To represent these causal patterns, we need a number of additional properties on node and link descriptions in the knowledge base. Three of these are time ranges: onset, delay, and persistence. Persistence is the range of time the state might remain true after the cause ceases. Thus, progressive states can be represented as having infinite persistence. Accumulative states have a persist property that gives the range of time needed for the state to resolve. For example, pedal edema can remain for days to weeks after the cause has been removed, dilatation of the heart takes weeks to many months to return to normal. Immediate causes have a persistence of zero.
Onset and delay refer to the initiation of a state. Accumulative states require time to start. This is captured by the onset property. The value is either zero for immediate states, or the range of time it takes for the state to become apparent. One might argue that even though it takes a day before one could observe pedal edema, the swelling was actually there earlier. However, the concern is on the relationship between states and observed findings, so requiring a time for onset makes sense even in such a case. Because the onset of the state is often a time in which the effects of the state may start to develop as well, the onset period is treated as part of the time of the state for reasoning about effects. The delay is the onset plus any additional time between the beginning of the cause and the beginning of the effect. Thus, if an effect has both a delay property and an onset property, the time between the cause and the effect is the delay. When the effect is observed it is considered to have been in existence for the onset time.
The most common type of delay is reflected in event-like causation. The delay of events is zero to infinity relative to their cause. Some of these also have an onset. For example, anemia can be caused by renal insufficiency. The delay is arbitrary but the onset is sufficiently gradual that it should be reasoned with as if it had been there for a week when it is observed. Other than events, it is unusual for a causation to have a delay in addition to an onset that makes any difference for diagnostic reasoning. The clearest instance is constrictive pericarditis. Pericardial calcification takes place over many months and does not have any effect until it starts to restrict the heart's filling capacity. At that point there are many effects that are observable.
The remaining properties are the max-exist (maximum existence time) and two binary properties, intermittent and self-limiting. The maximum existence is the maximum length of time a patient would stay in that state, even though the cause continued. The self-limiting property says that the state will return to normal even if no correcting state is present. For example, high sympathetic states only last a short time and return to normal (normal with respect to clinically important manifestations) without any therapy being directed at the state. The max-exist property is also needed for states that are not self-limiting when the continuation of the state is not compatible with life. For example, there are no effects that would be caused by months of septic shock because no patient would survive that long. Either the shock is successfully treated or the patient dies within a few days. The property intermittent implies that the state or finding does not always have to be observable. For example, many arrhythmias are intermittent and not observing them during a particular examination does not rule them out or mean that they can not have any effects.
In summary, the representation of causality in the knowledge base requires the following properties:
The relationships among these times are diagrammed in figure 3.
The rules for applying these properties are as follows:
These rules follow from the discussion above except for the third rule. This additional rule encompasses two practical considerations. First, causes have manifestations before (or contemporaneous with) their effects. From the representation it would be possible for a cause to have a longer onset than its effect making the effect observable before the cause. However, what it means for a state to be observable is that it has manifestations, so such a situation would not make sense. As a practical consideration, if a cause develops slowly, the effect will also develop slowly. Computationally, this is a useful constraint because it provides tighter constraints on what causes can produce a particular effect. Secondly, the cause and effect must overlap. In other words there are no remote causes. Computationally, this is also a useful constraint because it means that the effect must start before the cause ends regardless of how long a delay could otherwise exist. Physiologically, it is a matter of perspective. In situations where there appear to be remote causes, there is some underlying mechanism that covers the time period, although it may have fewer observable findings. For example, when an MI causes pericarditis a week later, there is an underlying process of myocardial modification going on that may continue for several weeks.
Besides causal links to states, there are also correcting links to many of the states. These are primarily the therapies that can counter the states, The model used in this program for the effects of correcting influences is the same as that for causes. That is, a therapy may take a period of time (onset) before it produces the desired effect. We have not come across sufficient reason to use persistence or other aspects of the representation to model the effects of the correcting influences. This also is a matter of perspective. One could model the effect of a surgical procedure, such as a valve replacement, as a therapy taking a short time with an infinite persistence, but it is just as easy to handle it as a therapy that always remains true. Most drugs have some persistence (the pharmacokinetic half-life), but that can be considered part of the treatment time.
With this information it is possible to compute the time limits either for a cause when the effect is known or an effect when the cause is known. When a node is instantiated for a case, it is given a temporal interval, representing the observable time of the node. These temporal intervals have earliest and latest beginning and ending times, similar to the representation used in the CHECK system[4]. These intervals are used in a number of ways. The findings attributed to the node and the causes contributing to it are used to refine the limits of the interval as a hypothesis is built. To determine the causal pathways to be added to the hypothesis, the intervals are used to determine consistency. To speed this process, the overall temporal constraints on causal pathways are pre-computed in the model. This alone eliminates as inconsistent about 20%of the pathways that were computed in the purely probabilistic model. In the context of a specific node or finding, the probability along a pathway from a known node or primary node is recomputed using the more specific temporal information from the case. Thus, all of the nodes added to a hypothesis are consistent with the temporal constraints of causation. This time mechanism combines the ideas about reasoning with time outlined in our earlier paper[8] with the probabilistic reasoning described more recently[9].