Learning cost-sensitive diagnostic policies from data

In its simplest form, the process of diagnosis is a decision-making process in which the diagnostician performs a sequence of tests culminating in a diagnostic decision. For example, a physician might perform a series of simple measurements (body temperature, weight, etc.) and laboratory measurements (white blood count, CT scan, MRI scan, etc.) in order to determine the disease of the patient. A diagnostic policy is a complete description of the decision-making actions of a diagnostician under all possible circumstances. This dissertation studies the problem of learning diagnostic policies from training examples. An optimal diagnostic policy is one that minimizes the expected total cost of diagnosing a patient, where the cost is composed of two components: (a) measurement costs (the costs of performing various diagnostic tests) and (b) misdiagnosis costs (the costs incurred when the patient is incorrectly diagnosed). The optimal policy must perform diagnostic tests until further measurements do not reduce the expected total cost of diagnosis. The dissertation investigates two families of algorithms for learning diagnostic policies: greedy methods and methods based on the AO* algorithm for systematic search. Previous work in supervised learning constructed greedy di agnostic policies that either ignored all costs or considered only measurement costs or only misdiagnosis costs. This research recognizes the practical importance of costs incurred by performing measurements and by making incorrect diagnoses and studies the tradeoff between them. This dissertation develops improved greedy methods. It also introduces a new family of learning algorithms based on systematic search. Systematic search has previously been regarded as computationally infeasible for learning diagnostic policies. However, this dissertation describes an admissible heuristic for AO* that enables it to prune large parts of the search space. In addition, the dissertation shows that policies with better performance on an independent test set are learned when the AO* method is regularized in order to reduce overfitting. Experimental studies on benchmark data sets show that in most cases the systematic search methods give better diagnostic policies than the greedy methods. Hence, these AO*-based methods are recommended for learning diagnostic policies that seek to minimize the expected total cost of diagnosis.