Extracting seizure frequency from epilepsy clinic notes: a machine reading approach to natural language processing.

Area of Science:

• Computational linguistics
• Clinical informatics
• Epilepsy research

Background:

• Seizure frequency and freedom are critical for epilepsy patient management.
• Extracting this data from unstructured clinical notes is challenging.

Purpose of the Study:

• To develop an automated method for extracting seizure-related information from clinical notes.
• To improve clinical decision-making and facilitate large-scale retrospective research in epilepsy.

Main Methods:

• A finetuning pipeline was created for pretrained neural models (BERT, RoBERTa, Bio_ClinicalBERT).
• Models were trained and tested on 1000 annotated clinical notes.
• Performance was evaluated for classifying seizure freedom and extracting seizure frequency and last seizure date.

Main Results:

• Finetuned models achieved near-human performance in classifying seizure freedom (BERTFT, Bio_ClinicalBERTFT >80% accuracy).
• Models demonstrated human-level performance in extracting seizure frequency and date (F1 scores >0.80).
• Significant performance gains were observed with approximately 70 annotated notes.

Conclusions:

• The novel machine reading approach performs at or near human levels for key clinical outcome extraction.
• This technology can support clinical practice and enable efficient retrospective clinical research.
• The finetuning pipeline requires minimal annotations for future clinical question answering.