Clinical reasoning is the essential function of the physician; optimal patient care depends on keen diagnostic acumen and thoughtful analysis of the tradeoffs between the benefits and risks of tests and treatments. As benchmarks for considering these issues, here are several real examples:
A 33-year-old man presents to the emergency room complaining of headache, facial flushing, and urticaria of the trunk, which he had never had before. Based on these limited data, most physicians would fail to make the correct diagnosis in this patient, but a clinician who had access to only this information suspects an unfamiliar disorder, namely scombroid poisoning. Twenty minutes later another man who ate bluefish in the same restaurant as the patient comes to the emergency room with the same complaints. Scombroid poisoning was the correct diagnosis.
A 62-year-old woman treated successfully 13 months earlier for exophthalmic goiter and thyrotoxicosis with propylthiouracil develops an alteration in her voice, regurgitation of fluids through her nose, and progressive weakness in her extremities. For 1 month, several physicians are unable to determine the nature of her problem, but then another physician immediately recognizes that the patient is suffering from a form of myasthenia gravis associated with recovery from hyperthyroidism. He verifies the diagnosis and treats the patient effectively for this disorder.
A 49-year-old man with cholangiocarcinoma develops renal failure and anemia after treatment with five cycles of fluorouracil, adriamycin, and mitomycin over a period of 1 year. A clinician unaware that such a complication is known to be caused by mitomycin nonetheless makes a diagnosis of hemolytic uremic syndrome and correctly attributes it to the mitomycin.
A 66-year-old woman with long-standing hypertension and cardiac failure admitted to the hospital with breathlessness is found to have distended neck veins, pulmonary edema, an S3 gallop, and a rough 5/6 holosystolic murmur radiating to the left axilla. After intubation for hypoxemia, she develops a picture of cardiogenic shock. Treatment with dopamine and then norepinephrine and nitroprusside fails to raise her blood pressure. Several physicians are baffled and expect the patient to succumb, but another physician notices a “spike and dome” configuration on her arterial tracing, a brisk carotid pulse with a bisferiens quality, and augmentation of the arterial pulse in the beat following a ventricular premature contraction. He diagnoses asymmetric septal hypertrophy, stops all drugs, including digoxin, and administers intravenous saline and phenylephrine. The patient recovers promptly from the hypotensive episode.
A gastroenterologist, in reviewing a histologic specimen from a 52-year-old man with acute appendicitis, believes that the diagnosis rendered by more than one pathologist (carcinoma of the appendix) is an incorrect interpretation, and he advises the patient against having the right hemicolectomy that has been recommended. The patient follows his advice and remains well with no evidence of cancer for more than 30 years.
A diagnosis of cirrhosis is made by two physicians in a 37-year-old schoolteacher with ascites, peripheral edema, temporal wasting, and palmar erythema. Liver biopsy is said to confirm the diagnosis. A consultant exhibits skepticism about the diagnosis because liver function tests are virtually normal. He extracts a previously undetected 4-year history of breathlessness and finds distended neck veins and a cardiac murmur, and an echocardiogram demonstrates that the correct diagnosis is silent mitral stenosis.
A 55-year-old man with a history of alcohol and benzodiazepine abuse is brought to the emergency room in a coma and is found to have profound metabolic acidosis. A physician systematically analyzes the patient’s electrolytes, anion gap, and osmolar gap, finds oxalatecrystals in the patient’s urine sediment, diagnoses intoxication with ethylene glycol, and treats the patient promptly. The patient recovers.
These brief vignettes are descriptions of actual occurrences. They display exceptional cognitive proficiency by some physicians when others had exhibited suboptimal performance, and they illustrate graphically the critically important nature of the reasoning processes of the physician,[ 1] both “intuitive” (rapid pattern recognition) and “analytic” (deliberative and exhaustive) reasoning as examples of the extremes of the “cognitive continuum.”[ 2 ], [ 3], [ 4], [ 5], [ 6]-[ 7] Few would contest the notion that no matter how competent a physician may be at other tasks, outcomes cannot be optimal if these reasoning skills are deficient. Indeed, the prime function of the physician is clinical reasoning: to suspect the cause of a patient’s symptoms and signs, to gather additional relevant information, to select necessary tests, and to recommend therapy.
Although no one would doubt that cognitive skills are the basis for these tasks, medicine has developed few methods to enhance the acquisition and development of these problem-solving skills. Instead of discussing how diagnostic hypotheses are initiated and refined and how testing and treatment decisions should be formulated, teachers of clinical medicine have substituted standardized histories and physicals, book chapters that list the myriad causes of individual symptoms, an apprentice system in which the student is expected to imitate others, formal approaches to recording patients’ problems, and lock-step algorithmic charts for blind guidance. None of these methods focuses on the essential reasoning processes that are critical to optimal performance.
The last several decades have witnessed considerable growth in our understanding of human reasoning and, in particular, clinical reasoning from fields generally not considered part of the fabric of clinical medicine. Research in these disciplines, namely cognitive science, decision theory, and computer science (in particular, artificial intelligence), provides insights into the critical cognitive processes that form the basis for both teaching and learning the principles that underlie diagnosis and management. These insights into the process of medical reasoning also form the basis for identifying errors in clinical cognition and improving the quality of medical care.[ 6],[ 8],[ 9]
Part I of this book describes many of the insights identified in recent years. It begins with a brief overview of the processes of diagnosis and management decision making and then elaborates on five aspects of the diagnostic process: generation (evocation) of diagnostic hypotheses, refinement of hypotheses, diagnostic testing, causal reasoning, and diagnostic verification. The book continues with a discussion of therapeutic decision making, evidence-based medicine, and cognitive errors in diagnosis. To introduce the reader to unfamiliar concepts derived from disciplines outside of medicine, we also consider some cognitive concepts underlying problem solving, knowledge, and memory. The final discussion offers some views on how to learn and how to teach the processes that are considered.
Throughout Part I, extensive references are given to the cases in Part II. These cases generally are in three parts: real clinical problems selected for their capacity to elicit significant aspects of clinical reasoning, prospective discussion of the problems by experts, and detailed analyses of the reasoning used. The analyses, focused around the specific clinical problems, elaborate extensively on the cognitive principles discussed in Part I. As with any new discipline, some of the expressions used in this book may not be familiar, and some have not achieved universal acceptance. For that reason, an extensive glossary is provided.
The book is about reasoning in clinical medicine. Because theories of human problem solving (including clinical problem solving) are incomplete, some of the concepts described here must be considered tentative. Although many of these concepts are new to medicine, they are sufficiently accepted by cognitive scientists to be adapted for use in learning and teaching. The cognitive aspects of diagnosis have been studied quite extensively, but few studies have been carried out on management decision making, that is, the process by which physicians make testing and treatment decisions.
This lack of information on physician behavior impedes our ability to assemble a comprehensive description of testing and therapeutic decision making. Nonetheless, we borrow generously from the principles inherent in prescriptive, or normative, approaches to decision making such as Bayes’ rule and decision analysis.[ 10],[ 11] These principles have been elaborated in sufficient detail to explain the rationale for many testing and treatment decisions.
Diagnosis is an Inferential Process
In the process of diagnosis, the clinician makes a series of inferences about the nature of malfunctions of the body. These inferences are derived not only from existing observations (historical data, physical findings, and “routine” studies), but also from invasive tests and responses to various interventions. Inferential or inductive reasoning proceeds until the clinician has identified a “working diagnosis,” a diagnostic category sufficiently acceptable to establish a prognosis, dictate a therapeutic action, or both.
When making diagnostic inferences from clinical data, the clinician uses many strategies to combine, integrate, and interpret the data. Clinicians make extensive use of rules of thumb or short-cuts (designated heuristics by cognitive scientists) in the process of gathering and interpreting information.
Rather than rely on statistical data on disease prevalence to generate diagnostic hypotheses from a set of findings, for example, they often assess the likelihood of diseases on the basis of the salience of the findings or familiarity: the resemblance of the findings in a given patient to those of a known disease. By reducing the need to ask an inordinately large number of questions, these rules of thumb make the task of information gathering manageable and efficient. By and large, judgments based on heuristics are accurate and appropriate, although on occasion they can be faulty.
Studies of human cognition suggest that problem-solving strategies depend on the nature of the clinical problem being addressed and even more on the expertise of the clinician. Nonexperts tend to use nonselective strategies that, although they are applicable across a wide range of clinical settings, are nonspecific, rather weak problem-solving methods and inefficient in generating specific hypotheses.
Experts, on the other hand, typically employ strong diagnostic problem-solving approaches tailored to a particular problem or situation usually in the domain of their expertise. Rather than casting their nets broadly, experts quickly focus on a problem by recognizing patterns, formulating problems in semantically meaningful “chunks,” gathering data relevant to a perceived specific solution of the problem, and applying familiar, “prepackaged” actions.
Diagnosis Based on Hypothesis Generation and Testing
At the inception of a diagnostic encounter, the first step is generation, or evocation, of one or more diagnostic hypotheses. The diagnostic process focuses on one or more evolving hypotheses. Typically, the clinician generates initial hypotheses merely from a patient’s age, sex, race, appearance, and presenting complaints, but sometimes such hypotheses emerge exclusively from physical findings or from laboratory data. Additional hypotheses are triggered as new findings emerge.
A diagnostic hypothesis can be either quite general (such as infection) or quite specific (such as acute inferior myocardial infarction). It can take several forms, including a state (inflammatory process), a clinical disorder (acute transplant rejection), a syndrome (nephrotic syndrome), or a specific disease entity (polycythemia vera). The formulation of a preliminary hypothesis on the basis of only a few observations is critically dependent on the cognitive ability to relate a new situation to past experience.
Diagnostic hypotheses serve an essential function: They form a context within which further information gathering takes place. This context, a diagnostic category of some kind (e.g., acute bacterial meningitis), provides a model against which a given patient’s findings can be assessed. The context is the framework for further hypothesis assessment. It specifies both the findings that should be present and those that should be absent if the patient has a given disorder.
Diagnostic reasoning proceeds by progressive hypothesis modification and refinement. Some hypotheses are made more specific, some previously triggered hypotheses are deleted, and some new ones are added. It is not clear how much of the diagnostic process is driven by hypotheses as described here and how much is driven simply by the availability of data from the patient’s history, the physical examination, or the laboratory. Quite likely, elements of cognitive approaches driven by hypotheses and those driven by data are frequently intermingled.
Verifying a diagnostic hypothesis is the penultimate task. It creates a working diagnosis that is used to plan further action. Because the diagnostic process is inferential, all diagnostic hypotheses (even those refined by extensive data gathering and interpretation) necessarily reflect a belief or a conviction by the physician regarding the nature of the condition from which the patient suffers.
Verifying a hypothesis is a kind of test of its validity. It involves assessing a hypothesis for its coherency (are all physiologic linkages, predisposing factors, and complications appropriate for the suspected disease in this patient?), its adequacy (does the suspected disease encompass all the patient’s findings—normal and abnormal?), and its parsimonious nature (is the suspected disease a simple explanation of all the patient’s findings?), often referred to as Ockham’s razor or the law of parsimony, from the fourteenth-century philosopher William of Ockham, who advocated “entia non sunt multiplicanda praeter necessitatem,” which can be interpreted as recommending that the simplest solution (i.e., the one with the fewest assumptions and factors) may be the best.
Verifying a hypothesis also requires eliminating competing hypotheses (can any other disease[s] explain the patient’s findings better than the current hypothesis?). This process produces one or more working diagnoses that form the basis for the next step in patient management—arriving at a certain forecast about the patient’s subsequent clinical course, taking no further action, ordering additional tests, or treating the patient. As noted later, such choices are a function not only of the probability that a patient is suffering from one or more given diseases, but also of the benefits to be derived from further testing, the risks of further testing, and the benefits and risks that accrue from treatment.
Alternate Concepts of Diagnostic Strategies
How much physicians use a general problem-solving approach such as the one just described is open to question. Quite likely, nonexperts rely on it considerably more than experts. The hypothesis generation/testing concept came into question because diagnostic accuracy seemed dependent more on a mastery of content (knowledge of disease and patterns of diseases) than on any specific strategy.[ 4],[ 12 ] Researchers in the field have subsequently tried to identify the nature of such knowledge structures and the mechanisms of their retrieval. Some suggest that diagnosis proceeds by matching the characteristics of a new case to a previously encountered specific instance or to a general resemblance of cases previously seen.[ 4],[ 12 ]
Others propose that clinicians develop mental models, abstractions, or prototypes and use a kind of pattern matching approach to diagnosis. Still others have hypothesized the existence of “illness scripts”—cognitive structures somewhat analogous to the frame structure of some artificial intelligence computer programs.[ 13],[ 14] Finally, some workers in the field have opined that a variety of methods (including all of the aforementioned) are used flexibly to solve diagnostic problems.[ 8],[ 12 ],[ 15] What this means is not certain, but what does seem to be clear is that in the absence of an extensive knowledge base for a disease or complaint, novice and experienced physicians alike are more likely to resort to the hypothesis generation and testing strategy discussed earlier.[ 12 ],[ 16],[ 17] Experts who have a finely honed knowledge of disease probably use this strategy principally when dealing with a particularly difficult diagnostic dilemma.[ 12 ]
The principles of diagnosis and therapy are inextricably intertwined. Because a diagnosis is an inference about a patient’s illness, we can never be absolutely certain that the disease label we assign to a patient’s illness is correct. For this reason, we will inevitably treat some patients who do not have the disease and inevitably fail to treat some who do. Both circumstances deprive some individuals of appropriate therapy.
To the extent that the treatment is effective but also produces harmful side effects, patients who have the disease for which the treatment is designated will derive the benefit of therapy, offset to some extent by the risk of therapy. Treated patients who do not have the disease, however, derive no therapeutic benefit but nonetheless are subjected to the risk.
Linking Diagnosis and Treatment
The interplay between diagnostic hypotheses and the benefits and risks of tests and treatments can be envisioned effectively in terms of decision thresholds, a concept derived from decision science. A threshold is the probability of a disease at the point at which two choices (e.g., treating vs. not treating; treating vs. further testing) have equivalent value.
The threshold is thus a benchmark for action: At disease probabilities lower than the threshold, one action is appropriate, whereas at disease probabilities greater than the threshold, a different action is appropriate. A threshold can be calculated using the methods of decision analysis from data on the benefits and risks of diagnostic/tests and treatments, or it can be estimated.
Thresholds define diagnostic and therapeutic interactions. When deciding whether or not to administer a treatment for a suspected disease, the efficacy and risks of the treatment for the disease determine how confident a physician must be in the diagnosis to make treating the patient a better choice than not treating.
For treatments with a high ratio of benefits to risks, the therapeutic threshold is quite low, and treatment can be given even when the probability of disease is relatively low (e.g., penicillin for suspected streptococcal throat infections). For treatments with a low ratio of benefits to risks, on the other hand, the therapeutic threshold is quite high, and the physician must be quite certain that the patient has a given disease before administering therapy (e.g., thrombolytic therapy for suspected myocardial infarction). Of course, low efficacy of treatment, high risk, or both can contribute to such a low ratio of benefits to risks.