LLM-parameterized simulated wearable data for seizure risk monitoring

This dataset provides entirely synthetic multimodal wearable physiological time-series data derived from a methodological study on seizure risk monitoring. The data were generated utilizing a simulation-based pipeline that blends LLM-parameterized patient profiles with a stochastic time-series generator. The LLM is only utilized to generate limited physiological parameters; all wearable signals are produced by a deterministic simulator. The collection contains roughly 98 hours of 1 Hz data from five simulated patients, which includes heart rate (HR), electrodermal activity (EDA), skin temperature (TEMP), and step rate. Data are provided as CSV files for each patient, with no temporal grouping or filtering. Binary seizure-risk labels are not explicitly kept; instead, they are calculated during downstream analysis using rule-based thresholding applied to physiological inputs, as described in the code. The dataset is designed for use in seizure risk monitoring, wearable health analytics, reinforcement learning, federated learning, rare-event detection, and class-imbalanced time-series analysis.