DIAGNOSIS AND TREATMENT

Diagnostic Reasoning
Jerome P. Kassirer, M D

Research in cognitive science, decision sciences, and artifi- probabilistic, causal, and deterministic reasoning stra-
cial intelligence has yielded substantial insights into the tegies. Finally, we consider the interrelationships be-
nature of diagnostic reasoning. Many elements of the diag- tween reasoning strategies and the circumstances in
which one strategy may be preferable to another. The
nostic process have been identified, and many principles of
discussion centers on the cognitive aspects of diagno-
effective clinical reasoning have been formulated. Three rea-
sis, not on the other aspects of the patient encounter
soning strategies are considered here: probabilistic, causal, (social and psychological factors, interviewing tech-
and deterministic. Probabilistic reasoning relies on the statis- niques, or precision of information).
tical relations between clinical variables and is frequently
used in formal calculations of disease likelihoods. Probabilis-
Elements of the Diagnostic Process
tic reasoning is especially useful in evoking diagnostic hy-
potheses and in assessing the significance of clinical findings In the diagnostic process, the clinician makes a series
and test results. Causal reasoning builds a physiologic model of inferences about the nature of malfunctions of the
and assesses a patient's findings for coherency and complete- body. These inferences are derived from existing ob-
servations (historical data, physical findings, and rou-
ness against the model; it functions especially effectively in
tine tests) as well as from invasive tests and responses
verification of diagnostic hypotheses. Deterministic reason-
to various manipulations. Inferential reasoning pro-
ing consists of sets of compiled rules generated from routine, ceeds until the clinician has discovered a diagnostic
well-defined practices. Much human problem solving may category sufficiently acceptable to either establish a
derive from activation and implementation of such rules. A prognosis, yield a therapeutic action, or both ( 1 ) .
deeper understanding of clinical cognition should enhance When making diagnostic inferences from clinical data,
clinical teaching and patient care. the clinician uses many strategies to combine, inte-
grate, and interpret the data (2, 3). Clinicians make
Annals of Internal Medicine. 1989;110:893-900. extensive use of rules of thumb, or shortcuts (heuris-
tics) in the process of gathering and interpreting infor-
From Tufts University School of Medicine and the New mation. These heuristics are essential: they reduce the
England Medical Center, Boston, Massachusetts. For need to ask an unnecessarily large number of questions
current author address, see end of text. and to order superfluous diagnostic tests, and they
make the task of information gathering manageable
and efficient (4, 5).
Studies of human cognition suggest that the kinds of
heuristics used depend on the nature of the clinical
lifficient and accurate diagnosis is one of the foremost problem being addressed and on the expertise of the
cognitive functions of the physician, and in the past clinician. Nonexperts tend to use nonselective search-
decade many fundamental features of diagnostic rea- and-seek strategies that, although they are applicable
soning have been disclosed. Probabilistic approaches across a wide range of clinical settings, are weak in
have dominated recent discussions of diagnosis, be- generating specific hypotheses (6-9). Experts, howev-
cause iterative engineering has converted them into er, typically employ strong diagnostic approaches tai-
practical clinical tools. Eclipsed by the probabilistic lored to their expertise. They quickly focus on a prob-
approach, however, are other powerful concepts that lem by recognizing patterns, formulating problems in
embody fundamental principles and important cogni- semantically meaningful "chunks," gathering data rel-
tive strategies of diagnostic reasoning. These strate- evant to a solution to the problem, and applying pre-
gies—causal reasoning and deterministic (rule-based) compiled actions (2, 7, 9-15).
reasoning—are derived from research in new disci- The diagnostic process focuses on one or more
plines that have only just begun to influence medical evolving diagnostic hypotheses (2, 3, 16, 17). A hy-
practice, namely cognitive science, cognitive psycholo- pothesis can be either general or specific. It can take
gy, and artificial intelligence. Many insights from several forms, including a state (inflammatory pro-
these fields can be applied directly to the medical diag- cess), a clinical disorder (acute transplant rejection),
nostic process. a syndrome (nephrotic syndrome), or a specific dis-
We discuss elements of diagnostic reasoning. First, ease entity (polycythemia vera) (1-3, 18). Hypothesis
we give an overview of the cognitive aspects of the generation, a process described as diagnostic trigger-
diagnostic process, and then discuss the principles of ing, is the first step at the inception of a diagnostic

©1989 American College of Physicians 893

Downloaded From: http://annals.org/pdfaccess.ashx?url=/data/journals/aim/19681/ by a Scott Memorial Library User on 04/05/2017

 encounter (2, 3, 16, 17, 19). Triggering, the formula- The probabilistic approach is based on associations
tion of a preliminary hypothesis on the basis of only a between clinical variables, usually described as statisti-
few observations is critically dependent on the cogni- cal relationships. This approach utilizes data such as
tive ability to relate a new situation to past experience. the prevalence of coronary disease in asymptomatic
Typically, the clinician triggers initial hypotheses 40-year-old men, the frequency of hypocomplemente-
merely from a patient's age, sex, race, and presenting mia in membranoproliferative glomerulonephritis, and
complaints, but sometimes such hypotheses emerge the frequency of false-positive tests for antibodies to
exclusively from physical findings or laboratory data the human immunodeficiency virus ( H I V ) in various
(2, 3, 16, 17). Additional hypotheses are triggered as populations at risk for the acquired immunodeficiency
new findings emerge. syndrome ( A I D S ) .
Hypotheses serve an essential function: They form a Causal reasoning is based on the physiologic or
context within which further information gathering cause-and-effect relations between clinical variables.
takes place (20, 21). This context, a diagnostic catego- For example, the diagnosis of hyperthyroidism is
ry of some kind (acute bacterial meningitis, for exam- strengthened in a patient with "soft" clinical manifes-
ple), provides a model against which a given patient's tations of the disease and a moderate elevation of plas-
findings can be assessed (2, 21, 22). The diagnostic ma thyroxine if the level of thyroid stimulating
category specifies the findings that should be present hormone is also found to be suppressed, because high
and those that should be absent if the patient has a levels of thyroid hormone inhibit pituitary production
given disorder. Diagnostic reasoning proceeds by pro- of thyroid-stimulating hormone; a possible diagnosis
gressive hypothesis modification and refinement. Some of superior vena caval syndrome is strengthened if a
hypotheses are made more specific, some previously mass lesion is identified in the right side of the me-
triggered hypotheses are deleted, and some new ones diastinum on chest roentgenogram, because mass le-
are added. Evidence suggests that only about seven sions in this location are known to obstruct the superi-
hypotheses are active at any one time, constrained by or vena cava.
the limited capacity of short-term memory (2, 23-25). Deterministic, or categorical reasoning, is based on
It is not clear how much of the diagnostic process is compiled strategies in the form of well-defined rules.
driven by hypotheses, and how much is driven by the Examples of this approach include: If plasma bicar-
availability of data from the physical examination or bonate is less than 8 m m o l / L and blood p H is low, the
laboratory (the so-called data-driven approach [7, 14, diagnosis is metabolic acidosis. If an elderly patient
2 6 ] ) . Possibly, elements of the hypothesis-driven and has patches of osteosclerosis in the pelvis, check the
data-driven approaches are intermingled. alkaline phosphatase level, and if it is elevated, the
Verifying a diagnostic hypothesis is the next, and diagnosis is Paget disease.
often penultimate task. Because the diagnostic process Although these three reasoning strategies are pre-
is inferential, all diagnostic hypotheses necessarily re- sented as separate and distinct entities with well-de-
flect a belief or a conviction by the physician regarding fined boundaries, we show how these strategies over-
the nature of the patient's condition. Verifying a hy- lap and complement each other.
pothesis involves assessing its coherency (for the pa-
tient in question, are all physiologic linkages, predis- Probabilistic Reasoning
posing factors, and complications appropriate for the
suspected disease), its adequacy (does the suspected Medical diagnosis and test interpretation involves con-
disease encompass all the patient's findings, normal siderable uncertainty, which can be represented as
and abnormal), and its parsimonious nature (is the probabilistic associations between clinical variables
suspected disease a simple explanation of the patient's (30, 31). A probability represents a belief in the state
findings) ( 2 ) . It also requires eliminating competing of a patient or in the meaning of a test result: We may
hypotheses in a process analogous to disproving com- believe, for example, that a patient may be either high-
peting scientific hypotheses (can any other diseases ly likely or highly unlikely to have a given disease, and
better explain the patient's findings) ( 7 ) . This process that a certain test result either increases or decreases
produces one or more working hypotheses that form the likelihood of that disease. In assessing clinical find-
the basis for the next step in patient management—ob- ings and test results, these relations among clinical
serving the patient, ordering additional tests, or treat- variables must be integrated. When combining such
ing the patient (27-30). data, clinicians ordinarily use ill-defined expressions
for likelihood such as "quite likely," "common," "fre-
quent," and "rare." However, because nonquantitative
terms do not have standardized meanings (32-34), the
Types of Diagnostic Reasoning
clinician's ability to combine clinical data character-
There exists no comprehensive description of diagnos- ized by such nonquantitative measures of uncertainty
tic problem solving (indeed, no generally accepted is compromised (35). A more satisfactory technique
theory of human problem solving exists), yet some of for interpreting clinical data uses formal probabilities
the elements used to carry out the inferences described and likelihood ratios, and combines such probabilities
above have been elaborated. We discuss three clinical quantitatively (36). We will discuss many of the im-
reasoning techniques: probabilistic, causal, and deter- portant principles of probability theory that relate to
ministic reasoning. interpretation of tests and other clinical findings.

894 1 June 1989 • Annals of Internal Medicine • Volume 110 • Number 11

Downloaded From: http://annals.org/pdfaccess.ashx?url=/data/journals/aim/19681/ by a Scott Memorial Library User on 04/05/2017

 Requirements

To combine clinical findings in the form of probabili- If all these precautions are heeded, the probabilistic
ties requires a rigorous definition of disease (30). This approach can be a valuable tool. Most important, it
definition, referred to as the "gold standard," often is provides an explicit means of test interpretation and
based on histologic findings or radiologic studies, but thus helps to illuminate the diagnostic process. Aside
sometimes it must be based on clinical criteria alone. from its value in analyzing complex and disparate
The first task in applying the probabilistic approach is data, the probabilistic approach is helpful in hypothe-
estimating the prevalence, or prior probability, of all sis generation because associations between clinical
diseases of interest from the patient's demographic findings trigger diagnostic hypotheses according to
data and presenting clinical findings. The prior proba- their frequency. For example, crushing chest pain in a
bility for an individual patient must reflect as closely 60-year-old man promptly conjures up a diagnosis of
as possible the population from which that patient is acute myocardial infarction; whereas acute pleuritic
derived. A carefully selected assessment of this proba- chest pain in a 20-year-old man triggers the possibility
bility is critical because its value has a major influence of spontaneous pneumothorax. The probabilistic ap-
on the interpretation of any clinical feature or test re- proach also teaches many important principles of di-
sult (30, 31, 37). The second task is to determine the agnostic testing: If the pretest probability of a disease
frequency of features associated with a defined disease is extremely low, a positive test result will rarely con-
entity, namely the conditional probabilities (30, 31, firm a diagnosis unless the test is highly specific (the
37). Typical questions that could be used to derive test result probably will be false-positive). When the
these data might include: How often is plasma IgA likelihood of a given disease is extremely high, a nega-
elevated in the entity; how often is it slightly, moder- tive test result usually does not exclude the presence of
ately, and markedly elevated? This information must the disease unless the test is highly sensitive.
be coupled with similar data about other diseases
under consideration and about findings in normal indi-
Taking the Next Step
viduals (38). The sensitivity and specificity of diag-
nostic tests are special terms for certain kinds of con- Both nonquantitative working hypotheses and calcu-
ditional probabilities (30, 31, 36). lated posterior probabilities are constructed for the
Given an appropriate gold standard and data on the purpose of taking action. The action may be only to
prior and conditional probabilities of the set of diseas- arrive at a certain forecast about the patient's subse-
es in question, clinical findings and test results can be quent clinical course, but most often it is a plan of
interpreted according to mathematical principles (30, further testing or treatment or sometimes no more
31, 36, 39). The output of this quantitative approach testing or treatment. These choices are a function of
is a revised view of the likelihood that a patient has the probability that a patient has one or more diseases,
one or more diseases, a concept known as the posterior the benefits to be derived from further testing, the
probability. The numerical values in these calculations risks of further testing, and the benefits and risks of
are ordinarily based on data from the literature, but the treatment.
when such information is unavailable, subjective as- These concepts are elaborated in the form of testing
sessments of prior and conditional probabilities may and treatment thresholds (27, 28). A threshold can be
be used. calculated using the methods of decision analysis from
data on the benefits and risks of diagnostic tests and
Caveats treatments (27, 28), or it can be estimated (29). A
threshold is the probability of a disease at which two
This process requires particular care. The list of possi- choices (for example, treating or not treating) have
ble diagnoses must be exhaustive, because if one diag- equivalent value. The threshold is thus a benchmark
nosis is left out it will never emerge as a potential for action: At disease probabilities lower than the
cause of the patient's symptoms or signs (39). Many threshold one action is appropriate, whereas at disease
test results cannot be described only as positive or neg- probabilities greater than the threshold a different ac-
ative, and for such tests it may be necessary to de- tion is appropriate.
scribe several levels of positivity. To do this, results Suspected pulmonary embolism provides an exam-
that are continuous variables usually must be broken ple of how thresholds can be applied. If at one ex-
into discrete intervals to use in calculations (39). To treme, the suspicion of pulmonary embolism, based on
avoid a repeat count of the same information, each clinical findings and a ventilation-perfusion scan is ex-
disease must be mutually exclusive of all other diseases tremely low, we should do no more tests and not give
and each conditional probability used in a calculation treatment. If, at the other extreme, the physician be-
must be independent of the others (30). Finally, cer- lieves there is a near certainty that a patient has had a
tain diseases cannot be appropriately considered as ei- pulmonary embolism, it is preferable to give anticoag-
ther present or absent. Because stages of diseases often ulants without more testing. Between these two ex-
have different manifestations, any quantitative analysis tremes, more testing with pulmonary arteriography
must recognize the clinical manifestations and test re- might or might not be the optimal strategy. Using data
sults in various stages of the disease in terms of sever- on the accuracy and risks of pulmonary arteriography
ity or time course, or both (40). and data on the efficacy and risks of anticoagulants,

1 June 1989 • Annate of Internal Medicine • Volume 110 • Number 11 895

Downloaded From: http://annals.org/pdfaccess.ashx?url=/data/journals/aim/19681/ by a Scott Memorial Library User on 04/05/2017

 the no-treatment/testing threshold and the testing/ clinical settings. Such a model is a coherent system
treatment threshold can be calculated, providing spe- capable of explaining its components, the range of pos-
cific guidance concerning how low the probability of sible variations, and the nature of findings in a particu-
pulmonary embolism must be before choosing not to lar patient. If the same model is not applicable for
give treatment and not to order an arteriogram, and another patient with a problem similar to that for
how high the probability must be before abandoning which it was created, additional physiologic features
arteriography and treating with anticoagulants. At might be required. Physiologic models exist in all ar-
probabilities of pulmonary embolism between these eas in medicine.
two thresholds, the optimal choice is to carry out arte- For example, in models for fluid and electrolyte
riography because the study results will determine equilibrium (47), assume that a patient with clinical
whether to administer anticoagulants. and laboratory findings suggestive of the syndrome
Implicit in the threshold approach is the concept of inappropriate secretion of antidiuretic hormone
that although a diagnosis has a greater or lesser degree ( S I A D H ) has a high urinary sodium excretion. Does
of uncertainty, this uncertainty should not impede the this finding influence the suspected diagnosis? We
clinician from making a testing or treatment decision. could assess this finding in a probabilistic framework
When formal threshold calculations have not been (for example, we might say that 8 5 % to 9 5 % of pa-
done, the choices to treat, to continue testing, or to tients with S I A D H have a high sodium excretion) or,
withhold testing or treatment should be made after alternatively, by understanding the pathophysiology of
consideration of the disease likelihood and the benefits the syndrome, we could examine how the finding
and risks of relevant tests and treatments. "fits" with the diagnosis. If our model of S I A D H con-
tains (as it should) the concepts that such patients are
volume expanded, that volume expansion promotes
Causal Reasoning
sodium excretion, and that sodium excretion in the
We are seldom explicit about the diagnostic value of syndrome typically matches sodium intake, we would
the cause-and-efFect relations among clinical variables readily understand that a high urine sodium excretion
although we expend enormous resources on the inves- is consistent with a diagnosis of S I A D H and adds to
tigation and teaching of pathophysiology. Yet causal the credibility of the diagnosis. We also would be in a
(physiologic) reasoning remains a fundamental un- position to explain a low urine sodium excretion if that
derpinning of diagnostic reasoning, and, when applica- were the finding instead. In this instance, despite the
ble, it enhances the diagnostic process with the rich- presence of S I A D H , urine sodium is low presumably
ness of its explanatory power. because the patient is ingesting little salt.
Causal reasoning relies on the cause-and-effect rela-
tions between clinical variables or chains of variables Judging Causality
( 1 , 18, 41-46). It derives from common sense notions
of causality. For example, if we see a field of shriveled Research in causality in nonmedical contexts provides
crops, we might draw several inferences. First, the a systematic framework for judging cause-and-effect
crops might be lacking nutrients; second, cold weather relations as they apply to medical diagnosis (43, 4 8 ) .
might have affected them; and third, pollutants might We often are alerted to the possibility that we should
have poisoned them. If, however, we are already aware use a causal model when abnormal findings or events
that the region has been subject to a prolonged violate normal physiologic expectations. This devia-
drought, we could conclude that the cause of the with- tion produces the context within which further data
ered plants is lack of water because we know that lack gathering and interpreting take place. To carry out
of water dehydrates crops. This and countless other this interpretive process, we generate a causal model,
examples in medicine are based on a causal or physio- typically a chain of physiologically related features
logic model of reality; the capacity to make inferences consisting of stimuli and their responses.
from the observed clinical findings also depends on the When invoking a causal hypothesis involving two or
principles embedded in this model. more variables, the links between stimuli and respons-
A causal model might be defined as a description of es must be assessed for their strength (43, 4 8 ) . Several
mechanisms (anatomic, physiologic, biochemical) criteria can be used to assess the strength of this link.
that can be used to simulate the normal workings of Is the entire causal chain credible? Does a change in a
the human body, its pathophysiologic behavior in dis- response correlate closely with the change in the stim-
ease, and the idiosyncracies of a particular patient ulus? Is there substantial congruity of duration and
(42). The causal model supports a clinician's perform- magnitude between response and stimulus? Is there
ance by simulating possible courses of the disease and close contiguity in time and space between a response
its modification by treatment. It also serves as a coher- and a stimulus (did one event follow another closely
ency criterion on hypotheses about the patient, and enough to allow us to accept that the first event caused
provides a common framework for explanations and the second)? When these tests are satisfied, one gains
discussion among clinicians ( 4 2 ) . confidence that a given response and a suspected stim-
When applying causal reasoning, clinical variables ulus (or a given stimulus and a suspected response)
are examined and included if they help explain the are related.
model. A model is created for each patient, although a This process of testing, validating, verifying, and
single model can apply for many patients in various falsifying causal connections is a fundamental aspect

896 1 June 1989 • AnnaIs of Internal Medicine • Volume 110 • Number 11

Downloaded From: http://annals.org/pdfaccess.ashx?url=/data/journals/aim/19681/ by a Scott Memorial Library User on 04/05/2017

 of medical practice. The final process in assessing a Deterministic Reasoning
causal hypothesis is testing it for alternate possible ex-
Deterministic or categorical reasoning consists pre-
planations (43, 48). Because a given causal hypothesis
dominately of compiled knowledge from any source in
appears to explain a set of findings, it does not neces-
the form of unambiguous rules. Such knowledge may
sarily prove that this causal chain is the correct one.
originate in probabilistic or causal associations be-
Alternative constructions of the causal chain must be tween clinical findings. Because so many diagnostic
sought and their strengths assessed before accepting problems that physicians confront daily are simple
one model and not another. and straightforward, the exercise of the diagnostic
process for such problems is routine, well defined and
Application in Diagnosis thus, as a practical matter, deterministic.

Causal reasoning can be applied in several steps of the

Production Rules
diagnostic process. It may be useful in the formulation
of a context: If the possibility of a pathophysiologic To specify reasoning in a deterministic fashion re-
state has been triggered by some findings, the state quires that we identify the rules that describe our rou-
may provide the context for further data gathering tine practices. Such deterministic assertions, referred
(for example, the context of possible volume contrac- to as production rules, are in this form: "If (certain
tion should lead a physician to take a detailed history conditions are m e t ) , then (a certain action is appro-
of fluid loss). A causal model also can be used to help priate)." Investigators have built an elaborate hypoth-
verify a diagnosis: In a patient suspected of having esis on the nature of human problem solving using
hyperthyroidism from clinical findings and a slightly these kinds of if-then rules (6, 52-56). They assume
elevated plasma thyroxine, the finding of a suppressed that production rules are the basic building blocks of
level of thyroid stimulating hormone verifies the diag- human cognition and that they specify both a condi-
tion that must be met and an action that derives from
nosis. Assessing the coherency of a diagnosis, namely
satisfaction of that condition (6, 52, 56). The condi-
determining whether the physiologic or causal associa-
tion is presumably a recognizable, semantically mean-
tions are reasonable and appropriate, is frequently a
ingful piece of information, and the action is a speci-
late step in gaining confidence in a diagnosis. This step fied, predetermined consequence that derives from the
involves determining whether a patient's findings are triggering of the condition (6, 25). Such production
consistent with recognized pathophysiologic manifes- rules, the theory argues, form the framework for de-
tations of a suspected disease. veloping the mental models that enhance human prob-
lem solving. Without them we would have to rely on a
Benefits complex storage system of examples and episodes.
Rules can have many purposes, and can describe
An important strength of causal reasoning is its capac- therapeutic approaches or prognostic implications. An
ity to provide an explanation for a given finding, espe- example of a rule constructed specifically to describe a
cially when the relation is not immediately obvious diagnostic strategy might be: If a young, otherwise
from either probabilistic associations or from previ- healthy black woman has hilar adenopathy, consider a
ously compiled concepts. Another strength is the ca- diagnosis of sarcoidosis. Another compiled rule from
pacity to use such a model to return to first principles our S I A D H model: If S I A D H is present, and the pa-
when findings do not fit an idealized pattern. A causal tient is ingesting an average diet, expect renal salt
model also makes it possible to unite various clinical wasting. Presumably many rules operate in parallel,
findings in a common framework: for example, the ef- compete with others for satisfaction, and either win or
fect of dietary sodium intake and sodium excretion in lose depending on their strength or relevance (6, 53).
patients with S I A D H . A causal approach provides a
consistency check among related findings: Two com- Clinical Algorithms
mon findings may have a strong probabilistic (or
statistical) relationship, yet they may be causally in- In medicine, deterministic reasoning often has been
consistent. Causal reasoning can help identify such imbedded in a printed, one- or two-page branching
discrepancies. Proper application of the causal ap- algorithm, or flow chart. An algorithm is an ordered
proach yields a rigorous guide to therapy because the set of instructions for carrying out a specific task
treatment can be based on efforts to reverse the string through a structure of simple steps. Most algorithms
are designed by authors who have simulated their per-
of events that produced the disordered state. Finally,
ception of the diagnostic process for a particular prob-
causal reasoning also has been applied in the develop-
lem (57-64), but some have been derived from deci-
ment of computer-based "expert systems" for medical
sion analysis or published studies of a clinical problem
diagnosis. One program used several causal models at (65). Many algorithms are designed to help the clini-
increasing levels of detail to deal with fluid and elec- cian interpret single findings, for example, hyperkale-
trolyte disorders (49, 50), and another used a network mia. In the typical algorithm, each nonterminal node
of causal associations to deal with general problems of requests information and requires decisive and
internal medicine (51). unequivocal answers, which are the labels of the

1 June 1989 • Annals ofInternal Medicine • Volume 110 • Number 11 897

Downloaded From: http://annals.org/pdfaccess.ashx?url=/data/journals/aim/19681/ by a Scott Memorial Library User on 04/05/2017

 branches leaving that node. The terminal nodes of the achieved some success but are still undergoing devel-
algorithm represent outcomes, each also precise and opment.
explicit.
Flow charts are used in various ways. In some in- The Choice of an Approach
stances the user encounters a node that requires a test
result. In a flow chart for the differential diagnosis of When is it appropriate to use the various kinds of clin-
hyperkalemia, for example, a node might have a label ical reasoning described? Presumably many strategies
that asks whether renal function is normal, whether are used, as a reflection of the rich interplay character-
adrenal function is normal, or whether the patient is istic of human reasoning. Because much of our medi-
receiving certain drugs. Therefore, in this approach cal practice is repetitive and straightforward, many of
diagnostic tests considered necessary are specified, al- our decisions are guided by precompiled deterministic
though the rationale for their application is not usual- rules. Probabilistic or rule-based reasoning is more
ly described. A flow chart also may specify which tests likely to be successful than causal reasoning in initiat-
are appropriate as outcomes, or terminal nodes, of the ing inferences about causes from observed effects (that
branching logic. is, triggering hypotheses). In turn, the probabilistic
An algorithm that accurately and realistically repre- and rule-based strategies can identify reasonable hy-
potheses that form the context for efficient causal rea-
sents actual clinical decision making has a substantial
soning. Because of its capacity to build an explanatory
payoff: Because it requires each user to follow the
model and its ability to permit us to return to diagnos-
specified pathway exactly, the user cannot forget to
tic and therapeutic first principles, causal reasoning is
ask an important question or do an essential test ( 6 6 ) .
a powerful approach especially when a complete or
Algorithms are useful for some diagnostic problems,
nearly complete model can be constructed.
particularly those in which the logic can be defined
clearly and precisely. To solve complex problems, The various reasoning strategies described are com-
however, algorithms frequently become so complicat- plementary. Causal models are exclusively dependent
ed that they cannot be represented readily as a printed on fundamental knowledge about physiologic function
and dysfunction. These models are specific to disease
figure.
entities and independent of the patient population. By
Algorithms also have other problems (63, 64, 67,
contrast, probabilistic models are dependent on the
68). They often do not deal effectively with uncertain-
specific population from which the patient is drawn.
ty, which is a common feature of medical practice. Because diagnostic hypotheses are critically dependent
Some are flawed because they are based on the faulty on disease prevalence, causal reasoning is a weak ap-
opinion of an individual rather than on data derived proach when the required task is triggering such hy-
from experimentation or rigorous clinical observa- potheses, whereas probabilistic reasoning is strong.
tions. The notations defining states or branch points Once some hypothesis has proposed a possible cause,
are ambiguous in some algorithms, thus yielding an however, causal reasoning allows us to verify whether
incorrect pathway through the branches. An algo- the cause can explain the observation. The S I A D H
rithm developed for one purpose may provide an example illustrates the interplay of these reasoning
incorrect result if used in another context when the strategies. Once the diagnosis was triggered, the causal
entry point is the same for both. For example, an algo- model made it possible to check the appropriateness of
rithm designed exclusively for the diagnosis of urinary a high urinary sodium excretion. Causal models also
tract infection and that has entry symptoms of dysuria help us to understand when certain findings do not fit
and urinary frequency will provide the wrong answer within the framework of a given hypothesis. Such a
if the patient has vaginitis. Finally, most algorithms signal becomes a trigger for generating new hypothe-
are unsuitable for patients with multiple complaints, ses.
complex clinical problems, or multiple interacting dis- Reasoning strategies also interact in validating pro-
ease entities. babilistic models. These models require that each dis-
ease under consideration be mutually exclusive of all
Rule-Based Expert Systems others and that conditional probabilities under consid-
eration be independent of each other. Because causal
Another deterministic approach to diagnosis consists models encode dependence among the parameters
of identifying lists of rules that clinicians use. Exhaus- they encompass and provide an understanding of the
tively specifying all rules for a problem domain (rheu- relations between variables, they can identify circum-
matology, for example) is a formidable, yet probably stances in which the independent assumptions of a
feasible exercise. But using a long list of production probabilistic model are invalid and can provide valu-
rules is impractical because it requires searching able guidance for correcting a poorly constructed
through the set of them to ensure that all the appropri- model.
ate findings and rules are identified. Although a physi- Our knowledge of the diagnostic process remains
cian hardly could be expected to sift through a list to too limited to precisely define the clinical situations in
achieve a diagnosis or to assess the need for a test, a which one kind of reasoning is most desirable. We
computer program can do this easily (52, 54, 69-71). may find that an efficient, effective, expert clinician
Several computer programs have been written using uses all of these approaches, even in the same diagnos-
sets of rules; these expert systems already have tic encounter. We still have much to learn about diag-

898 1 June 1989 • Annals ofInternal Medicine • Volume 110 • Number 11

Downloaded From: http://annals.org/pdfaccess.ashx?url=/data/journals/aim/19681/ by a Scott Memorial Library User on 04/05/2017

 nosis, but recent research efforts on the process bode diagnostic probabilities at which clinicians initiate testing and treat-
ment. Med Decis Making. 1983;3:155-68.
well for the enhancement of our diagnostic efficiency 30. Sox HC J r . Probability theory in the use of diagnostic tests: an
and accuracy. introduction to critical study of the literature. Ann Intern Med.
1986;104:60-6.
Grant Support: Supported in part by grant # 5 R01 LM04493 from the 31. Griner PF, Mayewski RJ, Mushlin Al, Greenland P. Selection and
National Library of Medicine. interpretation of diagnostic tests and procedures: principles and ap-
plications. Ann Intern Med. 1981;94:557-92.
Requests for Reprints: Jerome P. Kassirer, M D , 750 Washington Street,
32. Berwick DM, Fineberg HV, Weinstein MC. When doctors meet
Boston, MA 02111.
numbers. Am J Med. 1981;71:991-8.
Current Author Address: Dr. Kassirer: 750 Washington Street, Boston, 33. Bryant GD, Norman GR. The communication of uncertainty. Annu
Massachusetts 02111. Conf Res Med Educ. 1979;18:205-8.
34. Beyth-Marom R. How probable is probable? A numerical transla-
tion of verbal probability expressions. J Forecasting. 1982;1:252-69.
35. Moskowitz A J, Kuipers BJ, Kassirer JP. Dealing with uncertainty,
References risks, and tradeoffs in clinical decisions: a cognitive science ap-
1. Feinstein AR. An analysis of diagnostic reasoning. II. The strategy proach. Ann Intern Med. 1988;108:435-49.
of intermediate decisions. Yale J Biol Med. 1973;46:264-83. 36. Sox HC J r . Probability theory in the use of diagnostic tests: applica-
2. Kassirer JP, Gorry GA. Clinical problem solving: a behavioral anal- tion to critical study of the literature. In: Sox HC, Jr., ed. Common
ysis. Ann Intern Med. 1978;89:245-55. Diagnostic Tests: Use and Interpretation. Philadelphia: American
3. Elstein AS, Shulman LS, Sprafka SA. Medical Problem Solving: An College of Physicians; 1987:1-17.
Analysis of Clinical Reasoning. Cambridge: Harvard University 37. Gorry GA, Kassirer JP, Essig A, Schwartz WB. Decision analysis
Press; 1978. as the basis for computer-aided management of acute renal failure.
4. Tversky A, Kahneman D. Judgment under uncertainty: heuristics Am J Med. 1973;55:473-84.
and biases. Science. 1974; 185:1124-31. 38. Gorry GA, Pauker SG, Schwartz WB. The diagnostic importance of
5. Kahneman D, Slovic P, Tversky A, eds. Judgment Under Uncertain- the normal finding. N Engl J Med. 1978;298:486-9.
ty: Heuristics and Biases. Cambridge: Cambridge University 39. Pauker SG. Clinical decision making. In: Wyngaarden JB, Smith
Press; 1982. LH, Jr., eds. Cecil Textbook of Medicine. 18th ed. Philadelphia:
6. Holland JH, Holyoak KJ, Nisbett RE, Thagard PR. Induction: W.B. Saunders; 1988:74-9.
Processes of Inference, Learning, and Discovery. Cambridge: MIT 40. Ransohoff DF, Feinstein AR. Problems of spectrum and bias in
Press; 1986. evaluating the efficacy of diagnostic tests. N Engl J Med.
7. Langley P, Simon HA, Bradshaw GL, Zytkow JM. Scientific Dis- 1978;299:926-30.
covery: Computational Explorations of the Creative Process. Cam- 41. Kuipers B. Commonsense reasoning about causality: deriving behav-
bridge: MIT Press; 1987. ior from structure. Artificial Intelligence. 1984;24:169-203.
8. Newell A. Artificial intelligence and the concept of mind. In: Shank 42. Kuipers BJ, Kassirer JP. Causal reasoning in medicine: analysis of a
RC, Colby KM, eds. Computer Models of Thought and Language. protocol. Cognitive Science. 1984;8:363-85.
W.H. Freeman; San Francisco, 1973:1-60. 43. Kassirer JP, Kopelman RI. Judging causality. Hosp Pract.
9. Stillings NA, Feinstein MH, Garfield JL, et al. Cognitive Science: 1987;22:43, 45-6, 48, 50.
An Introduction. Cambridge: MIT Press; 1987. 44. Kassirer JP, Kopelman RI. The case for causal reasoning. Hosp
10. Eddy DM, Clanton CH. The art of diagnosis: solving the clinicopa- Pract. 1987;22:25-6, 30, 32, 35-6.
thological exercise. N Engl J Med. 1982;306:1263-8. 45. Patil RS. Review of causal reasoning in medical diagnosis. In: Pro-
11. Larkin JH, McDermott J, Simon DP, Simon HA. Expert and nov- ceedings of the Tenth Annual Symposium on Computer Applica-
ice performance in solving physics problems. Science. 1980; tions in Medical Care. Washington: IEEE Computer Society Press;
208:1335-42. 1986;ll-5.
12. DeGroot AD. Thought and Choice in Chess. 2d edition. The Hague: 46. Patil RS, Szolovits, Schwartz WB. Causal understanding of patient
Moulton Publishers; 1978. illness in medical diagnosis. In: Clancey WJ, Shortliffe EH, eds.
13. Kassirer JP, Kuipers BJ, Gorry GA. Toward a theory of clinical Readings in Medical Artificial Intelligence: The First Decade. Read-
expertise. Am J Med. 1982;73:251 -9. ing, Massachusetts: Addison-Wesley; 1984:339-60.
14. Patel VL, Groen GI. Knowledge-based solution strategies in medi- 47. Beck LH, Kassirer JP. Serum electrolytes, serum osmolality, blood
cal reasoning. Cognitive Science. 1986;10:91-116. urea nitrogen, and serum creatinine. In: Sox HC, Jr., ed. Common
15. Chi MTH, Feltovich PJ, Glaser R. Categorization and representa- Diagnostic Tests: Use and Interpretation. Philadelphia: American
tion of physics problems by experts and novices. Cognitive Science. College of Physicians; 1987:305-29.
1982;6:121-52. 48. Einhorn HJ, Hogarth RM. Judging probable cause. Psychol Bull.
16. Barrows HS, Bennett K. The diagnostic (problem solving) skill of 1986;99:3-19.
the neurologist: experimental studies and their implications for neu- 49. Patil RS. Causal Representation of Patient Illness for Electrolyte
rological training. Arch Neurol. 1972;26:273-7. and Acid-Base Diagnosis. Cambridge: MIT Laboratory for Comput-
17. Barrows HS, Norman GR, Neufield VR, Feightner JW. The clinical er Science; 1981.
reasoning of randomly selected physicians in general medical prac- 50. Szolovits P, Patil RS, Schwartz WB. Artificial intelligence in medi-
tice. Clin Invest Med. 1982;5:49-55. cal diagnosis. Ann Intern Med. 1988;108:80-7.
18. Feinstein AR. An analysis of diagnostic reasoning. I. The domains 51. Pople HE Jr. Heuristic methods for imposing structure on ill-struc-
and disorders of clinical macrobiology. Yale J Biol Med. 1973; tured problems: the structuring of medical diagnostics. In: Szolovitz
46:212-32. P., ed. Artificial Intelligence in Medicine. Boulder, Colorado:
19. Benbassat J, Cohen R. Clinical instruction and cognitive develop- AAAS/Westview Press; 1984:119-90.
ment of medical students. Lancet. 1982;1:95-7. 52. Davis R, Buchanan BG, Shortliffe EH. Production rules as a repre-
20. Keren G. On the importance of identifying the correct "problem sentation for a knowledge-based consultation program. Artificial In-
space." Cognition. 1984;16:121-8. telligence. 1977;8:15-45.
21. Kassirer JP, Kopelman RI. The critical role of context in the diag- 53. Newell A, Simon HA. Human Problem Solving. Engelwood Cliffs,
nostic process. Hosp Pract. 1987;22:67-98, 70-1, 74-6. New Jersey: Prentice Hall; 1972.
22. Pauker SG, Gorry GA, Kassirer JP, Schwartz WB. Towards the 54. Davis R, King J J. The origin of rule-based systems in Al. In: Bu-
simulation of clinical cognition: taking a present illness by computer. chanan BG, Shortliffe EH, eds. Rule-Based Expert Systems: The
Am J Med. 1976;60:981 -96. MYCIN Experiments of the Stanford Heuristic Programming Proj-
23. Miller GA. The magical number seven, plus or minus two. Psychol ect. Reading, Massachusetts: Addison-Wesley; 1984:20-52.
Rev. 1956;63:81-97. 55. Laird JE, Newell A, Rosenbloom PS. SOAR-an architecture for
24. Stillings NA, Feinstein MH, Garfield JL, et al. Topics in cognitive artificial intelligence. Artificial Intelligence. 1987;33:1-64.
psychology. In: Cognitive Science: An Introduction. Cambridge: 56. Anderson JR. The Architecture of Cognition. Cambridge: Harvard
MIT Press; 1987:73-86. University Press; 1983.
25. Kassirer JP, Kopelman RI. Memory: how we overcome its limita- 57. Weisberg LA, Strub RL, Garcia CA. Decision Making in Adult
tions. Hosp Pract. 1988;23:25-6, 28-9. Neurology. Philadelphia: B.C. Decker; 1987.
26. Langley P. Data-driven discovery of physical laws. Cognitive Sci- 58. Bready LL, Smith RB, eds. Decision Making in Anesthesiology.
ence. 1981;5:31-54. Philadelphia: B.C. Decker; 1987.
27. Pauker SG, Kassirer JP. Therapeutic decision making: a cost-bene- 59. Eiseman B, Wotkyns RS. Surgical Decision Making. Philadelphia:
fit analysis. N Engl J Med. 1975;293:229-34. W.B. Saunders; 1978.
28. Pauker SG, Kassirer JP. The threshold approach to clinical deci- 60. Feinstein AR. An analysis of diagnostic reasoning. III. The con-
sion making. N Engl J Med. 1980;302:1109-17. struction of clinical algorithms. Yale J Biol Med. 1974;47:5-32.
29. Eisenberg JM, Hershey JC. Derived thresholds: determining the 61. Greenfield S, Bragg FE, McCraith DL, Blackburn J. Upper-respira-

1 June 1989 • Annals of Internal Medicine • Volume 110 • Number 11 899

Downloaded From: http://annals.org/pdfaccess.ashx?url=/data/journals/aim/19681/ by a Scott Memorial Library User on 04/05/2017

 tory tract complaint protocol for physician-extenders. Arch Intern 67. Szolovits P, Pauker SG. Categorical and probabilistic reasoning in
Med. 1974;133:294-9. medical diagnosis. Artificial Intelligence. 1978;11:115-44.
62. Wood RW, Tompkins RK, Wolcott BW. An efficient strategy for 68. Szolovits P . Artificial intelligence and medicine. In: Szolovits P.
managing acute respiratory illness in adults. Ann Intern Med. Artificial Intelligence in Medicine. Boulder, Colorado: A A A S /
1980;93:757-63. Westview Press; 1984:1-19.
63. KomarorT AL, Black WL, Flatley M, Knopp RH, Reiffen B, Sher- 69. ShortlifTe EH. Computer-based Medical Consultations: MYCIN.
man H. Protocols for physician assistants: management of diabetes New York: Elsevier; 1976.
and hypertension. N Engl J Med. 1974;290:307-12.
70. Clancey WJ, Letsinger R. NEOMYCIN: Reconfiguring a rule-
64. Sox HC Jr, Sox CH, Tompkins RK. The training of physician's
based expert system for application to teaching. In: Clancey WJ,
assistants: the use of a clinical algorithm system for patient care,
audit of performance and education. N Engl J Med. 1973;288:818- Shortliffe EH, eds. Readings in Medical Artificial Intelligence: The
24. First Decade. Reading, Massachusetts: Addison-Wesley; 1984:361-
65. Komaroff AL, Pass TM, McCue JD, Cohen AB, Hendricks TM, 81.
Friedland G. Management strategies for urinary and vaginal infec- 71. Aikins JS, Kunz JC, Shortliffe EH, Fallat RJ. P U F F : An expert
tions. Arch Intern Med. 1978;138:1069-73. system for interpretation of pulmonary function data. In: Clancey
66. McDonald CJ. Protocol-based computer reminders, the quality of WJ, Shortliffe EH, eds. Readings in Medical Artificial Intelligence:
care and the non-perfectability of man. N Engl J Med. 1976; The First Decade. Reading, Massachusetts: Addison-Wesley;
295:1351-5. 1984:444-55.

900 1 June 1989 • Annals of Internal Medicine • Volume 110 • N u m b e r 11

Downloaded From: http://annals.org/pdfaccess.ashx?url=/data/journals/aim/19681/ by a Scott Memorial Library User on 04/05/2017