Development of a Natural Language Processing Model for Extracting Kidney Biopsy Pathology Diagnoses.

Area of Science:

• Nephrology
• Medical Informatics
• Computational Linguistics

Background:

• Kidney biopsy reports are unstructured text, requiring manual abstraction for diagnosis.
• Natural Language Processing (NLP) has not been extensively validated for extracting kidney biopsy diagnoses.

Purpose of the Study:

• To develop and evaluate an accurate NLP model for extracting kidney biopsy diagnoses from free-text reports.
• To improve the efficiency and scalability of kidney biopsy data analysis.

Main Methods:

• Utilized PubMed Bidirectional Encoder Representations from Transformers (BERT) for NLP model development.
• Processed 3,042 kidney biopsy reports from 2,666 patients using Structured Query Language (SQL) and Python.
• Split reports into training (80%) and testing (20%) sets for model development and validation.

Main Results:

• The NLP model achieved an average Area Under the Receiver Operating Curve (AUROC) of 0.95 across all diagnoses.
• For the 20 most common diagnoses, the model achieved an AUROC of 0.97 and an F1 score of 0.72.
• High interrater reliability (Cohen kappa = 0.76) was observed for manual abstraction.

Conclusions:

• An accurate and scalable NLP system for extracting kidney biopsy diagnoses from free-text reports has been demonstrated.
• This system can significantly aid epidemiologic studies and patient identification for clinical trials.
• The NLP model offers a valuable tool for advancing kidney disease research.