Optimizing, diagnostic, and therapeutic strategies using decision-theoretic planning: principles and applications.

OBJECTIVE Decision-theoretic planning is a new technique for selecting optimal actions. The authors sought to determine whether decision-theoretic planning could be applied to medical decision making to identify optimal strategies for diagnosis and therapy. METHODS An existing model of acute deep venous thrombosis (DVT) of the lower extremities--in which 24 management strategies were compared--was converted into a set of conditional-probabilistic actions for use by the DRIPS decision-theoretic planning system. Actions were grouped into an abstraction/decomposition hierarchy. A utility function was defined in accordance with the existing DVT management model to incorporate the costs and risks of the diagnostic tests and treatments. RESULTS From 18 primitive actions (such as "perform venography" and "treat if venography shows thigh DVT"), a total of 312 possible concrete plans were encoded within the abstraction/decomposition hierarchy. The DRIPS planning system used abstraction techniques to eliminate 136 possible plans (44%) from consideration. It determined that, given the parameters specified, the most cost-effective management strategy was "no tests, no treatment." This result differed from the published result of "perform ultrasonography, treat if positive." In reviewing the original article, it was determined that DRIPS had revealed an error in the manually constructed decision trees used in that manuscript. At values of $75,000 and greater for the cost of death, the optimal strategy became "impedance plethysmography (IPG), don't wait, perform venography if IPG is positive, and treat only if venography shows thigh DVT." CONCLUSION Decision-theoretic planning is applicable to medical decision making and may be an extremely useful technique for complex decisions. The use of inheritance abstraction makes the technique computationally tractable for complex planning problems, and the modular nature of the data entry may help eliminate errors that appear in manually encoded decision trees.