Combining Data and Knowledge by MaxEnt-Optimization of Probability Distributions

We present a project for probabilistic reasoning based on the concept of maximum entropy and the induction of probabilistic knowledge from data. The basic knowledge source is a database of 15000 patient records which we use to compute probabilistic rules. These rules are combined with explicit probabilistic rules from medical experts which cover cases not represented in the database. Based on this set of rules the inference engine PIT (Probability Induction Tool), which uses the well-known principle of Maximum Entropy [5], provides a unique probability model while keeping the necessary additional assumptions as minimal and clear as possible. PIT is used in the medical diagnosis project Lexmed [4] for the identification of acute appendicitis. Based on the probability distribution computed by PIT, the expert system proposes treatments with minimal average cost. First clinical performance results are very encouraging.