Classification of cervical biopsy free-text diagnoses through linear-classifier based natural language processing.

Area of Science:

• Computational pathology
• Medical informatics
• Machine learning in healthcare

Background:

• Cervical cancer screening relies on validated algorithms, necessitating accurate correlation of cytology, HPV tests, and biopsy diagnoses.
• Manual categorization of free-text cervical biopsy reports is time-consuming, prone to errors, and can introduce bias.
• Advances in machine learning and natural language processing (NLP) offer potential solutions for automating pathology report classification.

Purpose of the Study:

• To develop and validate a machine learning classifier using NLP to categorize free-text cervical biopsy diagnoses.
• To assess the accuracy and robustness of the NLP classifier compared to manual annotations and on external datasets.
• To demonstrate the utility of NLP in streamlining pathology classification tasks for cervical cancer screening.

Main Methods:

• An efficient NLP framework, FastText™, was applied to an annotated cervical biopsy dataset.
• A supervised classifier was trained to assign accurate categories to free-text biopsy interpretations.
• The classifier's performance was evaluated against manual annotations and on an independent external dataset.

Main Results:

• The machine learning classifier achieved high concordance (>99.6%) with manually annotated data.
• Discrepant cases were reviewed by expert pathologists, providing insights into classifier performance.
• The classifier demonstrated robustness on an untrained external dataset, achieving 97.7% concordance.

Conclusions:

• NLP-based machine learning provides an efficient and accurate method for classifying cervical biopsy diagnoses.
• This approach can significantly reduce the labor and potential bias associated with manual categorization.
• The study highlights the practical benefits and limitations of applying NLP to real-world pathology classification.