PREDICTING SURGICAL OUTCOME USING BAYESIAN-ANALYSIS

Surgical outcome analysis is best performed using Bayesian statistics. The ability of this type of analysis to take into consideration multi ple parameters affecting surgical outcome is a marked improvement over single-condition probabilities that ignore the many degrees of freedo m in the dynamics of a surgical intervention. To illustrate the power of a Bayesian analysis a surgical population of 1017 patients undergoi ng cholecystectomy, colon resection, and appendectomy was developed. E ach patient was assigned to a mutually exclusive outcome group (D-1, s urvival; D-2, survival with complications; D-3, nonsurvival), A priori outcome probabilities for the population were D-1 = 0.917, D-2 = 0.06 6; and D-3 = 0.017. A conditional probability matrix (CPM) was then de veloped for 59 patient parameters (Sj) that may have affected outcome. The CPM contained the conditional probability that a parameter was pr esent given the known outcome P(Sj/Di), Once the CPM was matured Bayes ian analysis allowed one to predict the surgical outcome given any set or combination of patient parameters P(Di/Sj). Posterior probabilitie s generated by the Bayes analysis allowed one to investigate the effec t of a single parameter or any group of parameters on outcome. Criteri on based validity testing based on comparison of Bayesian outcomes ver sus the surgeons perception of outcomes for computer simulated surgery on 15 artificial patients suggests that this type of analysis provide s insightful and educational data to the operating surgeons (V-mortali ty = 0.547, SEE = 24.46; V-morbidity = 0.319, SEE = 25.86), Objective outcome analysis or surgical peer review cannot be fairly accomplished unless the statistical methodology takes into consideration all of th e parameters affecting outcome. This study concludes that Bayes Theore m provides the ideal statistical framework for performing an outcome a nalysis that considers the many parameters affecting the results of a surgical intervention. (C) 1998 Academic Press.