The Heart Failure Program (HFP) is a computer system which acts as an ``intellectual sounding board'', operating in the domain of cardiovascular disorders, assisting the physician by providing differential diagnoses for findings, predicting effects of therapy, and suggesting additional measurements, all with detailed graphical physiologic explanations. The mechanism for differential diagnosis consists of three parts, 1) an input interface that takes the findings about the patient in menu form, 2) a knowledge base in the form of a probability network of causal relationships between pathophysiologic states and findings, and 3) a reasoning mechanism designed to find likely explanations for the findings in terms of causal pathways through the pathophysiologic states. The result of differential diagnosis is an ordered list of complete explanations for the findings (called hypotheses) with relative probabilities.
The input interface is a dynamically expanding menu with entries divided
into categorical and numeric values, with arbitrary constraints among
categorical values and appropriate precision for numeric values along
scales. The intention is to capture the information pertinent to the
cardiovascular disease without requiring the system to do reasoning
outside of the domain and to display the relevant patient information in
an effective manner. It is assumed that the data has been interpreted
and filtered by the user. For many test results an interpretation is
entered rather than the test value, such as hypoxemia rather than a
pO value.
The HFP knowledge base (KB) is a clinically defined physiologic model of the cardiovascular system. From the perspective of diagnosis, the model consists of data structures representing physiologic states and measurement categories (the categories of patient information), constituting the general diagnostic knowledge about the domain and a template from which the more specific knowledge about a case is generated. Using the information from a case, the states and measurement categories are instantiated as nodes and findings representing the relationships that potentially exist in the case - essentially the superset of all possible diagnostic hypotheses for the patient. The states include diseases, qualitative states of physiologic parameters, and therapies. The measurement categories represent the observables entered in the input: the history items, symptoms, and lab results. The states are linked by probability relations, as are the relations between states and values in measurement categories. When a case is entered, the states and measurement values are instantiated as nodes and findings. The probability relations between them may be conditional on input values or on the nodes in a hypothesis. These probability relations are partially evaluated to provide the constraint implied by the input values.
There are two essential features of this knowledge representation that make the reasoning mechanisms tractable but also limit the expressive power of the KB. The first is the essentially binary nature of the physiologic states. For example, the low cardiac output node is either true or false. There are no degrees of severity. However, a parameter is not restricted to two states, so there is also a high cardiac output node with the constraint that high and low can not be simultaneously true. The probabilities between nodes can be adjusted for values in the input or even other nodes included in a hypothesis, overcoming some of this restriction, but essentially there is no representation of severity in the model. The second is lack of time relationships between nodes. For example, there is no way to represent and reason about a finding that was present yesterday but absent today. A hypothesis is a snapshot in time. This restriction is partially alleviated by having explicit nodes for some chronic states with different characteristics than their acute counterparts. For example, the KB has both high left atrial pressure and chronic high left atrial pressure represented. The only way to deal with effects that persist for some time after their causes cease is to make the cause node represent the average state of the cause over a longer time period and adjust the probabilities for findings that track the cause more closely. For example, the high left atrial pressure node represents an average value over hours so that it can act as a cause for pulmonary congestion, which takes hours to resolve. Otherwise, a single normal pressure reading would rule out a cardiac cause for pulmonary congestion. Without a time representation, the usual assumptions about the independence of diseases are also suspect. That is, the probability of disease combinations involving chronic diseases is actually much higher than simply multiplying the probabilities would indicate.
The KB covers the common and some not so common causes of heart failure or hemodynamic disturbance including myocardial ischemia and infarction, congestive, restrictive, and hypertrophic cardiomyopathy, the valvular diseases, atrial and ventricular septal defects, constrictive pericarditis and tamponade. It also has non-cardiac diseases that cause the same symptoms or complicate the hemodynamic situation such as pulmonary, renal, liver, or thyroid diseases, anemia and infection. The non-cardiac diseases are not further differentiated. That is, there is a node for primary liver disease but not for any specific types of liver disease.
The probabilities on the links between nodes are combined using a
``noisy-or'' combination rule[4] except for special links
called worsening factors, which increase the probability of another
cause but are insufficient to produce the effect alone, and correcting
factors, which decrease the probability. Thus, if the causes are ,
the worsening factors
, the correcting factors
, and at least one
of the causes in
is true, the probability of a node is
. Similarly, each finding has a
probability of being produced by nodes. The model is similar to those
investigated by Pearl[4] as Bayesian probability networks.
However, this model has forward loops (excluded by Pearl), some
probabilities that are conditional on other nodes in the hypothesis, and
nodes with multiple paths between them (handled only in exponential time
by Pearl's methods). Thus heuristic methods are necessary to reason
about the model.
Our solution to the differential diagnosis problem is to generate complete hypotheses (causal paths from primary causes) for the findings and present the user with a list of hypotheses and their relative total probabilities for comparison. In comparing hypotheses we discovered that the natural notion of different hypotheses requires that they differ in some significant node, nodes which we have labeled diagnostic. The algorithm is as follows: 1) check the input for definite implications, findings that require nodes to be true or false (known diagnoses, therapies, or pathognomonic findings); 2) collect the abnormal findings from the input; 3) find all of the diagnostic or primary nodes that could account for each finding; 4) rank the diagnostic and primary nodes by the number of findings they account for; 5) use the better of these as the initial nodes for generating small covering sets of primary nodes; 6) for each covering set, order the findings by the difference between the first and second highest probability path to it (since the best path to those findings are most likely to remain best as the hypothesis evolves); 7) for each finding, find the best path from the partial hypothesis and add it; and 8) prune the hypothesis of unneeded primary nodes and extra paths that decrease the probability. Finally, the probabilities of the hypotheses are computed by multiplying the probabilities of the nodes given the other nodes in the hypothesis and they are rank ordered and presented to the user. It is unnecessary to normalize by the probability of the findings as long as we are only interested in the rank order and relative probabilities of the hypotheses.
Figure 1 shows a simple, but otherwise typical, best hypothesis from
case 13 in the format HFP displays it on the screen.
This hypothesis can be summarized as having causal pathways that show:
The patient has chronic mitral regurgitation which is either primary with p=0.005 or secondary to chronic low LV systolic function with p=0.05. The chronic mitral regurgitation causes chronic high LA press with p=0.8, which in turn causes atrial stretch with p=0.5. The patient has paroxysmal atrial fibrillation which is either primary with p=0.008 or caused by atrial stretch with p=0.3. However, the digitalis decreases the probability of paroxysmal atrial fibrillation by p=0.3. Paroxysmal atrial fibrillation causes atrial fibrillation rhythm with p=0.4 but the digitalis decreases the probability by p=0.5.The hypothesis accounts for all of the abnormal findings and shows the causal paths that explain the hypothesized mechanisms.