BACKGROUND: Cost-effectiveness analyses of test/treatment strategies with sequential decisions typically require repeated analyses of sub-models that estimate utility in effectively identical contexts. We developed a technique for improving the efficiency of such analyses by storing intermediate results of a decision tree at the specific positions where they can be reused. We refer to these nodes as “Elephant Nodes”, since they both remember the utility for each context in which they appear, and require potentially large amounts of additional storage. When subtrees involve time-consuming utility calculations such as Markov processes, the tradeoff between space and time make this technique a viable option.

METHODS: We implemented Elephant Nodes in the Decision Maker software package (www.SimPal.org) so that the modeler needs to provide only a binding to a special variable, "REMEMBER", on a branch that may benefit from memory. The value of this variable indicates an index for the context in which the associated utility values should be stored. If an identical series of events reoccur via another path the utility does not need to be recomputed. This index value can be encoded into binding expressions at design-time through automatic means. Similarly, embedded decision nodes can be automatically transformed into Boolean nodes determined by a top-level strategy decision.

The strategies can be enumerated taking into account all possible combinations of embedded decisions in all event contexts. To enumerate the contexts for which subtree utilities are identical, we consider differences in upstream paths, combined with downstream differences due to top-level decision strategy. "REMEMBER" indices can thus be expressed as formulas composed of top-level strategies and a handful of Boolean variables which flag the events that differentiate the context.

RESULTS: To demonstrate this technique, we applied an automated transformation script to decision models for the management of solitary pulmonary nodules. One tree involved 11 decisions up to 3 deep, resulting in 42 strategies. Another involved 10 decisions up to 4 deep, resulting in 3534 strategies. The first tree having simple utilities required 85% fewer terminal node evaluations, only 66 out of 442. The second tree, with Markov processes as utilities yielded a time savings of 96%. In addition, the automated transformation results in less effort for the analyst, fewer errors, and a more comprehensive consideration of all possible strategies.