UC San Diego

Electroencephalography (EEG) reports contain clinically important information regarding epilepsy and brain function. However, these findings are typically documented as unstructured free text, limiting their accessibility for large-scale clinical research, quality improvement initiatives, and decision-support systems. We propose a hybrid natural language processing (NLP) pipeline that leverages both fuzzy-matching and open-weight generative large language models (LLMs) to extract relevant features from UCSD Neurology’s highly heterogeneous and time-ambiguous EEG reports. Our best-performing quantized method, Qwen3 14B Q6_K with one-shot quote-extraction prompting, achieved a macro F1 of 0.72 (0.845, micro F1), macro sensitivity of 0.832, specificity of 0.962, and a mean processing report time of 7.2s.