Automated information extraction from echocardiography reports for the development of a congenital heart disease database

Abstract

Background: Congenital Heart Disease (CHD) is common in Thailand. Manual patient registry is time and labor intensive.

Objective: The aim of this study was to evaluate an automated system for a CHD registry from echocardiography reports.

Methods: An automated data extraction workflow was developed using N8N as the primary integration platform. The system processed image-based echocardiography reports, beginning with Optical Character Recognition (OCR) to convert images to text. Subsequently, an agentic AI, powered by the ChatGPT-4o model, extracted key clinical information, transforming unstructured text into structured JSON data according to a specifically designed instruction set. This structured data was then automatically populated into a Google Sheet, which worked as the patient registry database. Accuracy was evaluated by comparing the system output against data manually verified by a pediatric cardiologist. Efficiency was assessed by comparing the consuming time of automated process with manual data entry method.

Results: A total of 301 echocardiography reports were processed. The accuracy of the automated system to identify the main diagnosis and additional diagnoses were 86.7% and 88.0% respectively. The automated system significantly reduced the average processing time from 84.0 seconds with manual entry to 12.3 seconds per report (a 71.7-second reduction, p value <0.001). The system had a process failure rate of 0.997% and a minor extraction issue rate of 9.9%.

Conclusion: The new developed system proves to be an accurate, efficient, and reliable method for CHD patient registry from existing clinical reports.