System and method of prediction through the use of latent semantic indexing

• US 10,224,119 B1
• Filed: 09/23/2014
• Issued: 03/05/2019
• Est. Priority Date: 11/25/2013
• Status: Active Grant

• Current Assignee: Quire, Inc.
• Original Assignee: Quire, Inc.
• Inventors: Heinrich, Kevin Erich, Homayouni, Ramin, Silver, Bradford Ryan
• Primary Examiner(s): Kanaan, Maroun P
• Application Number:
  US14/494,582
• Time in Patent Office: 1,624 Days
• CPC Class Codes: G06F 16/3334 Selection or weighting of terms from queries, including natural language queries
• CPC Class Codes: G06F 16/3344 using natural language analysis
• 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

A predictive modeling method implemented on a computer for predicting patient outcomes and conditions from medical database records of a population of patients, and an optimization process of iterative variation of parameters of the method to achieve a best precision fit. Individual patient documents are created by concatenation of unstructured text fields from the patient'"'"'s medical record, and these are processed using Natural Language Processing. A patient document corpus is built, and terms in the corpus are weighted and mapped to standard vocabularies. A term-by-document matrix is built and its dimensionality is reduced by Latent Semantic Indexing. Patient and term queries are combined and scored, producing a ranked list. The parameters of the model are iteratively optimized for an input list of patients with corresponding health score values.

Claims

1. An optimizing process for optimizing a predictive modeling method implemented on a computer for predicting patient outcomes and conditions from medical database records of a population of patients, said predictive modeling method comprising the steps of:

(a) providing medical database records of a population of patients, each patient of

said population of patients having corresponding outcome health score values;

(b) processing said medical database records by using Natural Language Processing;

(c) building a patient document corpus from said medical database records processed

by using Natural Language Processing;

(d) weighting terms in said corpus by assigning a weight to each term in the corpus,

said weight representing said term'"'"'s frequency for a patient'"'"'s document with

respect to said term'"'"'s frequency across all documents in said corpus;

(e) Constructing a high-dimensional and sparse term-by-document matrix from said

weighted terms, each said term in said corpus being represented as a vector across

said population of patients;

(f) performing matrix factorization of said term-by-document matrix using Latent

Semantic Indexing to reduce the dimensionality of said term-by-document matrix

into a lower-dimensional matrix concept space;

(g) querying said lower-dimensional matrix concept space using a term or

combination of terms to produce a single ranking of patients in said corpus using

a similarity score;

(h) given a single threshold of said similarity score, combining multiple said single

rankings to re-rank said population of patients in said corpus based on relatedness

to multiple queries;

(i) optimizing said predictive modeling method through iterative variation of certain

parameters to achieve a best precision fit, said certain parameters comprising;

(1) the number of patients used for each said query of said multiple queries;

(2) said threshold of said similarity score;

(3) a frequency of association to query of said patients of said corpus;

(4) a recall value of said patients returned by said query; and

(5) a precision value of said patients returned by said query; and

transmitting data associated with optimized re-ranked population of patients to one or more practitioners.

2. The optimizing process of claim 1, where said processing of said medical database records is performed by concatenation of unstructured text fields from said individual patient'"'"'s medical records, and where processing said medical database records by using Natural Language Processing is performed on individual patient documents.

3. The optimizing process of claim 1, where said lower-dimensional matrix concept space is queried using a given patient'"'"'s document in said corpus to rank other patient'"'"'s documents in said corpus to produce a ranking of said other patients in said corpus using said similarity score.