Semi-Automated Detection, Annotation, and Prognostic Assessment of Ictal Chirps in Intracranial EEG from Patients with Epilepsy

Abstract

We analyzed the spectro-temporal characteristics of ictal "chirp" events in intracranial EEG (iEEG) recordings from 13 epilepsy patients, using a custom derivative dataset of 22,721 spectrograms that we generated from the Epilepsy-iEEG-Multicenter Dataset. Ictal chirps, transient frequency-modulated patterns, were semi-automatically annotated to assess their relationship with seizure onset zones (SOZs) and surgical outcomes. Preprocessing included notch filtering (60 Hz, 120 Hz) and bandpass filtering (1-60 Hz), followed by segmentation into 60-second windows. Spectrograms were generated via Short-Time Fourier Transform (STFT) with a Hann window (87.5% overlap) and converted to dB scale. Chirps were annotated by manually drawing bounding boxes, followed by automated ridge detection, model fitting, and feature extraction (start/end time-frequency, duration, direction, RMSE, R2). Spatial, Spectro-temporal, and clinical features were analyzed using heatmaps, hierarchical clustering, statistical tests (Mann-Whitney U), and outcome prediction models. Patient-channel mappings revealed clustering of chirp patterns among specific patient pairs, correlating with shared clinical profiles. Flow-based analysis demonstrated prognostic value: very high spectral durations in SOZ regions were associated with favorable surgical outcomes (80.43% success rate), whereas very high temporal durations in SOZ correlated with poorer outcomes (51.35% risk). Statistical comparisons showed significant differences between SOZ and non-SOZ chirps: SOZ chirps exhibited longer spectral durations (10.13 {+/-} 6.35 Hz vs. 8.51 {+/-} 5.66 Hz, *p* < 0.001), shorter temporal durations (6.76 {+/-} 5.83 s vs. 7.14 {+/-} 5.39 s, *p* = 0.006), and higher spectro-temporal ratios (2.66 vs. 1.92, *p* < 0.001). Distribution analyses further indicated that prolonged temporal chirps were more prevalent in non-SOZ regions.