Authors :
Presenting Author: Rohan Bhansali, BS – Beth Israel Deaconess Medical Center
M. Brandon Westover, MD, PhD – Beth Israel Deaconess Medical Center
Daniel Goldenholz, MD, Phd – Beth Israel Deaconess Medical Center
Rationale:
Traditional methods for detecting anti-seizure medication side effects rely on structured reporting systems that may underestimate the frequency of adverse effects or miss them altogether. This study developed and validated an artificial intelligence-based approach to systematically identify temporal associations between levetiracetam initiation and adverse effects using unstructured clinical notes.Methods:
We employed a two-phase methodology using open-weight large language models. First, we generated synthetic clinical notes using gpt-oss-120b with ground truth clinical data and stylistic parameters to create diverse note compositions mimicking real clinical documentation patterns. We then employed DeepSeek-R1:70b to analyze these notes and identify side effects temporally correlated with initiation of a masked medication that we called ‘Seizurel.’ After validation demonstrated accurate approximation of ground truth side effect frequencies, we applied this methodology to de-identified clinical notes from 600 Mass General Brigham patients initiating levetiracetam therapy. The medication name was masked as ‘Seizurel’ again to prevent model bias from pre-trained knowledge about levetiracetam’s known side effect profile.Results:
Analysis of 575 patients revealed the following side effect frequencies: altered mental state, confusion, and delirium including cognitive impairment (12%); sedation, somnolence, and drowsiness (6%); agitation, irritability, and aggression (6%), aphasia, word-finding difficulties, and speech problems (3%); fatigue and lethargy (3%); weakness (3%); headache (2%); dizziness and vertigo (2%); nausea and vomiting (2%); and thrombocytopenia (2%).Conclusions:
Our LLM-derived frequencies demonstrated reasonable congruence with established literature values. Published levetiracetam frequencies from Lexicomp include drowsiness (8-15%), irritability (6-12%), aggressive behavior (1-10%), confusion (2-3%), fatigue (10-11%), dizziness (5-9%), nausea (5%),and headache (14-19%). Our methodology successfully identified several side effects within comparable frequency ranges, particularly for neuropsychiatric symptoms. The higher frequency of altered mental status and cognitive impairment detected through AI analysis compared to traditional reporting may reflect the methodology’s ability to capture more subtle cognitive changes documented in clinical notes but not formally reported as adverse effects. Speech and language difficulties, identified in 3% of patients, represent an important finding that may be underrecognized in classical pharmacovigilance systems.
Large language model-based analysis of clinical notes provides a promising approach for comprehensive pharmacovigilance in epilepsy care. This methodology successfully identified known levetiracetam side effects while revealing potentially underreported cognitive and speech-related adverse effects. The approach enhances sensitivity for detecting subtle neuropsychiatric effects that may be documented in clinical notes but missed by traditional reporting mechanisms.
Funding:
NIH NINDS Funding Number: K23NS124656