Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

A Back-Off Method for Identifying Relevant Content for ICD-10-CM Codes

IP.com Disclosure Number: IPCOM000245414D
Publication Date: 2016-Mar-08
Document File: 3 page(s) / 176K

Publishing Venue

The IP.com Prior Art Database

Abstract

The effectiveness of computer-assisted coding (CAC) workflows is dependent on accurate identification of relevant content for automatically identified ICD-10 codes. Some auto-coding methods generate relevant content as part of code identification while other methods leverage circumstantial textual cues that would not be readily recognized as relevant by medical codes. It is therefore important to identify relevant content for codes identified by these latter methods retrospectively. This paper describes end-to-end methods for identifying relevant content for ICD-10-CM codes.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 31% of the total text.

Page 01 of 3

A Back-Off Method for Identifying Relevant Content for ICD-10-CM Codes

Abstract

The effectiveness of computer-assisted coding (CAC) workflows is dependent on accurate identification of relevant content for automatically identified ICD-10 codes. Some auto-coding methods generate relevant content as part of code identification while other methods leverage circumstantial textual cues that would not be readily recognized as relevant by medical codes. It is therefore important to identify relevant content for codes identified by these latter methods retrospectively. This paper describes end-to- end methods for identifying relevant content for ICD-10-CM codes.

Introduction

When using computer assisted coding systems for ICD-10 coding, a medical coder can choose a "code-centric" view, where a list of codes suggested by auto-coders is displayed, or a "document-centric" view, which displays the medical documentation to be coded. In the document-centric view, the system indicates relevant content for auto-suggested codes (i.e., portions of the text mentioning information relevant to some auto-suggested code). When working in the document-centric view, the coder must select a specific piece of relevant content in order to view auto-suggested codes. In the code-centric view, a coder selects a particular code to browse relevant content for that code. Frequently, auto-suggested codes without associated relevant content are not displayed to the coders in either view. Thus, identification of relevant content for auto-suggested codes is critical for effectiveness of computer-assisted workflows.

Some auto-coding methods implemented in natural language processing (NLP) engines generate relevant content as an integral part of their algorithmic design. This holds especially for rule-based methods. However, this is not the case for many other methods, and in particular those using machine learning techniques. Since machine learning models are able to leverage a large number of varied textual cues to predict a code, including highly circumstantial cues, the pieces of information used to make a prediction are often not recognized by humans as relevant content for that code. Displaying these pieces of information as relevant content is liable to create confusion and potentially decrease customer satisfaction and trust. At the same time, identification of appropriate relevant content for auto-suggested codes may be treated as a separate task from predicting the codes themselves and may be performed retrospectively after the codes for a given document have been selected. Retrospective identification of relevant content for ICD-10 diagnostic codes is described in further detail.

Implementation

Relevant Content from Automatically Generated Code-To-Concept Crosswalk

This method is based on a newly created knowledge resource between an automatically generated "crosswalk" from ICD-10 CM codes and medical concept identifiers (IDs). The IDs used in this resource...