ABSTRACT
BACKGROUND/PURPOSE: The authors compared 3 quantitative methods for assisting clinicians in the differential diagnosis of abdominal pain in children, where the most common important endpoint is whether the patient has appendicitis. Pretest probability in different age and sex groups were determined to perform Bayesian analysis, binary logistic regression was used to determine which variables were statistically significantly likely to contribute to a diagnosis, and recursive partitioning was used to build decision trees with quantitative endpoints. METHODS: The records of all children (1,208) seen at a large urban emergency department (ED) with a chief complaint of abdominal pain were immediately reviewed retrospectively (24 to 72 hours after the encounter). Attempts were made to contact all the patients' families to determine an accurate final diagnosis. A total of 1,008 (83%) families were contacted. Data were analyzed by calculation of the posttest probability, recursive partitioning, and binary logistic regression. RESULTS: In all groups the most common diagnosis was abdominal pain (ICD-9 Code 789). After this, however, the order of the most common final diagnoses for abdominal pain varied significantly. The entire group had a pretest probability of appendicitis of 0.06. This varied with age and sex from 0.02 in boys 2 to 5 years old to 0.16 in boys older than 12 years. In boys age 5 to 12, recursive partitioning and binary logistic regression agreed on guarding and anorexia as important variables. Guarding and tenderness were important in girls age 5 to 12. In boys age greater than 12, both agreed on guarding and anorexia. Using sensitivities and specificities from the literature, computed tomography improved the posttest probability for the group from.06 to.33; ultrasound improved it from.06 to.48; and barium enema improved it from.06 to.58. CONCLUSIONS: Knowing the pretest probabilities in a specific population allows the physician to evaluate the likely diagnoses first. Other quantitative methods can help judge how much importance a certain criterion should have in the decision making and how much a particular test is likely to influence the probability of a correct diagnosis. It now should be possible to make these sophisticated quantitative methods readily available to clinicians via the computer.
