

Because the hallmark of effective communication is understanding the audience, we should tailor risk-related discussions to the specific patient's needs and knowledge, recognizing that a brief targeted disclosure is likely to be mutually satisfying. We can allay their general concerns by noting that anesthesia care has never been safer, that it may be among the least hazardous parts of overall patient care, that common postanesthesia problems tend to be transient, that serious complications have become very uncommon, and that our expanding perspective on anesthesia care (eg, postoperative pain management) means that they are also likely to be more comfortable and satisfied with their care. We can convey to patients that as a result of myriad improvements, anesthesia methods are more similar than different with regard to risk, and often their anesthesia preferences can be honored.

Problems with Risk Numeracy

We should avoid placing undue emphasis on specific risk estimates, in part because they are based on the experiences of large populations and may not predict well a given individual's clinical outcome in a given situation (Box 25-1). Although patient-specific risks can be estimated from regression equations and patients often want to know their "chances" (even if they do not ask), the more problematic issue is that quantitative information about risk is meaningful only to the small minority having facility with probabilities and numerical concepts.194-196 Rather than regarding an outcome occurring 1 in 10 as "common," 1 in 100 as "uncommon," and 1 in 1000 as "rare," patients tend to focus on the numerator almost to the exclusion of the denominator (ie, odds are less important than the possibility that an event can occur). Framing is also a barrier to communicating risk, with "90% survival" perceived as better than "10% mortality."

Evolving guidance for enhancing effective risk communication includes understanding the patient's experience and expectations, presenting relative risks of competing options rather than risk estimates in isolation, using graphics ("decision aids") to help the presentation, encouraging a balanced discussion of options and uncertainties, developing recommendations informed by clinical judgment and patient preferences, and continually checking for understanding and agreement.197,198 For some patients, comparing the risks of anesthesia to other well-known events, such as parachute jumping or flying in commercial aircraft, may be helpful (Box 25-4), but both the clinician and the patient must recognize that these broad generalities may not apply to a given patient.

The foregoing dissection of sources of risk emphasizes the need to move beyond the former narrow definition of anesthesia care as the use of drugs, methods, and devices, to embrace a far broader perspective of the overall anesthesia care process, from preoperative assessment to postoperative care, and its role in the overall system of care (see Chapter 3 for a discussion of the system of anesthesia care, within a larger system of overall care). A real benefit of adopting a more comprehensive view of anesthesia care has already been demonstrated by the mortality and morbidity risk reduction achieved by modulating the sympathetic nervous system by administering β-adrenergic blockers to high-risk patients when appropriate26,199 and preventing hypothermia,100,101 decreasing surgical wound infections by administering perioperative supplemental oxygen,200 decreasing perioperative mortality by adopting optimal clinical practice patterns189 and immediate presurgical checklists,179 and by having an anesthesiologist directing the anesthesia care.87 The identification of an association between higher cancer recurrence and metastasis-free survival rates and use of anesthesia approaches known to reduce stress201 suggests that perhaps only our formerly narrow notions of "anesthesia" risk are preventing further reduction in anesthesia risk.

• Arbous MS, Meursing AEE, van Kleef JW, et al. Impact of anesthesia management characteristics on severe morbidity and mortality. Anesthesiology. 2005;102(2):257-268.
• Gold BS, Kitz DS, Lecky JH, Neuhaus JM. Unanticipated admission to the hospital following ambulatory surgery. JAMA. 1989;262(21):3008-3010.
• Li G, Warner M, Lang BH, Huang L, Sun LS. Epidemiology of anesthesia-related mortality in the United States, 1999-2005. Anesthesiology. 2009;110(4):759-765.
• Lindenauer PK, Pekow P, Wang K, et al. Perioperative beta-blocker therapy and mortality after major noncardiac surgery. N Engl J Med. 2005;353(4):349-361.
• Lunn JN, Devlin HB. Lessons from the confidential enquiry into perioperative deaths in three NHS regions. Lancet. 1987;2(8572):1384-1386.
• Pronovost PJ, Jenckes MW, Dorman T, et al. Organizational characteristics of intensive care units related to outcomes of abdominal aortic surgery. JAMA. 1999;281(14):1310-1317.
• Rodgers A, Walker N, Schug S, et al. Reduction of postoperative mortality and morbidity with epidural or spinal anesthesia: results from overview of randomized trials. BMJ. 2000;321(7275):1493-1505.
• Sessler DI, Sigl J, Manberg PJ, et al. A broadly applicable risk stratification system for predicting duration of hospitalization and mortality. Anesthesiology. 2010;113(5):1026-1037.
• Silber JH, Kennedy SK, Even-Shoshan O, et al. Anesthesiologist direction and patient outcomes. Anesthesiology. 2000;93(1): 152-163.
• Silber JH, Williams SV, Krakauer H, Schwartz JS. Hospital and patient characteristics associated with death after surgery: a study of adverse occurrence and failure to rescue. Med Care. 1992;30(7):615-629.