Clinical Record AI Coding With Highlight Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The variability in format and content of clinical records makes it difficult to automate or support the manual process of coding, as they often include images, handwriting, and special symbols, leading to inefficiencies and errors in assigning codes for healthcare claims.
Innovation Solution
A multi-stage Artificial Intelligence (AI) engine is employed to identify content highlights within clinical records, associate them with candidate codes, and correct errors through a neural network trained at different granularities, using optical character recognition (OCR) to process text and perform binary classification to suggest accurate codes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual coding review is performed to ensure accuracy, then coding precision is improved, but productivity deteriorates due to time-consuming manual processes
Solution Approach 1:
The coding process is segmented into multiple stages: initial code suggestion generation, confidence scoring, selective human review of only low-confidence cases, and final validation. This segmentation allows automated processing of high-confidence cases while maintaining human oversight for uncertain cases, thereby improving both productivity and precision simultaneously.
Solution Approach 2:
The system implements feedback loops where coded data is continuously validated against clinical records, and coding accuracy is measured and fed back to improve the automated coding algorithm. Human coders' corrections are also fed back to refine the system, creating a continuous improvement cycle that enhances precision without proportionally increasing manual effort.
2Productivity
If automated coding tools are implemented to improve productivity, then coding productivity is improved, but reliability deteriorates due to errors in automated code assignment
Solution Approach 1:
Instead of fully automating the coding process, the system applies partial automation with confidence-based filtering. The automated tool generates suggestions for all records, but only low-confidence suggestions require human review. This partial automation approach maintains high productivity while ensuring reliability through selective human validation of uncertain cases.
Solution Approach 2:
The system introduces an intermediary confidence scoring mechanism that acts as a mediator between automated coding and human review. This intermediary layer identifies which automated suggestions are reliable enough to accept automatically and which require human verification, thereby maintaining both productivity and reliability.
3Measurement precision
If comprehensive review of all clinical records is performed to ensure accurate coding, then coding precision is improved, but loss of time increases due to lengthy review processes
Solution Approach 1:
The system performs partial review by using confidence scoring to identify only those cases that require human verification. High-confidence automated suggestions are accepted without human review, while only low-confidence cases undergo manual verification. This selective approach maintains coding precision while dramatically reducing the time loss associated with reviewing all records comprehensively.
4Measurement precision
If multiple coding options are generated to improve precision, then device complexity increases due to multi-stage processing, but ease of operation deteriorates due to complex system management
Solution Approach 1:
The multi-stage coding system operates autonomously without requiring complex user management. The confidence scoring and selective review processes happen automatically, and the system self-manages the workflow between automated processing and human verification. This self-service approach maintains high coding precision while preserving ease of operation by eliminating the need for users to manage system complexity.
Data Source
AI summary
A method includes receiving a record containing clinical information associated with a patient; and automatically processing the record using a multi-stage Artificial Intelligence (AI) engine to generate at least one suggested code for one or more portions of the clinical information; wherein the multi-stage AI engine comprises: a first module configured to identify content highlights within the one or more portions of the clinical information and associate the content highlights with candidate codes, respectively; and a second module configured to remove ones of the content highlights and/or ones of the candidate codes that have been identified and/or associated in error to generate corrected content highlights that are associated with the corrected candidate codes, respectively.