Abstract | Automatic data processing to identify EGFR mutations in pathology reports of patients with non-small cell lung cancer (NSCLC)
Authors: Betzabel Cajiao Garcia (Groningen, Netherlands) Bart Koopman (Groningen, Netherlands) Vincent De Jager (Groningen, Netherlands) Clara L. Oeste (Leuven, Belgium) Ed Schuuring (Groningen, Netherlands) Anthonie Jan Van der Wekken (Groningen, Netherlands) Stefan Willems (Groningen, Netherlands) Leon Van Kempen (Edegem, Belgium)
Presented at ESMO MAP 2023.
ℹ️ Click HERE to visit our Lung Cancer Insights overview for more information on the LynxCare datasets within our European hospital network.
Background
Non-small cell lung cancer (NSCLC) is the most common subtype of lung cancer. Driver mutations in epidermal growth factor receptor (EGFR), which occur in ∼10-15% of NSCLC, can be targeted by specific therapies. Real-world data can provide valuable information regarding the prevalence of these mutations, including their subtypes. However, despite comprehensive data availability in the Dutch Pathology Registry (Palga), manual extraction of EGFR mutation status from narrative pathology reports is time-consuming. Therefore, we used machine learning and natural language processing (NLP) to identify pathology reports that state the presence of an EGFR mutation.
Methods
The NLP algorithm was trained and validated on manually curated datasets of semi-structured pathology reports from the Palga archive to generate a structured OMOP CDM database. Afterwards, pathology reports of patients with metastatic, non-squamous NSCLC in 2019-2020 were requested from the Palga registry. The output of the algorithm was compared to results of the manual extraction.
Results
The algorithm identified 839 (10.9%) reports that mention an EGFR alteration. Manual analysis indicated 875 reports, resulting in a data extraction accuracy of 95.9% (95% CI 92.7-99.2). The 36/875 (4.1%) reports that were not identified by the algorithm were all listed as variants of unknown significance (VUS) by the reader. In the EGFR-mutated patient groups, 73.0% (639/875) had a common EGFR mutation (i.e., exon 19 deletion (41.4%, 362/875) or p.(Leu858Arg) mutation (31.7%; 277/875)). Exon 20 insertions were detected in 8.1% (71/875) of patients. Automatic data processing was 48 times faster than complete manual extraction.
Conclusions
NLP algorithms allow rapid data extraction from pathology reports, thereby offering a time-efficient and cost-effective alternative to manual data processing. In turn, this approach enables rapid insight in current biomarker testing rates and prevalence of (actionable) mutations.
Background
Non-small cell lung cancer (NSCLC) is the most common subtype of lung cancer. Driver mutations in epidermal growth factor receptor (EGFR), which occur in ∼10-15% of NSCLC, can be targeted by specific therapies. Real-world data can provide valuable information regarding the prevalence of these mutations, including their subtypes. However, despite comprehensive data availability in the Dutch Pathology Registry (Palga), manual extraction of EGFR mutation status from narrative pathology reports is time-consuming. Therefore, we used machine learning and natural language processing (NLP) to identify pathology reports that state the presence of an EGFR mutation.
Methods
The NLP algorithm was trained and validated on manually curated datasets of semi-structured pathology reports from the Palga archive to generate a structured OMOP CDM database. Afterwards, pathology reports of patients with metastatic, non-squamous NSCLC in 2019-2020 were requested from the Palga registry. The output of the algorithm was compared to results of the manual extraction.
Results
The algorithm identified 839 (10.9%) reports that mention an EGFR alteration. Manual analysis indicated 875 reports, resulting in a data extraction accuracy of 95.9% (95% CI 92.7-99.2). The 36/875 (4.1%) reports that were not identified by the algorithm were all listed as variants of unknown significance (VUS) by the reader. In the EGFR-mutated patient groups, 73.0% (639/875) had a common EGFR mutation (i.e., exon 19 deletion (41.4%, 362/875) or p.(Leu858Arg) mutation (31.7%; 277/875)). Exon 20 insertions were detected in 8.1% (71/875) of patients. Automatic data processing was 48 times faster than complete manual extraction.
Conclusions
NLP algorithms allow rapid data extraction from pathology reports, thereby offering a time-efficient and cost-effective alternative to manual data processing. In turn, this approach enables rapid insight in current biomarker testing rates and prevalence of (actionable) mutations.
