Automated epilepsy and seizure type phenotyping with pre-trained language models

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Informatics

Background:

• Epilepsy is a common neurological disorder with diverse seizure and epilepsy types impacting prognosis and treatment.
• Electronic health records (EHRs) offer rich longitudinal data but lack detailed epilepsy phenotypes in structured formats.
• Unstructured clinical notes contain crucial phenotypic information often missed by traditional data analysis.

Purpose of the Study:

• To evaluate transformer-based language models for automated epilepsy and seizure type phenotyping from clinical notes.
• To benchmark model performance against expert epileptologist agreement.
• To deploy the best-performing model for large-scale phenotyping and analysis of epilepsy patient data.

Main Methods:

• Two transformer models, fine-tuned BERT and DeepSeek-R1, were assessed for phenotyping epilepsy and seizure types.
• A subset of clinical notes was annotated by epileptologists to establish a ground truth.
• The top-performing model was applied to a large cohort of clinical notes (77,049 notes from 18,566 patients).

Main Results:

• DeepSeek-R1 demonstrated performance comparable to expert agreement in classifying epilepsy and seizure types.
• DeepSeek-R1 outperformed BERT, especially on more granular classification tasks.
• Large-scale deployment revealed clinical patterns like diagnostic stabilization and seizure type co-occurrence.

Conclusions:

• Automated phenotyping of epilepsy using language models transforms unstructured EHR data into a valuable research resource.
• This approach enables large-scale, longitudinal analysis of epilepsy trajectories and treatment outcomes.
• The findings support future clinical and translational research in epilepsy care.