FROM BROAD LUNG CANCER ICD-10 CODES TO CONFIRMED NON-SMALL CELL LUNG CANCER (NSCLC): HUMAN-IN-THE-LOOP LARGE LANGUAGE MODEL (LLM) EXTRACTED CLINICAL NOTES ENABLE HISTOLOGY AND STAGE STRATIFICATION NOT ACHIEVABLE FROM STRUCTURED DATA ALONE USING...
Author(s)
Isabella Even-Chen, BA1, ilan behm, MPH2, Rahul Das, PhD3, Atharva Manjrekar, MS4, Woojun Daniel Park, PhD5, Allison Perry, PhD1.
1Norstella, New York, NY, USA, 2Norstella, Englewood, CO, USA, 3Norstella, Yardley, PA, USA, 4Norstella, West Hartford, CT, USA, 5Norstella, Houston, TX, USA.
1Norstella, New York, NY, USA, 2Norstella, Englewood, CO, USA, 3Norstella, Yardley, PA, USA, 4Norstella, West Hartford, CT, USA, 5Norstella, Houston, TX, USA.
OBJECTIVES: To characterize the extent to which human-in-the-loop large language model (LLM) extracted clinical notes enable true NSCLC diagnosis confirmation and histology subtype identification beyond what structured claims and EHR data alone can provide, and to assess implications for real-world cohort construction.
METHODS: A cross-sectional analysis was conducted using NorstellaLinQ’s US real-world linked structured EHR, and clinical notes (June 2023-June 2026). Among 126,727 prevalent patients with at least one lung cancer ICD-10 code (C34%) in EHR, LLM-extracted clinical notes were applied to confirm NSCLC diagnosis and tumor histology. Confirmed NSCLC patients were further stratified by disease stage (early vs late) and biomarker status using note-derived data. Structured EHR fields were assessed for their ability to provide diagnosis confirmation, histology subtype, and stage independent of clinical notes.
RESULTS: Of 126,727 EHR patients with a lung cancer ICD-10 code, LLM-extracted clinical notes confirmed NSCLC in 44,540 (35%), demonstrating that the majority of coded lung cancer patients do not have confirmed NSCLC documentable from structured fields alone. Among confirmed NSCLC patients, approximately 96% had disease stage available exclusively from clinical notes: 7,650 (18%) were early stage and 35,197 (82%) were late stage. The most frequently documented biomarkers were PD-L1 (19,633 patients), EGFR (9,854), ALK (7,686), and KRAS (5,904).
CONCLUSIONS: Structured EHR fields alone cannot confirm NSCLC diagnosis, identify histology subtype, or determine disease stage in real-world populations. LLM-extracted clinical notes enable these critical stratifications, revealing that the majority of coded lung cancer patients lack confirmed NSCLC and that late-stage disease predominates among confirmed cases. These findings have direct methodological implications for RWE study design in NSCLC and for the validity of evidence submitted to HTA reviews of stage- and histology-specific therapies.
METHODS: A cross-sectional analysis was conducted using NorstellaLinQ’s US real-world linked structured EHR, and clinical notes (June 2023-June 2026). Among 126,727 prevalent patients with at least one lung cancer ICD-10 code (C34%) in EHR, LLM-extracted clinical notes were applied to confirm NSCLC diagnosis and tumor histology. Confirmed NSCLC patients were further stratified by disease stage (early vs late) and biomarker status using note-derived data. Structured EHR fields were assessed for their ability to provide diagnosis confirmation, histology subtype, and stage independent of clinical notes.
RESULTS: Of 126,727 EHR patients with a lung cancer ICD-10 code, LLM-extracted clinical notes confirmed NSCLC in 44,540 (35%), demonstrating that the majority of coded lung cancer patients do not have confirmed NSCLC documentable from structured fields alone. Among confirmed NSCLC patients, approximately 96% had disease stage available exclusively from clinical notes: 7,650 (18%) were early stage and 35,197 (82%) were late stage. The most frequently documented biomarkers were PD-L1 (19,633 patients), EGFR (9,854), ALK (7,686), and KRAS (5,904).
CONCLUSIONS: Structured EHR fields alone cannot confirm NSCLC diagnosis, identify histology subtype, or determine disease stage in real-world populations. LLM-extracted clinical notes enable these critical stratifications, revealing that the majority of coded lung cancer patients lack confirmed NSCLC and that late-stage disease predominates among confirmed cases. These findings have direct methodological implications for RWE study design in NSCLC and for the validity of evidence submitted to HTA reviews of stage- and histology-specific therapies.
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
RWD95
Topic
Clinical Outcomes, Epidemiology & Public Health, Real World Data & Information Systems
Topic Subcategory
Health & Insurance Records Systems
Disease
Oncology