Indexing of large scale patient set

• US 10,394,850 B2
• Filed: 06/25/2015
• Issued: 08/27/2019
• Est. Priority Date: 12/19/2012
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Wang, Fei, Wang, Jun
• Primary Examiner(s): Vital, Pierre M
• Assistant Examiner(s): Sultana, Nargis
• Application Number:
  US14/750,216
• Publication Number:
  US 20150293956A1
• Time in Patent Office: Days
• Field of Search: 707741
• CPC Class Codes: G06F 16/13 File access structures, e.g. distributed indices arrangements of input from, or output to, record carriers G06F3/06
• CPC Class Codes: G06F 16/2228 Indexing structures
• CPC Class Codes: G06F 16/2246 Trees, e.g. B+trees
• CPC Class Codes: G06F 16/278 Data partitioning, e.g. horizontal or vertical partitioning
• CPC Class Codes: G06F 16/285 Clustering or classification
• CPC Class Codes: G06F 16/31 Indexing; Data structures therefor; Storage structures
• CPC Class Codes: G06F 16/316 Indexing structures
• CPC Class Codes: G16H 10/60 for patient-specific data, e.g. for electronic patient records
• CPC Class Codes: G16H 50/70 for mining of medical data, e.g. analysing previous cases of other patients

Abstract

Systems and methods for indexing data include formulating an objective function to index a dataset, a portion of the dataset including supervision information. A data property component of the objective function is determined, which utilizes a property of the dataset to group data of the dataset. A supervised component of the objective function is determined, which utilizes the supervision information to group data of the dataset. The objective function is optimized using a processor based upon the data property component and the supervised component to partition a node into a plurality of child nodes.

Claims

1. A method for indexing data, comprising:

formulating an objective function to index a dataset, a portion of the dataset including supervision information;

determining a data property component of the objective function which utilizes a property of the dataset to group data of the dataset, the determining the data property component including maximizing a variance of the dataset by applying a principal component analysis (PCA) and maximum variance unfolding (MVU), the (MVU) comprising forming an inner product of a data matrix and a learned partition hyperplane and maximizing overall pairwise distances in a projected space to reduce computational costs during the indexing, and the PCA comprising determining an eigenvector from a data covariance matrix with a largest corresponding eigenvector;

determining a supervised component of the objective function which utilizes the supervision information to group data of the dataset; and

optimizing the objective function using a processor based upon the data property component and the supervised component to partition a node into a plurality of child nodes.

2. The method as recited in claim 1, wherein optimizing the objective function includes recursively optimizing the objective function at each node to provide a binary space partitioning tree.

3. The method as recited in claim 1, wherein determining the data property component includes clustering data around a median of the dataset.

4. The method as recited in claim 1, wherein determining the supervised component includes minimizing pairwise distance of similar data identified by the supervision information and maximizing pairwise distance of dissimilar data identified by the supervision information.

5. The method as recited in claim 1, wherein determining the supervised component includes applying the dataset as a linear prediction function.

6. The method as recited in claim 1, wherein the objective function includes a tradeoff parameter to balance contributions from the data property component and the supervision component.

7. The method as recited in claim 1, wherein the supervision information includes must-links indicating pairs of data that are similar to each other and cannot-links indicating pairs of data that are dissimilar to each other.

8. The method as recited in claim 1, wherein the dataset includes patient medical information.

9. The method as recited in claim 1, wherein the objective function includes a tradeoff parameter to balance contributions from the data property component and the supervision component.

10. The method as recited in claim 1, wherein determining the data property component includes applying maximum variance unfolding by forming a projection as an inner product of a data matrix and a learned partition hyperplane and maximizing overall pairwise distances in a projected space.

11. The method as recited in claim 1, wherein determining the data property component includes applying maximum variance unfolding such that a binary partition produces a balanced tree.

12. The method as recited in claim 11, wherein applying maximum variance unfolding comprises finding directions under which a projected data variance is maximized such that the binary partition of the directions increases a likelihood of producing the balanced tree.