Chapter 3: Refinement of Diagnostic Hypotheses

Where Refinement Begins and Ends

( Case 1, Case 12 , Case 13, Case 16, Case 17, Case 37, Case 38)

After hypotheses are evoked, the process of hypothesis refinement, also known as “case building,” ensues. Hypothesis refinement is an evolving, sequential process of data gathering and interpretation. Repeated inferences yield a series of provisional approximations (intermediate diagnostic hypotheses) that are revised continually in an iterative process until one or more diagnostic hypotheses satisfactorily explain all available clinical data. The process begins with a small number of hypotheses generated from a set of clinical findings. It proceeds by elaborating questions that elicit further data and by interpreting the data obtained. Initial hypotheses are revised, refined, and often made more specific.

Some hypotheses are added and some are deleted. The process of hypothesis refinement uses a variety of reasoning strategies (probabilistic, causal, and deterministic) and often involves the use of diagnostic tests to discriminate among existing hypotheses. After relevant data are accumulated, diagnostic refinement merges into diagnostic verification—the process in which one or more hypotheses are accepted as sufficiently valid to permit further decision making (testing, therapeutic, or prognostic).

This chapter considers details of the refinement process. Subsequent chapters describe the probabilistic approach to combining clinical data and the use of causal (physiologic) reasoning in the process. Causal reasoning, which depends on the cause-and-effect relations between clinical variables, is discussed later because it functions chiefly in the later phases of the diagnostic process as we attempt to verify our hypothesis.

Context and Diagnostic Classification

( Case 7, Case 9, Case 63, Case 64)

The context within which problem solving occurs is a function of the cognitive representation of the problem in memory. As noted in Chapter 2 , this context can be as general as “infection” or as specific as “unexplained hypoglycemia.” The context frames the problem, constrains the number of possible explanations, sets a limit on the number of operations to be applied to the problem, and serves as a basis for expectations.[ 30], [ 31]–[ 32 ] These expectations are the predictable, anticipated findings and are based on some mental model of the disease.

When an attempt is being made to classify a given patient within a diagnostic hypothesis, the varied characteristics of the clinical disorder become the basis for such expectations. Given a hypothesis of acute appendicitis, for example, features expected in appendicitis (right-lower-quadrant tenderness, leukocytosis) are sought as more clinical data are obtained. In addition, features not expected in appendicitis (disorientation, cough, normal white cell count) can be evaluated and explained. Thus, the representation of disease entities in memory is a critical factor.[ 8],[ 12 ],[ 13]

A central question is how new instances (a patient with certain clinical manifestations) are compared to existing entities in memory and indexed or classified. Clearly, the characteristics of the context critically determine the efficiency and accuracy of diagnostic refinement. If a clinical entity (appendicitis again as the example) is defined narrowly according to its textbook description or according to its typical or classical descriptions, features that occur in its variants (such as diarrhea when the appendix is in a retrocecal location) might be considered to exclude the diagnosis. Repeated experience with variations of disease entities fills out the expectations, that is, the normal and abnormal findings that are associated with a given entity.

How is such experience stored and accessed? For many years, it has been assumed that information about a new case is compared to some case prototype, or abstract model, stored in memory.[ 6],[ 33] It has been further assumed that the abstract description is sufficiently detailed to contain all variations of the disease, as well as rules about how the disease relates to other diseases or conditions. This theory argues that memory consists of abstract descriptions that evolve by compiling and compressing information into a single model, or prototype, as we encounter more and more patients with a certain disease. To the extent that a single abstract model could exist, a new case would be assessed by comparison to this abstract description.

Another theory, based on studies in the domain of case-based reasoning, supposes that knowledge is stored in a symbolic structure known as a script.[ 6],[ 13],[ 34] In medicine a disease script would comprise patient-specific scenarios containing personal features, predisposing factors, causative agents, and clinical manifestations tied together both by causal links and chronological relations. A script might consist of a description of an illness, the natural course of the illness, the possible interventions, the sequences of events, and the outcomes. Scripts could vary from representing clinical data in a highly physiologic format on one hand (i.e., containing a detailed causal model or physiologic or anatomic model) to a smaller, more efficient, highly compiled format on the other (i.e., containing only relations between findings in the form of diagnostic labels, e.g., radiologic, pathologic, or dermatologic findings).

One recently developed concept holds that much of the indexing or classification by physicians of new instances is carried out not against a single prototype of the disease but against multiple stored prototypes or even actual recalled cases (instance scripts or exemplars) of the disease seen by a physician in the past. Given the range of manifestations seen in a set of patients with a single disease entity (i.e., the polymorphism of that disease), the notion that multiple cases are stored in memory for later comparison with new cases is attractive. A more detailed discussion of the structure of memory is given in Chapter 10, section The Structure of Memory.

Hypothesis Evolution

( Case 1, Case 10, Case 12 , Case 16, Case 33)

Although initial diagnostic hypotheses provide the framework for data gathering, they may or may not survive. When new data are consistent with an existing mental model (however it is constructed), the hypothesis remains active and may become even more specific. A hypothesis of “infection” may evolve into “urinary tract infection,” then into “pyelonephritis,” and finally into “left-sided Escherichia coli pyelonephritis.” Alternatively, hypotheses deemed interesting initially may be dropped quickly when further data fail to support them. The process should not be viewed as an orderly one in which hypotheses that initially are quite vague always are progressively specified.

Although this pattern does occur, others are observed as well. An initial hypothesis may be highly specific (e.g., Cushing syndrome), and it may not change as more information is obtained. Usually, diagnostic hypotheses become more or less credible with each new clinical datum, but hypotheses may disappear only to reappear later. A given hypothesis may be considered highly probable when only a few cues are available; later it may be nearly dismissed only to become prominent again when all available data are obtained. A diagnostic hypothesis may have to be abandoned when data appear that are inconsistent with it.

In such instances, replacement hypotheses must be generated to account for the data. It seems quite likely that clinicians do not simply continue to collect hypotheses indefinitely during a diagnostic encounter, only to narrow down to one or two after all information has been gathered. Rather, evidence is strong that the cognitive limitation of working memory to a small number of items constrains the number of hypotheses in active memory.[ 27],[ 30] This constraint probably pertains to the concept of differential diagnosis, as discussed later.

Sequence of Data Collection

( Case 14, Case 24, Case 45)

Clinical data need not be accumulated according to a fixed pattern. Although data are typically sought first from the history, then from the physical examination, and then from the laboratory, this pattern of data gathering is more a matter of historical precedent than of cognitive necessity. In fact, data may first emerge from a patient’s physical appearance (gait, tremor, or facial features), from the laboratory (an unexpected low hematocrit or a high serum calcium), or from a test (a blood pressure measurement made in a shopping mall). Hypothesis refinement demands no special sequence of data collection, although some optimal sequence probably does exist.

Initially, expert clinicians do focus heavily on data from the patient’s history and previous records (a particularly rich data source), but they readily switch to an aspect of the physical examination or a diagnostic test in the interest of gathering a pertinent piece of data whenever appropriate. On the other hand, conceding that it is appropriate to gather data out of sequence does not invalidate either the traditional questions asked as part of a “review of systems” or the “routine” physical examination. Such approaches have valid goals, including gathering of baseline data, avoidance of errors in drug administration, identification of risk factors, case finding for diseases that are uncommon but important to identify, and disclosure of critical psychological and social issues.

The sequence of data accumulation has increased in importance, given the foreshortened pace of medical diagnosis, especially in emergency departments, where the rapid triage of patients often begins with a brief acquisition of a patient’s presenting complaint and is followed immediately by the ordering of one or more diagnostic tests. Whether it is more efficient, and just as accurate, to “short-circuit” the diagnostic process in this manner has never been evaluated. Until we learn more about the benefits and risks of this approach, we continue to recommend the process described previously.

Reducing Diagnostic Uncertainty

( Case 9, Case 12 , Case 15, Case 18, Case 38)

Early in the process of hypothesis revision when only a small set of cues is available, the number of possible disorders that could explain this set of cues often is quite large. At this stage, diagnostic uncertainty is at its highest (i.e., differentiation among the various diagnostic hypotheses is at its nadir) and the number of questions that a physician might ask to elicit the data needed to narrow the number of hypotheses is at its peak. The process that the physician uses to gather data follows no preordained pattern and in this framework can be characterized as unstructured problem solving, yet diagnostic hypotheses do lend some structure to the process.

Most of the time a lock-step or algorithmic method cannot substitute for this unstructured approach simply because of the large problem space (i.e., the constrained environment that guides the possible operations and solutions to a problem) in which the problem must be solved. Questioning is guided by hypotheses, which may be related to probabilistic relations between clinical variables. Diagnostic efficiency requires that the questions that are asked are the ones most likely to reduce diagnostic uncertainty. To do so requires that the data obtained from such a question, whether positive or negative, should produce the largest change in disease probability.

Several strategies for eliciting information are used.[ 35], [ 36], [ 37]–[ 38] One is a confirmation strategy, in which information is sought that might be expected to enhance a highly likely hypothesis.[ 19] Another is a disconfirming or elimination strategy, in which information is sought to reduce the likelihood of an unlikely hypothesis. Of course, when either of these strategies alters the likelihood of any hypothesis, the likelihood of one or more remaining hypotheses also must change. A reduction in the likelihood of a leading hypothesis, for example, forces remaining hypotheses to be more prominent.

When only a few possibilities remain, a discrimination strategy can be invoked to seek specific information to discriminate among these remaining hypotheses.[ 19] Frequently these few diseases bear close resemblance to each other in their clinical manifestations (e.g., constrictive pericarditis and severe biventricular failure; or polyarteritis nodosa and systemic atheroembolism) and are often mistaken for one another. In such instances, differences in the prevalence of the disorders, subtle differences in the clinical characteristics of each, and the results of specific laboratory tests may be required to discriminate among completing diagnostic entities. In some instances, the response to therapy becomes a final discriminator.

The process of hypothesis refinement can be carried out mathematically, but expert clinicians rarely rely on formal probabilistic models as they engage in diagnostic reasoning. Instead, they use a variety of rules of thumb or heuristics previously described. These simplifications are useful short-cuts, and although they are not precise reflectors of prevalence or other probabilistic associations between clinical variables, they are convenient and frequently correct. As uncertainty increases, physicians rely even more on their clinical intuition.[ 39]

The goals of questioning and data accumulation are several: to identify highly likely diagnostic hypotheses, to disprove unlikely hypotheses, to forge causal links between clinical phenomena, to differentiate among existing hypotheses, and, as noted before, to find hypotheses that are particularly critical to preserving a patient’s well-being (diagnostic imperatives).[ 40]

The Differential Diagnosis

( Case 10, Case 18, Case 38)

As attempts are made to refine hypotheses, clinicians often assemble a list of surviving, competing hypotheses commonly known as a differential diagnosis. However, no single definition of a differential diagnosis is universally accepted. Such lists are assembled early in the process from single or multiple cues, and they may or may not be ordered according to some hierarchy (such as physiologic categories or disease probabilities).

Some clinicians define a differential diagnosis as a small final set of remaining hypotheses for which the discrimination strategy described previously is used.[ 41],[ 42 ] We prefer to consider the entire process of hypothesis refinement as one that differentiates among diagnostic possibilities. According to this definition, a differential diagnosis comprises the entire evolving, sequential, and iterative diagnostic process from generation of hypotheses to establishment of the working diagnosis.

Relation to Formal Probabilistic Approach

( Case 23, Case 27, Case 42 )

It is useful to set this evolving process against an explicit process of diagnostic revision that is based on probability theory and that uses Bayes’ rule for recalculating the likelihood of various diseases. This comparison is of particular value because of the close parallelism between the implicit reasoning processes that physicians use to revise and refine diagnostic hypotheses with new information and the formal, prescriptive process that calculates these revisions.[ 12 ]

Bayesian analysis requires that a physician assembles a complete set of diagnostic hypotheses that could explain a given set of clinical findings. For each hypothesis, a set of relevant attributes is identified (historical findings, physical findings, complications, predisposing factors, laboratory results) that might help discriminate among the diagnoses. The pretest or prior probability of each diagnostic hypothesis is specified numerically, as is the probability that each attribute is found in each disease entity (the conditional probability). Then, a calculation is made of the likelihood of each disease entity, given the disease prevalence and the probability of each clinical attribute. The resulting revised probabilities (the posterior probability) represent the likelihood of various disease entities, given the prevalence and the presence of the specified attributes.

This process requires that all possible diseases be specified prospectively because omitting even rare possibilities may eliminate the correct diagnosis. As long as a complete set of diagnostic hypotheses is assembled at the onset of an analysis, Bayes’ rule can be applied sequentially as information is gathered. Thus, Bayesian analysis is best applied after considerable data are already available. Bayesian analysis seeks to combine information as a clinician would, but according to formal mathematical rules. A detailed example of how Bayesian analysis is used in diagnostic hypothesis revision when multiple diseases and multiple attributes of these diseases are under consideration is given in Chapter 4, section Bayesian Revision for Multiple Diseases with Multiple Attributes.