EUVIMEDEUVIMED entdecken
Crossref

Abstract 9599: Vision-Enabled GPT5.4 Mini for Automated Extraction of Echocardiographic Measurements from Imaging Data

Akshay Arora, Khemra Keo, Jennifer Priem, Jasjit Kaur Banwait, Joseph Sahulcik, jason ettlinger, Lisa Maria Cross, Elisa Priest, Javed Butler

Circulation · 02.11.2026 · 154 (Suppl_1)

DOI: 10.1161/circ.154.suppl_1.9599 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Circulation, erschienen am 02.11.2026. Verfasst von Akshay Arora und 8 weiteren Autor:innen.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Introduction

Echocardiographic (echo) measurements are critical for cardiovascular assessment, yet many remain embedded within images or unstructured reports. Manual extraction is labor-intensive and prone to error. While traditional OCR methods often struggle to differentiate measurement type and imaging context (e.g., mitral vs. tricuspid valve gradients or apical four-chamber vs. Doppler studies), recent advances in vision-enabled GPT-5.4 models enable accurate extraction of clinically relevant echocardiographic measurements directly from images. Research Question: Can a GPT-guided OCR model accurately recover quantitative echo measurements from imaging data?

Methods

We developed a vision-enabled GPT-5.4-mini-OCR framework [Figure] to automatically extract quantitative echo measurements directly from echo DICOM images using the pixel data while filtering non-clinical content, including annotations and device-generated metadata. The echo images were de-identified to remove protected health information before analysis. In addition to measurement extraction, the model identified relevant imaging contexts, including echo view (e.g., apical four-chamber), measurement focus (e.g., mitral valve), and functionality (e.g., 2D, Pulsed-Wave Doppler). Model performance was evaluated by comparing OCR-extracted measurements, including measurement names, values, and units, to corresponding EHR records.

Results

A total of 53,353 echo sequence images from 599 patients (mean age = 78.1 ± 10.6 years, BMI = 30.5 ± 6.9; 49.8% female) were analyzed using the OCR pipeline and compared with structured EHR-derived echo measurements. Across major echo categories, OCR achieved 87.4% overall agreement with EHR data. Agreement was highest for Right Ventricle measurements (91.2%), followed by Tricuspid/Pulmonic Valve (89.8%), Aortic Valve (88.3%), Great Vessels (87.4%), Mitral Valve (80.7%), Left Ventricle (78.8%), and Atria (76.4%). Manual review of discordant cases revealed that many discrepancies were attributable to differences in measurement naming conventions between OCR-generated labels and EHR abbreviations rather than true extraction errors, suggesting that actual OCR performance may be underestimated. On average, 13.2% of OCR-extracted measurements were absent from structured EHR data (range 2.3% - 29.5%), demonstrating OCR's ability to enhance echo data completeness.

Conclusions

Vision-enabled GPT-based OCR accurately extracted and contextualized echo measurements directly from imaging data, helping address gaps in structured EHR capture. Although some discordance resulted from differences in OCR and EHR measurement naming conventions, measurement adjudication improved data quality and reliability. This approach enables scalable recovery of clinically relevant echo data for research and clinical applications.

Zitieren

APA

Arora, A., Keo, K., Priem, J., Banwait, J. K., Sahulcik, J., ettlinger, J., Cross, L. M., Priest, E., & Butler, J. (2026). Abstract 9599: Vision-Enabled GPT5.4 Mini for Automated Extraction of Echocardiographic Measurements from Imaging Data. Circulation, 154(Suppl_1). https://doi.org/10.1161/circ.154.suppl_1.9599

Vancouver

Arora A, Keo K, Priem J, Banwait JK, Sahulcik J, ettlinger J, et al. Abstract 9599: Vision-Enabled GPT5.4 Mini for Automated Extraction of Echocardiographic Measurements from Imaging Data. Circulation. 2026;154(Suppl_1). doi:10.1161/circ.154.suppl_1.9599

BibTeX
@article{Arora2026,
  author = {Arora, Akshay and Keo, Khemra and Priem, Jennifer and Banwait, Jasjit Kaur and Sahulcik, Joseph and ettlinger, jason and Cross, Lisa Maria and Priest, Elisa and Butler, Javed},
  title = {Abstract 9599: Vision-Enabled GPT5.4 Mini for Automated Extraction of Echocardiographic Measurements from Imaging Data},
  journal = {Circulation},
  year = {2026},
  volume = {154},
  number = {Suppl_1},
  doi = {10.1161/circ.154.suppl_1.9599}
}
Kostenlos starten →