System and method for integration of medical information

1. A computer implemented method for integrating medical information using the Unified Medical Language System (UMLS) as a hierarchical tree-structured taxonomy of medical terms, comprising the steps of:

providing a source ontology and a target ontology, said ontologies comprising formalizations of domain concepts associated with a medical terminological vocabulary, each concept being associated with data, each said ontology being distinct from the UMLS;

learning concept classifiers for said source and target ontologies from training data;

forming associations that map elements between the source and target ontology using the UMLS and said concept classifiers, wherein associated source ontology elements and target ontology elements are mapped to a same UMLS concept, wherein forming associations comprises the steps of:

providing an ontology;

initializing a test concept with a root concept of said ontology;

initializing a current UMLS concept with a root concept of UMLS;

classifying said test concept using the data associated with said test concept against all UMLS concepts at the hierarchy level of said current concept;

selecting a best UMLS concept from said hierarchy level that has a best classification of said test concept;

determining whether the data associated with said test concept can be better classified with any child concept of said best UMLS concept, wherein

if said test concept cannot be better classified with any child concept of said best UMLS concept, associating said test concept to said best UMLS concept, and

if said test concept can be better classified with a child concept of said best UMLS concept, reinitializing the current UMLS concept with said child UMLS concept that better classifies the data associated with said test concept, and repeating said steps of classifying said test concept, selecting a best UMLS concept, and determining whether the test data can be better classified with a child concept until either said current UMLS concept is a leaf UMLS concept or said test concept cannot be better classified to any children of the current UMLS concept;

ranking said associations in terms of a number of source and target concepts to which each UMLS concept is mapped; and

presenting a predefined number of highest ranked associations to a user,

wherein the steps of providing a source ontology and a target ontology, learning concept classifiers, forming associations, ranking said associations, and presenting associations to a user, are performed by a computer processor.

2. The method of claim 1 , wherein learning classifiers comprises the steps of:

providing an ontology;

starting from a root concept of said ontology,

collecting training data for said concept of said ontology by aggregating said training data from all descendants in the ontology tree;

using multiple instance learning to learn a Naive Bayes classifier for said concept;

improving the classifier by using shrinkage along a path from the root of the ontology to said concept; and

associating the learned classifier to the concept, wherein said steps of collecting training data, using multiple instance learning, improving the classifier, and associating the learned classifier to the concept are repeated for each concept in the ontology.

3. The method of claim 2 , wherein said ontology is UMLS.

4. The method of claim 2 , wherein said ontology is said target ontology.

5. The method of claim 2 , wherein said ontology is said source ontology.

6. The method of claim 1 , further comprising reinitializing said test concept to another concept in said ontology, and repeating said steps of initializing a current UMLS concept, classifying said test concept, selecting a best UMLS concept, and determining whether the test data can be better classified with a child concept until said test concept can be associated to a best UMLS concept.

7. The method of claim 1 , wherein said ontology is the target ontology.

8. The method of claim 1 , wherein said ontology is the source ontology.

9. The method of claim 1 , wherein ranking said associations comprises the steps of determining, for each concept in UMLS, a product of a number of source ontologies and a number of target ontologies to which said UMLS concept is mapped by an association; and sorting said UMLS concepts in terms of decreasing product values, wherein said predefined number UMLS concepts with highest product values are selected.

10. The method of claim 1 , further comprising the step of retraining those classifiers presented to said user that are selected by said user for correction.