A computer-derived protocol to aid in selecting medical versus surgical treatment of horses with abdominal pain.

In order to determine which variables are useful in identifying horses with abdominal pain requiring surgery, data were analysed from 219 horses presented at one veterinary teaching hospital. Using multiple stepwise discriminant analysis with a recursive partitioning algorithm, we obtained a decision tree that identifies surgical and non-surgical patients. The prevalence of surgical patients was 79 per cent in this population. The sensitivity, specificity, and positive and negative predictive values of this decision tree were 99 per cent, 55 per cent, 90 per cent and 99 per cent respectively. Compared to the clinical decision, this decision tree yielded more false positives (11 per cent) but almost eliminated false negatives (1 per cent). This decision tree was validated by the jack-knife method and also by evaluation using a new sample in a second veterinary teaching hospital in which the prevalence of surgical patients was 55 per cent. This led to sensitivity, specificity and positive and negative predictive values of 93 per cent, 73 per cent, 81 per cent and 89 per cent respectively.