The findings that have proven the most difficult to interpret are murmurs. Murmurs present a particular challenge for the probability network formalism because there can be multiple murmurs and they have varying location and extent. For example, mitral regurgitation can have a systolic murmur that is usually holosystolic but can be shorter in duration. It is usually loudest at the apex, but sometimes in the third, fourth, or fifth left interspace. It is often also heard in the left axilla, or possibly in a number of other locations. It can be of any intensity, while most other murmurs have restrictions on their intensity. Mitral regurgitation can also have a diastolic rumble murmur in association with the systolic murmur when severe. Since the descriptions for the murmurs from different valvular lesions overlap considerably, we use a scheme for adjusting the probability that a murmur is caused by a particular lesion. The probability is multiplied by a factor for each of the characteristics reflecting how typical the value is for that disease. Since there may be multiple values for other locations, only the most typical is used.
This scheme has proven fairly effective for setting the probabilities to reflect how likely a disease is to cause a particular murmur, but there are still several problems. First, it is very difficult to fill in the table of allowed murmur characteristics. It is easy to describe the typical murmurs of particular lesions. It is hard to imagine all of the values that would still be consistent with a particular lesion. In the process of reviewing the cases, we made seven changes to the table of murmur characteristics. Second, the location and extent of murmurs means that a single murmur may be mistaken for two and two murmurs may be mistaken for one. The program independently considers each murmur and will often attribute multiple murmurs to the same lesion, more often than it should, since the intent on the part of the physician is that separately described murmurs (of the same type) do not seem to be coming from the same lesion. The problem of finding two causes for a singly described murmur is harder. Usually the evidence that this is happening is multiple alternate locations, some of which are rare or impossible for any single lesion. Separating the murmurs is complicated because there is no way to tell where the primary location of the second murmur might be. In cases 100 and 153 two murmurs are probably described as one, since the multiple locations specified are not consistent with any single murmur. Case 139 has a murmur description that is just inconsistent. In that case, the cardiologists discount the location and make the proper attribution using other findings. The program ignored the murmur entirely but still had the correct diagnosis as the second hypothesis on the final run and the first hypothesis on several other runs.