Natural Language Processing in Diagnostic Texts from Nephropathology.

Area of Science:

• Nephropathology
• Computational Pathology
• Medical Informatics

Background:

• Accurate diagnosis in nephropathology is crucial for patient management.
• Image analysis is a key component of modern nephropathology.
• This study explores text-based analysis as an alternative or complement to image analysis.

Purpose of the Study:

• To investigate the efficacy of Natural Language Processing (NLP) methods in predicting final diagnoses from written nephropathological descriptions.
• To evaluate NLP as a surrogate for image analysis in nephropathology.
• To assess the performance of different NLP techniques for diagnostic prediction.

Main Methods:

• Utilized 1107 unlabelled nephropathological reports.
• Separated reports into microscopic description and diagnosis sections.
• Employed unsupervised clustering (HDBSCAN) and text classification (transformer encoder, SVM) for analysis.

Main Results:

• HDBSCAN with Bag-of-Words (BoW) achieved optimal clustering.
• Transformer encoder and Support Vector Machine (SVM) models performed best for diagnosis prediction.
• Achieved F1-scores up to 0.892 for certain diagnoses, with weaker performance for others.

Conclusions:

• Textual histomorphological descriptions alone can predict some nephropathological diagnoses.
• Predictive accuracy varies, with some diagnoses being more challenging to predict than others.
• Findings align with image analysis studies where complex pattern combinations affect diagnostic certainty.