The process of generating hypotheses is heuristic in HFP. This has led to three kinds of errors in the cases: not finding the best causal path, inappropriate handling of conditional probabilities, and not accounting for unlikely combinations of nodes. In cases 180 and 231, there were better causal paths to account for the findings but taking the findings in the heuristic order and searching for causal paths, they are missed. In cases 121, 148, and 209, there are conditions on the causal probabilities that depend on multiple nodes (as mentioned in section 5.2). In the generation of the hypotheses, these nodes do not assume truth values until after the wrong paths have been chosen. Both of these problems can be addressed with a mixed strategy of developing the hypotheses from both the diseases and the findings, rather than exclusively using the findings to search for causal pathways.
In cases 102 and 202 there are unlikely combinations of nodes. In case
102 there are COPD and hypocapnia (low pCO). Since COPD often
causes hypercapnia (high pCO
) and always tends to increase the
pCO
even when it stays in the normal range, this is unlikely.
Similarly, case 202 has both low blood volume and high cardiac output,
which are almost incompatible. Both of these problems require further
modifications to the KB.
One final error illustrates the considerations that go into a diagnosis. In case 96, HFP misses pulmonary embolism in the hypothesis and leaves pleuritic chest pain unexplained. Muscloskeletal or other non-cardiopulmonary causes for the chest pain may indeed have a higher probability than the desired hypothesis with pulmonary embolism, but it is so important that the program catch the possibility of the treatable disease that we left the pulmonary embolism as a mandatory part of the diagnosis. This kind of problem must be handled outside of the current mechanism for hypothesis generation using a scheme for assessing the utility of the hypothesis as well as its probability.