CLOSING GAPS IN REAL-WORLD EVIDENCE: UNLOCKING CLINICAL VALUE FROM UNSTRUCTURED EHR NOTES
Author(s)
Zuzana Dostalova, MSc, Marley Boyd, Jr., MS, Mike Temple, MD, Jeffrey Brown, PhD.
TriNetX, LLC, Cambridge, MA, USA.
TriNetX, LLC, Cambridge, MA, USA.
OBJECTIVES: Structured EHR fields systematically miss clinically critical data: 87% of smoking status records and virtually all biomarker results are absent. These gaps bias research and obscure treatment patterns in precision oncology. We applied validated AI extraction to recover this information at scale across 6 healthcare organizations.
METHODS: Clinical notes were processed through a six-step validated pipeline: note preparation, AI-based concept tagging, expert clinical review and correction, performance measurement, iterative model improvement, and full-dataset extraction. Target precision was ≥0.9; 87.04% of extracted facts met both NER and Assertion confidence thresholds. NSCLC patients were identified via AI-extracted histology (confidence ≥0.6). Biomarker status used an ever-positive approach across ALK, EGFR, and PD-L1 (≥1% TPS). Treatment concordance was assessed within 6 months of first biomarker test.
RESULTS: Of 7,741,982 patients with clinical notes, structured fields captured 22% of smoking status records; AI raised this to 67%. Among 28,256 NSCLC patients, 7,832 (27.7%) had AI-extracted biomarker data. ALK was tested in 32.9% (11.1% positive); EGFR in 37.3% (25.5% positive); PD-L1 in 93.2% (77.4% positive, 40% with high expression ≥50% TPS). Concordance within 6 months: ALK-positive patients received an ALK inhibitor in 23.2% versus 0.2% of ALK-negative patients; EGFR-positive 28.3% versus 0.6%. Notably, 20.9% of PD-L1-negative patients received a checkpoint inhibitor, compared with 35.9% of PD-L1-positive patients (40.5% high-expressors, 32.4% low-expressors), suggesting checkpoint inhibitor use independent of biomarker status, which is consistent with the current NCCN guidelines.
CONCLUSIONS: AI extraction uncovers clinically meaningful signals absent from structured EHR data — including treatment of biomarker-negative patients with targeted therapies and checkpoint inhibitor use irrespective of PD-L1 status. Combined with recovery of smoking status missingness, these findings demonstrate the research value of unstructured notes for precision oncology at scale. Ongoing development will extend this approach to breast cancer (ER, PR, HER2, Ki-67, BRCA1/2) and colon cancer (MSI, KRAS/NRAS) biomarkers.
METHODS: Clinical notes were processed through a six-step validated pipeline: note preparation, AI-based concept tagging, expert clinical review and correction, performance measurement, iterative model improvement, and full-dataset extraction. Target precision was ≥0.9; 87.04% of extracted facts met both NER and Assertion confidence thresholds. NSCLC patients were identified via AI-extracted histology (confidence ≥0.6). Biomarker status used an ever-positive approach across ALK, EGFR, and PD-L1 (≥1% TPS). Treatment concordance was assessed within 6 months of first biomarker test.
RESULTS: Of 7,741,982 patients with clinical notes, structured fields captured 22% of smoking status records; AI raised this to 67%. Among 28,256 NSCLC patients, 7,832 (27.7%) had AI-extracted biomarker data. ALK was tested in 32.9% (11.1% positive); EGFR in 37.3% (25.5% positive); PD-L1 in 93.2% (77.4% positive, 40% with high expression ≥50% TPS). Concordance within 6 months: ALK-positive patients received an ALK inhibitor in 23.2% versus 0.2% of ALK-negative patients; EGFR-positive 28.3% versus 0.6%. Notably, 20.9% of PD-L1-negative patients received a checkpoint inhibitor, compared with 35.9% of PD-L1-positive patients (40.5% high-expressors, 32.4% low-expressors), suggesting checkpoint inhibitor use independent of biomarker status, which is consistent with the current NCCN guidelines.
CONCLUSIONS: AI extraction uncovers clinically meaningful signals absent from structured EHR data — including treatment of biomarker-negative patients with targeted therapies and checkpoint inhibitor use irrespective of PD-L1 status. Combined with recovery of smoking status missingness, these findings demonstrate the research value of unstructured notes for precision oncology at scale. Ongoing development will extend this approach to breast cancer (ER, PR, HER2, Ki-67, BRCA1/2) and colon cancer (MSI, KRAS/NRAS) biomarkers.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
RWD176
Topic
Methodological & Statistical Research, Real World Data & Information Systems
Topic Subcategory
Distributed Data & Research Networks
Disease
Oncology, Respiratory-Related Disorders (Allergy, Asthma, Smoking, Other Respiratory)