A multivariable prognostic model for equine colic patients

A survey of 1965 equine colic cases was conducted from August 1985 to July 1986 ten equine referral hospitals located througout the U.S.A. Two-thirds of the cases were randomly selected for model development (1336), while the remaining cases (629) were used only for subsequent validation of the model. The following outcomes were defined: (1) died or killed prior to discharge from the hospital; or (2) alive at the time of discharge. Only variables which were significant (P<0.05) in an initial bivariate screening procedure and for which there were less than 400 missing values were considered in the multivariable modelling. A multivariable logistic regression model was constructed using a stepwise algorithm. The model used 666 cases and included the following variables: peripheral pulse (normal or weak), pulse rate, surgical or medical treatment, packed cell volume, self-inflicted trauma (absent or present) and capillary refill time. For each horse in the validation data set which had the appropriate variables recorded (n=335), the estimated expected probability of death (expected value) was calculated using the logistic regression equation. Using Bayes theorem, the post-test probability was calculated from the expected value (an estimate of the test odds) and the present probability (the case-fatality rate at each institution). Nomograms of predictive values for different case-fatality rates and expected values were constructed. Hosmer-Lemeshow goodness-of-fit statistics indicated that the model fitted the model data set well but the validation set poorly. However, when the observed case-fatality rates were compared with the average post-test probabilities for 0.10 increments of post-test probability, qualitatively, the model's performance was better.