The Heart Failure Program (HFP) is a computer program designed to assist
physicians in reasoning about patients with cardiovascular
disease[3][2][1]. It takes as input the description
of the patient's history, physical examination findings, and test
results at the level of detail one would find in the physician's
description in the patient record. This information is used by the
program to generate a differential diagnosis for the case. It attempts
to generate a hypothesis corresponding to each likely combination of
diseases that could account for the findings. The program uses a
pseudo-Bayesian network as a knowledge base representing the physiologic
causality of the cardiovascular hemodynamics from which it generates
hypotheses that explain the findings. A small portion of the network is
sketched in figure 1.
Each arrow between nodes represents a causal link with a conditional probability. Each hypothesis is a subnetwork of the knowledge base instantiated with the findings of the case. This subnetwork is a complete explanation from primary causes through the pathophysiologic mechanisms to the findings. The corresponding portion of a hypothesis is shown on the left side of figure 2. The only findings not in the hypothesis causal network are those considered normal or better explained by something outside of the domain. For each hypothesis generated, the program can compute a probability by combining the probabilities in the network. The differential diagnosis consists of these hypotheses ordered by the computed probability. To limit the differential to likely hypotheses, the list is cut off when the probabilities fall below 1%of the best hypothesis. Thus, the differential may consist of one or many hypotheses. This type of diagnostic hypothesis is much more informative than those of earlier programs, such as AI/RHEUM[4], QMR[5], or DXplain[6], which provide hypotheses consisting of a single word or phrase. The structure of the HFP hypothesis explains the findings by showing the causal mechanisms producing them and thus provides a justification that the physician can evaluate to decide whether the conclusions are reasonable. The causal mechanisms are modeled at a level of detail consistent with that at which a physician might explain the findings to a colleague. This enhances the usefulness of the hypotheses for justification because the concepts fit naturally with the understanding of the physician. It is not a handicap for reasoning because this is presumably the level at which the human physician does similar reasoning and the physician is the best model we have for this kind of reasoning.
We previously conducted a formative evaluation of the program[7]. This evaluation used 242 cases collected from discharge summaries and compared the program diagnoses to diagnoses collected from cardiologists using the same information. We examined the maximum potential accuracy of the program by iteratively revising the knowledge base and rerunning the cases. Ultimately, the program was able correctly diagnose 90%of the cases. The main reasons that the remaining 10%could not be diagnosed correctly were the lack of reasoning about temporal relationships, the distinctions of different severities of diseases, and inappropriate combining of probabilities of multiple diseases. As a result of this evaluation we have developed methods for reasoning with the additional constraints provided by temporal and severity relationships to deal with some of the limitations encountered. We also developed an appropriate framework for computing the probability of multiple diseases[8].
The additions to the program include both enhancements of the knowledge base and changes in the reasoning methods. In the knowledge base the diseases and pathophysiologic states (all referred to as nodes) are subdivided into levels of severity and subtypes with additional constraints on the causal links. The severity and subtypes distinguish qualitatively different hemodynamic consequences. The link constraints reflect what severities and time bounds are required for the cause to produce the effect. For example, aortic regurgitation has three levels of severity reflecting situations in which there is a murmur without any hemodynamic consequences, situations in which the heart has compensated by dilating, and those in which the aortic regurgitation has caused deterioration of systolic function. These differences are reflected in the different probabilities and constraints on the links conditioned on the severity. Aortic regurgitation is also divided into two subtypes: primary valvular regurgitation and secondary regurgitation due to dilation of the aortic root. This distinction allows the effects to be different or have different probabilities for different subtypes. General temporal constraints on the nodes reflect limitations on causality, such as how long it takes for the nodes to become true (eg, minutes for an acute myocardial infarction (MI) to years for aortic stenosis), how long the node will remain true after the cause ends, and how long a patient might have the state (eg, an ``acute'' MI only exists for two days by definition before it becomes a ``recent'' infarct). The causal links include statements relating the severities and temporal constraints. For example:
Thus, high pulmonary vascular resistance (PVR) has two severities,
distinguished by being reversible and irreversible, rather than a
particular measured resistance. A number of possible criteria could be
used for defining severity, such as right ventricular dilatation, but
reversibility was chosen because in the case of PVR it is correlated
with severity and has important implications for management. The
possible causes include pulmonary embolism, which if severe ()
causes high PVR immediately and is irreversible 30%of the time. Low
arterial oxygen levels for a limited time (a week or less) can cause
reversible high PVR 30%of the time. Finally, chronic (a year or more)
severe (
) high left atrial pressure (as mitral stenosis might cause)
can cause either severity of high PVR. The severities of pulmonary
embolism and high left atrial pressure are also defined in terms of
qualitative differences in their effects.
These constraints are enforced by the reasoning mechanisms at two levels: 1) The causal pathways that are computed when the knowledge base is loaded are pruned using the constraints, and 2) the patient specific nodes that are generated from the findings to build the hypotheses carry the constraints which guide the process of building the hypotheses. In the completed hypothesis, each node has one or more data structures called severities that represent a constraint on disease severity and subtype true over a specified temporal interval. For example, a severity for high PVR includes the following:
Each of these structures is linked to the corresponding severity structures of the causes and effects. More than one severity structure is necessary when the evidence for a node supports different constraints on severity over different time intervals, such as a chronic disease with acute worsening.
This work of incorporating temporal relations and severity constraints had progressed to the point where it was appropriate to evaluate the performance of the program.