A Multi-Institutional Natural Language Processing Pipeline to Extract Performance Status From Electronic Health...

Area of Science:

• Medical Informatics
• Computational Linguistics
• Oncology

Background:

• Performance status (PS) is crucial for cancer patient assessment but often buried in clinical notes.
• Natural Language Processing (NLP) offers a solution for extracting this vital data from electronic medical records (EMRs).

Purpose of the Study:

• To design and validate a multi-institute NLP pipeline for automatic extraction of performance status (PS) from free-text patient notes.
• To assess the pipeline's accuracy and utility in real-world clinical data.

Main Methods:

• Developed an NLP pipeline using a rule-based approach with post-processing.
• Trained and validated the pipeline on data from Harris Health System (HHS) and the Veteran Affairs' Corporate Data Warehouse (VA-CDW).
• Tested PS documentation compliance with American Society of Clinical Oncology (ASCO) Quality Metric and analyzed disparities in stage IV non-small cell lung cancer (NSCLC) reporting.

Main Results:

• Achieved high accuracy: 92% on HHS data and 98.5% on VA data.
• Demonstrated 98.5% accuracy for stage IV NSCLC patients.
• Found a PS documentation rate exceeding 85%, surpassing ASCO Quality Metrics, with no observed disparities.

Conclusions:

• The developed NLP pipeline effectively extracts performance status from diverse healthcare institutions' free-text notes.
• This tool holds potential for use in longitudinal cancer data registries and improving clinical decision-making.