Application of information theory to decision analysis in potentially prostaglandin-responsive neonates

To define settings in which use of prostaglandin E₁ before transfer from a community hospital to a tertiary care center benefits neonates with possible heart disease, information theory was used to predict the probability of a favorable response to prostaglandin therapy from the limited information of clinical variables. Records of 250 patients, newborn to 7 days old, with suspected heart disease were reviewed to assess six clinical variables (cyanosis, respiratory distress, heart murmur, pulse contour, hepatomegaly and prematurity). According to the anatomic and hemodynamic cardiovascular condition, each case was categorized as to whether a favorable response to prostaglandin E₁ could be anticipated. Information content of each clinical variable with respect to prostaglandin responsiveness was determined, and patients were classified according to the most informative clinical variable. Stepwise extraction of information proceeded until remaining clinical variables added no significant information. Bayes' rule gave estimates of probability of prostaglandin-responsive defect in final subgroups for use in decision analysis. Cyanosis, murmur, small volume pulses and prematurity gave information about prostaglandin-responsive defects. Decision analysis indicated that frequency of poor outcome is minimized by early prostaglandin treatment of cyanotic term infants with a murmur or poor pulses, regardless of how ill they appear, and by treating any critically ill term newborn who has either cyanosis or poor pulses. Acyanotic patients with normal pulses are best untreated with prostaglandin until after definitive diagnosis is made. Advantage to either course was not seen in some small subgroups. Information theory with decision analysis is a rigorous approach to identify relevant clinical variables and define their roles in critical decisions in pediatric cardiology.