Chapter 12. Clinical Decision-Making

•  Decision analysis is a formal process used to determine the preferred course of action from two or more potential approaches to clinical management.
• Decision analysis is most appropriate when there is both some uncertainty about the preferred course of action and a meaningful trade-off of risks and benefits for the alternative management strategies.
• A decision tree or decision diagram is the underlying structure of the clinical situation, including all uncertainties and choices, as well as all outcomes.
• The expected utility is the estimated typical or average outcome for a population of patients managed with a particular strategy.
• Substituting a range of values for a particular probability in a decision tree and determining the impact on the expected utility is referred to as a sensitivity analysis.

Cholangitis, an infection of the biliary ductal system, classically presents with a triad of fever, jaundice, and pain in the right upper quadrant of the abdomen. The clinical illness arises in the presence of bacterial colonization of the bile, most commonly because of obstruction of biliary tract flow. Historically, the most common underlying cause of obstruction was blockage of the biliary tract by stones, although malignant strictures have become increasingly important contributors in recent years. The increased pressure within the bile ducts resulting from obstruction produces bacterial reflux into the hepatic veins and perihepatic lymphatics, with subsequent bacterial spread into the circulating bloodstream. The bacteria associated with cholangitis include Escherichia coli, Klebsiella species, and the enterococci, with a shift more recently to include Enterobacter and Pseudomonas species. Infections typically involve multiple species of organisms.

The clinical manifestations of cholangitis range from asymptomatic illness to severe toxic symptoms, including septic shock. Fever is present in almost all patients, typically accompanied by chills. Jaundice and abdominal pain in the right upper quadrant of the abdomen are part of the classical description of presenting symptoms, although they may not be present in the absence of obstructive stones. Complications of advanced disease can include confusion, liver abscess, low blood pressure, and kidney failure. Laboratory examination usually reveals an elevation of leukocytes in the blood, with moderate elevation of bilirubin and liver enzymes.

The clinical management of patients with cholangitis is adapted to the severity of the illness, with severely ill patients requiring intensive care and monitoring. General supportive care typically includes stopping oral intake, starting intravenous fluids, and after taking appropriate cultures, initiating antibiotics for a broad range of potential pathogens. Relieving the elevated pressure within the biliary tract also is important, particularly for patients with severe symptoms. This can be accomplished through a variety of means, including using an endoscope to drain the tract. If there is an obstructing stone, this can be removed with an endoscope, or if an obstruction cannot be removed, a stent can be inserted to maintain flow of bile. Surgical drainage is associated with a high risk of morbidity and mortality, and therefore generally is a choice of last resort.

Alcoholic hepatitis is an inflammatory process that occurs in persons with long-term heavy consumption of alcohol, although only a minority of heavy consumers will develop this problem. As with cholangitis, the clinical severity of alcoholic hepatitis ranges widely from asymptomatic illness at one extreme to severe incapacitating disease at the other extreme. Patients often experience loss of appetite, nausea and vomiting, weight loss, and malaise. Fever is present in about half of all patients. The physical examination typically reveals jaundice, enlargement of the liver, and tenderness in the right upper quadrant of the abdomen. With advanced disease, patients may present with abnormal fluid collections in the soft tissues (edema) or abdominal cavity (ascites), bleeding, or impaired central nervous system functioning.

The laboratory evaluation of patients with alcoholic hepatitis often reveals an elevated level of leukocytes in the blood. Bilirubin levels can be elevated to varying extents, typically accompanied by abnormal levels of liver enzymes. To the extent that hepatic function is disordered, prolonged prothrombin time, depressed levels of serum albumin, or elevated blood ammonia may also be observed. Microscopic examination of the liver reveals infiltration with leukocytes, accompanied by degeneration and necrosis of liver cells. The damaged liver cells often include material that stains with eosin dye and that are referred to as Mallory bodies.

The clinical management of patients with alcoholic hepatitis involves initial supportive therapy. The recommended regimen typically includes abstinence from further consumption of alcohol, rest, and nutritional therapy. In addition, patients who present with severe complications, such as bleeding, ascites, or central nervous system dysfunction, may require specific therapy directed at these disorders. Other therapeutic approaches have an uncertain benefit. Some patients may benefit from corticosteroid treatment. The role of liver transplantation for patients with end-stage alcoholic liver disease is still a matter of debate.

The practicing physician is a problem solver whose clinical task is to maximize benefit to the patient by correctly diagnosing illness and instituting appropriate treatment. The problem-solving process in medicine involves

• 1. Collection and evaluation of diagnostic information
• 2. Formulation of diagnostic hypotheses
• 3. Consideration of alternative hypotheses
• 4. Appropriate use of an efficacious therapy or procedure.

This process is made difficult, in part, because of inherent uncertainty in the information both for diagnosis and for treatment. For example, information elicited from patients may be false, inconclusive, and of uncertain validity; signs often are not perfect indicators of the presence of specific diseases; and even laboratory and other diagnostic tests can yield misleading information—either false-negative or false-positive results. In addition, treatment modalities are far from perfect, and there is always uncertainty about how efficacious a therapy may be in an individual patient. This uncertainty that bedevils the practice of medicine may be ameliorated, in part, by effective use and measurement of probabilities to assess the validity of diagnostic data and the efficacy of therapies, thus demonstrating that the practice of medicine is probabilistic—not deterministic—in nature.

An understanding of the derivation, calculation, and use of probabilities is important in making clinical decisions, and thus is an important topic in clinical education. The subjects presented in this book have direct applicability to clinical decision-making. Although developed in populations, the concepts and measurements of disease frequency (incidence, prevalence, risk, and probability) have direct applicability to probability of disease in single patients. The sensitivity and specificity of information collected from the patient’s history, the physical examination, and laboratory tests are probability measures that play a significant role in determining the diagnostic utility of the work-up. In addition, the clinical trial assesses the efficacy of a therapy in terms of proportions of patients who benefit, so this probability of benefit derived from a population of patients can be used to predict the outcome of treatment in a single patient.

Decision analysis is an explicit process that considers medical choices, medical outcomes, and the uncertainty in the clinical and test data used to make decisions. The process is explicit in that the medical alternatives present and the choices available in the management of a patient’s care may be clearly specified and demonstrated to others: physicians, students, and with care, the patient. The elements in a formal decision analysis include the following:

• 1. An underlying structure termed the decision tree (or decision diagram)
• 2. The probabilities for uncertain events
• 3. The incorporation of test results
• 4. The medical outcomes of the alternative choices to be made.

Table 12–1 summarizes the considerations: the decision problem must be explicitly stated, the problem must be structured over time, the information about uncertainties and the value of outcomes must be characterized, and a preferred course of action must be demonstrated.

• 1. Clinical measures: a variety of patient outcomes might be considered, including, among others, functional status, probability of survival (or death), risk of complications, and years of life expectancy.
• 2. Economic measures: models that include the costs of management alternatives are useful for assessing the cost effectiveness of treatment alternatives; these models typically take the societal perspective of costs, but other perspectives are possible.
• 3. Utility measures: quality-of-life measures, termed utilities, represent how patients value life in a particular health state. Utilities typically are placed on a relative scale, with the extremes being zero, corresponding to death, and one, corresponding to perfect health.

The process of assigning values to the elements of a decision tree can be challenging. For clinical events, probabilities may be obtained from systematic reviews of the medical literature. For many situations, however, published information may be limited or nonexistent. In the absence of published data, the parameters of interest may be estimated by the collection of new information, or by reference to existing databases, such as disease registries, electronic medical record systems, or reimbursement claims. Under certain circumstances, clinical opinion may be used, particularly if the analysis reveals that the conclusions are not highly sensitive to these estimates.

A variety of approaches can be used to estimate utilities, including surveys in which subjects rate health states, or to make judgments about willingness to trade risk for benefit in moving from one health state to another. For example, persons may be asked to determine whether they are willing to accept a more aggressive treatment regimen by weighing the risk of death against the gain of additional functional status.

In constructing a decision tree, it is useful to bear in mind two competing considerations. On the one hand, the tree must be complex enough to represent the key events and probabilities in the decision-making process. On the other hand, the tree must be simple enough to be able to be understood. Inevitably, the tree is a model of reality, and as such cannot capture all of the complexities of a particular clinical situation. To reach a valid conclusion, however, the model must include the key elements needed to determine relative risks and benefits.

Alternative paths that may be chosen are termed clinical scenarios. For each path, an Expected Utility (expected outcome) is a numeric value that may be calculated as described below. The Expected Utility is interpreted as the average, or the expected result, if the decision-maker follows a specific path (clinical scenario). The Expected Utility can combine medical data in the form of probabilities with human values in the form of utilities. The integration of evidence-based medicine, with both patient preferences and economic considerations, allows for a holistic approach to health care decision-making.

The purpose of a decision analysis is to identify a preferred path among two or more clinical scenarios. The preferred path may be selected as the best outcome on the basis of clinical response, utility, or cost effectiveness.

The numeric calculations used to derive the Expected Utility are simple in concept. At each chance node, the probabilities and numeric outcomes of each branch of the path are multiplied, and the results from separate branches are summed. This process is termed averaging out. In this manner, calculations are performed from right to left to obtain the Expected Utility, which is always presented in the same numeric terms as the outcome. When certain pathways clearly will not, for medical reasons, be considered, the branch is folded back using double slash lines to indicate a path that is not considered appropriate. The calculations of Expected Utilities, which may become tedious to perform, are easily handled by available computer programs.

One of the challenges of a formal decision analysis is deciding how many pathways should be included. In many clinical situations, a large number of combinations of tests and interventions is possible. Rather than considering all possible options, it is advisable to limit the number of pathways considered to a subset that is clearly different from other subsets and covers the full range of options that might be pursued.

The element of time also is fundamental to the conduct of a decision analysis. It is common practice to emphasize current events more than future ones. Assigning greater value to near-term events is referred to as discounting or valuing time. For example, if an intervention will save money many years down the road, the present value of those savings is less because of inflation. Similarly, near-term clinical benefits may be valued more highly than clinical benefits that will be experienced in the future. Many of the aspects of a formal decision analysis are best illustrated by considering a specific clinical problem. The situation introduced in the Patient Profile affords us an opportunity to examine a decision analysis in detail.

The patient had a history of heavy use of alcohol and presented with fever, jaundice, and abdominal pain. The clinical presentation—along with laboratory findings of leukocytosis and elevated levels of bilirubin and liver enzymes—suggested a diagnosis of either cholangitis or alcoholic hepatitis. Beyond initial supportive care, endoscopic or surgical intervention would be warranted to relieve the biliary obstruction precipitating cholangitis, but these procedures, particularly surgery, would be contraindicated for a patient with alcoholic hepatitis. This decision problem can be examined using the process of decision analysis. The various probabilities of uncertainty and outcome are presented in Table 12–3.

The structure of the tree (Figure 12–2) indicates the outcomes in terms of probabilities of survival for each alternative, as well as probabilities of each clinical state or test result (the branches of the chance nodes). Note that if a biopsy is done, the decision for surgery or no surgery awaits the biopsy result, exactly as it would in the clinical setting. Note also that the summation of all probabilities at any chance node always equals 1 (100%). Once constructed, the tree is solved by multiplying the outcome values (survivals in this case) by their respective probabilities. Around each chance node, the products of the branches are added to give the value at that chance node. After all the chance nodes from a single decision arm are summed, the Expected Value (expected utility) of that decision is achieved. This can be compared with the values of the other branches of the choice node; then the best numeric choice may be selected as the most favorable option. The other potential choices are excluded or folded back (by convention with double slash lines). Values of each chance node are shown in that node.

The calculations are straightforward. For example, the Expected Utility at the surgery node, expressed as the probability of survival, is calculated as

In this model, although there is a small risk of mortality associated with performing a diagnostic liver biopsy, the information gained from the biopsy helps to optimize the subsequent treatment decision (Figure 12–3). On the basis of this probabilistic argument, the clinical preference would be to perform a diagnostic liver biopsy. Note, however, that using this approach does not ensure the best outcome for the patient, as there is a small (0.2%) but real risk of a fatal complication from the biopsy. Nevertheless, after weighing the relative risks and benefits of performing the diagnostic biopsy, the most favorable outcome is likely to be achieved if a liver biopsy is performed first. This is reflected in Figure 12–3 by the Expected Utility for liver biopsy (0.76), which is somewhat more favorable than medical management alone (0.70), and clearly better than surgical intervention.

This probabilistic approach contrasts with the traditional method of decision-making, in which the outcome is frequently the only criterion for assessing the method, so that good results may positively reinforce bad decision methods, and bad results may negatively reinforce good decision methods. It should also be noted that other outcomes could have been used, such as surgical morbidity, biopsy morbidity, and others. To the extent that morbidity parallels mortality, the optimal approach will remain the same. However, circumstances can be envisioned in which there may be discrepancies in Expected Utilities depending on the choice of outcome. For this reason, it may be useful to consider different choices of outcome.

A criticism of formal decision analysis has been that the input data (probabilities) may not be verifiable, and the values may be subject to doubt. Data obtained through the conduct of large and well-designed studies typically are the most reliable, and hence when available, are preferred for use in decision analysis. For many clinical scenarios, however, carefully designed studies have not been conducted, so it is necessary to rely on data of less certain reliability and validity. The latter type of information can be collected by retrospective review of outcomes for a series of patients treated at a particular facility, and in the absence of any observational data, on clinical impression based on prior experience with treated patients.

The use of data that are little more than a “best guess” is subject to question. A method for handling this problem in decision analysis is called sensitivity analysis. The term sensitivity analysis should not be confused with the sensitivity of a diagnostic test. It also should not be confused with the use of the same term in the context of meta-analysis. In meta-analysis (Chapter 7: Clinical Trials), we referred to sensitivity analysis in the context of examining the pattern of results across subgroups of studies to understand possible sources of heterogeneity in results.

In applying sensitivity analysis to the problem illustrated above, for example, among patients with alcoholic hepatitis the impact of allowing the probabilities of death following surgery can be considered to range between 50% and 75%. When these values are inserted into the tree and the problem is solved, performing a diagnostic liver biopsy remains the preferred clinical scenario. It can be concluded, therefore, that the inferences from the decision analysis are invariant to at least this range of postsurgical mortality probabilities for patients with alcoholic hepatitis.

Any of the variables (probabilities) may be subjected to sensitivity analysis, and if a reasonable maximum or minimum value changes the inference from the analysis, this circumstance should be considered in choosing the preferred clinical scenario. Sensitivity analysis can be performed easily with a personal computer and appropriate software. The analysis may be done using either a spreadsheet program or a more specific program, of which there are several (see references). The computer allows rapid calculations over a wide range of probabilities, so that the answers are quickly available, often in easily understood graphs.

It is also necessary to consider that any test is less than 100% sensitive and specific in classifying disease. If the liver biopsy misses 5% of alcoholic hepatitis and incorrectly labels other diseases as alcoholic hepatitis 2% of the time, then these considerations must be included in the decision tree. The new tree will be slightly more complicated, but it can easily accommodate these differences. For this particular problem, sensitivity analysis of the possibility of alcoholic hepatitis using values for sensitivity and specificity of the biopsy ranging from 50% to 100% does not change the preference for performing the diagnostic test. The likelihood that the biopsy is positive has little influence on the decision to test over a wide range of probabilities.

In contrast, it can be shown that the preferred clinical scenario will vary somewhat with the probability of alcoholic hepatitis versus cholangitis. The results of a sensitivity analysis for alcoholic hepatitis over a range of probabilities from 0% to 100% is given in Table 12–4. Over most of the range of values (10–90%), there is clear benefit to obtaining the diagnostic liver biopsy. Only at the extremes of probability of alcoholic hepatitis is there any question about the value of the diagnostic biopsy. When it is certain that the patient has cholangitis, proceeding directly to decompression of the biliary obstruction offers an Expected Utility comparable with that of obtaining the diagnostic liver biopsy first. On the other extreme, when it is virtually certain that the patient has alcoholic hepatitis, proceeding directly to medical management offers an Expected Utility comparable with that of obtaining the diagnostic liver biopsy first.

As might be suspected, the estimated probability that the liver biopsy itself will produce a fatal complication also influences the decision of whether or not to test. The results of a sensitivity analysis of surviving a liver biopsy using estimated probabilities ranging from 90% to 94% are shown in Table 12–5. In this problem, the sensitivity analysis indicates that the preferred alternative is liver biopsy unless the risk of a fatal complication is greater than 8%.

Reaching optimal decisions in clinical practice is complex, in part because the decision may involve the use of information that is imperfect or of uncertain validity. The estimation of the degree of uncertainty in clinical information, based on the concept of probability estimates, and the use of these estimates in a logical structure are useful tools for making objective decisions in medicine. Formal decision analysis requires the specification of alternative management strategies and provides an explicit probabilistic method for choosing the most favorable anticipated course of action.

The elements of a formal decision analysis include the following: (1) identification of an underlying set of management options, (2) application of probabilities to uncertain events, (3) incorporation of test results, and (4) inclusion of anticipated clinical outcomes for alternative management options. Within the formal structure of the process, referred to as a decision diagram or tree, numerical values are inserted for the anticipated likelihoods of diagnoses, test results, and treatment outcomes. The clinical end points can be assessed in a variety of different ways, such as morbidity, mortality, or quality of life. The alternative management options within the decision diagram are referred to as clinical scenarios, each of which is associated with an Expected Utility (or outcome). The Expected Utilities of different scenarios are compared to arrive at the most favorable approach to management. In many circumstances, precise estimates of alternative diagnoses, test performance, or clinical outcomes are unavailable. In these situations, the impact of assuming different values for the unknown parameters can be explored through sensitivity analysis.

Some clinicians may object to the use of probability estimates, arguing that their use may imply more precision than is warranted by the clinical information, that it dehumanizes the doctor–patient relationship, that it may somehow increase liability, and so on. Although these concerns may limit the widespread use of formal decision analysis, this technique can serve as a useful tool to aid in appropriate patient management. As the stethoscope aids an experienced clinician in making a diagnosis, decision analysis provides insight into alternative management strategies, taking advantage of clinical and epidemiologic methods, many of which are outlined in this book.

For the situation described in the Patient Profile, the clinical presentation suggested a diagnosis of either cholangitis or alcoholic hepatitis. Cholangitis requires surgery to relieve biliary obstruction, whereas surgery is contraindicated in alcoholic hepatitis. By framing this management problem in a formal decision analysis, performing a liver biopsy to reduce uncertainty about the diagnosis was shown to result in a more favorable outcome (in terms of mortality) than proceeding directly to either surgical or medical management.

Sensitivity analysis further demonstrated that the decision to perform an initial liver biopsy would be preferred under a fairly wide range of expected mortality probabilities for alcoholic hepatitis and test accuracy. There were only two circumstances in which an initial liver biopsy did not appear warranted: when there was minimal baseline uncertainty about the diagnosis, and when the anticipated likelihood of a fatal complication of the biopsy reached an appreciable level. Using the available information, the physician in the Patient Profile would be justified in opting for a diagnostic liver biopsy before proceeding further in patient management.