ABSTRACT NUMBER: ESOC2026A2214 COMBINING A LARGE LANGUAGE MODEL WITH TRADITIONAL SOFTWARE ENGINEERING TOOLS FOR AUTOMATED RES-Q REGISTRY VARIABLE EXTRACTION

Abstract Background and aims Despite substantial advances in digital health data infrastructure, quality improvement efforts still rely on resource-intensive manual data collection. Using a combination of a large language model (LLM) and conventional programming tools, we developed and evaluated the accuracy of a pilot algorithm that enables automated data extraction of the international Registry of Stroke Care Quality (RES-Q) variables. Methods We created a reference dataset containing 198 variables for 100 ischemic stroke cases, manually extracted from the electronic health records (overall 13,872 data points), taking an average of 20 minutes per patient. In automated extraction, 126 variables were extracted using conventional software engineering, and 72 using LLM. The accuracy of categorical variables was compared using Cohen’s κ, numeric variables using Spearman’s correlation, and free-text variables were assessed descriptively. Results The accuracy of the current version of the pilot algorithm was 89.2%, compared to the ground-truth dataset. Discrepancies included 457 (3.3%) missing and 1,042 (7.5%) incorrect values. 60 variables, including age, sex, prestroke modified Rankin scale score, index thrombectomy, mTICI score, puncture and reperfusion timestamps, showed Cohen’s κ score of 1.0 or Spearman’s coefficient of 1.0 (Figure 1). Overall, estimated programming time was 416 hours, and 19 algorithm iterations were performed using 5-30 patients’ data in each iteration. 1.2% of values were found to be incorrectly entered manually into the reference dataset. Conclusions Automated tools can potentially improve stroke registry data collection efficiency by reducing data extraction time, with accuracy for some variables comparable to or better than trained staff. Further improvements in the algorithm are anticipated. Conflict of interest “All authors: nothing to disclose” Figure 1 - belongs to Results