974-P: Glucose Data Extraction from Continuous Glucose Monitoring Documents Using Natural Language Processing

Introduction and Objective: Continuous glucose monitoring (CGM) is growingly used in diabetes clinical and research settings; however, manually extracting the valuable data (e.g., time above, in, or below range) from CGM reports is time-consuming. Natural language processing (NLP) is an automatic approach to extracting data from unstructured sources (e.g., images), but its application in CGM remains unexplored. We aimed to evaluate the accuracy of extracting CGM data using NLP. This work is crucial to automatically obtaining valuable data from CGM reports. Methods: We analyzed 23,319 CGM reports from the EHR at NYU Langone Health 2012-2022. The steps of our algorithm pipeline consist of: 1) perform optical character recognition (OCR) to obtain textual data; 2) determine the type of document based on keywords in OCR results; 3) extract variables of interest from the text according to the specific types of documents; 4) save the extracted information into a CSV file. Two experts in using CGM for research and clinical practice conducted an independent manual review of 1% of the documents. We calculated accuracy (correct extraction of CGM data ) by comparing the algorithm against manual review. Results: Among the 23,319 documents analyzed, 36.8% were Freestyle Libre and 63.2% Dexcom. In our preliminary experiments, we randomly selected 10 reports and tested PaddleOCR and easyOCR for information recognition. We finally chose PaddleOCR as it demonstrated a text block recognition accuracy of 99.64% (546/548), which significantly outperformed easyOCR that achieved only 94.89% (520/548). For information extraction, the agreement in evaluating Libre results between two experts was 99.93%. When comparing algorithm accuracy with manual review, the accuracy for Libre was 99.87%, and for Dexcom was 100.00%. Conclusion: Using an NLP approach to extract value glucose data from CGM reports is feasible and accurate, which is useful for clinical practice and diabetes research. Disclosure Y. Zheng: None. Y. Song: None. E. Iturrate: None. B. Wu: None. S. Zweig: None. S.B. Johnson: None. Funding New York Regional Center for Diabetes Translation Research (NY-CDTR) Pilot and Feasibility (P&F) Program Funding (P30DK111022-08 )