Machine learning-based natural language processing to extract PD-L1 expression levels from clinical notes.

Area of Science:

• Oncology
• Biomedical Informatics
• Natural Language Processing

Background:

• Programmed death-ligand 1 (PD-L1) expression is critical for predicting patient response to cancer immunotherapy.
• PD-L1 status is often buried in unstructured clinical notes, hindering its use in large-scale studies.
• Automating PD-L1 extraction is essential for advancing cancer immunotherapy research.

Purpose of the Study:

• To develop and evaluate a machine learning-based natural language processing (NLP) tool.
• To extract PD-L1 expression values from unstructured clinical notes within the Veterans Affairs electronic health record system.
• To enable population-level analysis of PD-L1 status for cancer immunotherapy research.

Main Methods:

• Development of a machine learning model for NLP.
• Training and validation using data from the Veterans Affairs electronic health record system.
• Evaluation of the model's performance in extracting PD-L1 expression values.

Main Results:

• The NLP tool demonstrated high performance across various granularity levels.
• Mean precision for PD-L1 positive was 0.859, recall 0.994, and F1 score 0.921.
• For numeric PD-L1 values, the mean absolute error was 0.537 (scale 0-100).

Conclusions:

• An accurate NLP method for deriving PD-L1 status from clinical notes was developed.
• This tool significantly reduces the manual effort required for medical record review.
• The method will facilitate future population-level studies in cancer immunotherapy.