Comparing three natural language processing methods for the automatic identification of epilepsy patients from French...

Area of Science:

• Medical informatics
• Natural Language Processing
• Epilepsy research

Background:

• Manual review of clinical notes is the standard for identifying epilepsy patients but is time-consuming.
• Natural Language Processing (NLP) shows promise for automating patient phenotyping from unstructured text.

Purpose of the Study:

• To develop and validate NLP algorithms for identifying epilepsy patients using clinical notes.
• To compare the performance of keyword-based, rule-based, and pretrained language models for epilepsy detection.

Main Methods:

• A cohort of 109,448 patients was selected from a French clinical data warehouse.
• Sentences from clinical notes were labeled, and 3000 patients were manually reviewed by a neurologist.
• Three methods (basic keyword, rule-based, pretrained language model) were compared using F1 scores.

Main Results:

• The pretrained language model achieved the highest F1 score of .95 at both sentence and patient levels.
• The pretrained language model outperformed rule-based (.87 and .93) and basic (.81 and .82) methods.
• High performance was achieved by both rule-based and pretrained language models.

Conclusions:

• Developed algorithms can automatically identify epilepsy patients from unstructured clinical notes.
• These NLP tools support large-scale phenotyping and comorbidity detection in French data warehouses.
• Automated methods offer an efficient alternative to manual review for epilepsy identification.