Automatic generation of text-based summaries of intracranial eeg seizure events using large language models (llms)

Abstract

Precise localization of epileptogenic tissue from intracranial electroencephalography (iEEG) is essential for successful epilepsy surgery, but methods used in the clinical environment remain labor-intensive, time-consuming, and difficult to standardize. This thesis presents a system for the automatic generation of reports with iEEG findings by combining signal processing, deep learning, and large language models (LLMs) techniques. The system consists of two branches: high-frequency oscillation (HFO) detection for epileptogenic zone localization and seizure detection for ictal event detection. HFO analysis was first developed with the OpenNeuro ds003498 dataset and then adapted to the HUP dataset, while seizure detection was initially explored on the SWEC-ETHZ dataset and subsequently extended to HUP. For HUP seizure detection, a CNN-transformer architecture was used and improved with model simplification, regularization, and windowing changes. Across 43 subjects, the final model achieved mean subject-level accuracy and balanced accuracy of 0.83, with median values of 0.87. The median F1-score was 0.84, while median AUROC and AUPRC reached 0.97 and 0.95, respectively, although performance remained heterogeneous across subjects. The structured outputs of both branches were exported as JSON files and used as inputs to four LLM system versions. Version 1 focused on report generation, Version 2 introduced channel-level seizure evidence and deterministic cross-modal fusion, Version 3 added schema-validated reasoning and deterministic report rendering, and Version 4 extended the system with bounded tool orchestration and planning. The framework demonstrates a step toward more automated and standardized iEEG reporting.