NATURAL LANGUAGE PROCESSING (NLP) TO IDENTIFY TREATMENT PATTERNS IN ADVANCED UROTHELIAL CARCINOMA (AUC): ADVANCING REAL-WORLD EVIDENCE GENERATION THROUGH AUTOMATED CLINICAL DATA EXTRACTION
Author(s)
Diego Teyssonneau, MD1, Veronique Lorgis, MD2, Laurent Paret, PhD3, Lorraine HOUVET, MSc4, Arthur Perez, PharmD4, ISABELLE DELAROZIERE, PhD4, Jason Hoffman, PharmD5, Mairead M. Kearney, MBA, MPH, MSc, MD6, Benoit Van Hille, PhD3, Titouan Lacombe, BSc4, Camille Chaussy, BSc4, Laura Luciani, PharmD, MSc3, Stéphane Culine, MD, PhD7.
1Institut Bergonié, Bordeaux, France, 2Burgundy Cancer Institute, Dijon, France, 3Merck Santé S.A.S., an affiliate of Merck KGaA, Lyon, France, 4OSPI, Lyon, France, 5EMD Serono Research & Development Institute, Inc., an affiliate of Merck KGaA, Billerica, MA, USA, 6Merck Healthcare KGaA, Darmstadt, Germany, 7Saint Louis Hospital, Paris, France.
1Institut Bergonié, Bordeaux, France, 2Burgundy Cancer Institute, Dijon, France, 3Merck Santé S.A.S., an affiliate of Merck KGaA, Lyon, France, 4OSPI, Lyon, France, 5EMD Serono Research & Development Institute, Inc., an affiliate of Merck KGaA, Billerica, MA, USA, 6Merck Healthcare KGaA, Darmstadt, Germany, 7Saint Louis Hospital, Paris, France.
OBJECTIVES: Advances in NLP and generative artificial intelligence (AI) offer new opportunities accelerating real-world evidence (RWE) generation and support patient-centered decision-making. We report a scientific framework for semi-automated extraction of complex clinical variables, from the SEQUOIA_RW study in patients treated with avelumab first-line maintenance (1LM) for aUC until 2025 in France.
METHODS: Pseudonymized unstructured data available in electronic medical records (EMR), such as medical notes, scanned lab reports, and general practitioner letters, were processed through an algorithm configuration based on study protocol and transformed into a suitable format for data management. Quality checks were completed prior to algorithm deployment on all documentation. Manual annotations allowed calculation of variable-level performance metrics including F1-score (harmonic mean of precision and recall) and completeness rate. Results presented coupled F1-Score;completeness. For sites unable to support automated extraction, manual electronic case report form (eCRF) based data capture was implemented. Both data sources were integrated into a harmonized multicenter database for study analyses.
RESULTS: AI-assisted extraction process was applied to 269 patient files from 7 study sites representing 42,877 documents. Almost 200k data points were extracted taking a mean time of 5.5 minutes per patient. Avelumab 1LM initiation date (index date) was assessed as (100%; 100%). Other important variables such as chemotherapy administration dates, history of peri-operative therapy, radiotherapy and surgery, cancer stage, Human epidermal growth factor receptor 2 and fibroblast growth factor receptor status showed coupled scores of (85-96%;100%). Binary variables distributions were similar between AI-assisted extraction and eCRF-based data collection.
CONCLUSIONS: This study confirmed AI-assisted extraction has the potential to transform unstructured clinical data into reliable RWE. By leveraging recent advances in NLP and agentic tools, this scalable framework reduced manual data abstraction burden and enhanced efficiency of converting routine EMR-based clinical information into research-ready datasets. Ultimately, this cutting-edge methodology may support patient-centered clinical innovation.
METHODS: Pseudonymized unstructured data available in electronic medical records (EMR), such as medical notes, scanned lab reports, and general practitioner letters, were processed through an algorithm configuration based on study protocol and transformed into a suitable format for data management. Quality checks were completed prior to algorithm deployment on all documentation. Manual annotations allowed calculation of variable-level performance metrics including F1-score (harmonic mean of precision and recall) and completeness rate. Results presented coupled F1-Score;completeness. For sites unable to support automated extraction, manual electronic case report form (eCRF) based data capture was implemented. Both data sources were integrated into a harmonized multicenter database for study analyses.
RESULTS: AI-assisted extraction process was applied to 269 patient files from 7 study sites representing 42,877 documents. Almost 200k data points were extracted taking a mean time of 5.5 minutes per patient. Avelumab 1LM initiation date (index date) was assessed as (100%; 100%). Other important variables such as chemotherapy administration dates, history of peri-operative therapy, radiotherapy and surgery, cancer stage, Human epidermal growth factor receptor 2 and fibroblast growth factor receptor status showed coupled scores of (85-96%;100%). Binary variables distributions were similar between AI-assisted extraction and eCRF-based data collection.
CONCLUSIONS: This study confirmed AI-assisted extraction has the potential to transform unstructured clinical data into reliable RWE. By leveraging recent advances in NLP and agentic tools, this scalable framework reduced manual data abstraction burden and enhanced efficiency of converting routine EMR-based clinical information into research-ready datasets. Ultimately, this cutting-edge methodology may support patient-centered clinical innovation.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
MSR86
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
No Additional Disease & Conditions/Specialized Treatment Areas, Oncology