Investigates the relationship between chronic cannabis use and temporal lobe epilepsy using EEG-derived machine learning features. Spectral and connectivity biomarkers extracted from HD-EEG recordings are used to characterize network-level differences in affected populations, with classification performance evaluated across seizure onset zones.

Identifies and evaluates a comprehensive set of engineering and clinical biomarkers for machine learning based seizure prediction in hospitalized patients. Examines the interplay between EEG-derived signal features and clinical covariates to improve prediction performance in acute care settings, with emphasis on features that generalize across patient populations and monitoring configurations.

Presents a survival analysis of 117 hospitalized patients with electrographic or electroclinical seizures undergoing cEEG at the University of Wisconsin Hospital between 2018 and 2022, establishing data-driven thresholds for monitoring duration after seizure detection. Demonstrates that status epilepticus is the only clinical feature significantly associated with increased recurrence risk, and quantifies the EEG duration required to reduce that risk below 5% — 36.8 hours in patients with status epilepticus and 21.2 hours in those without.

Presents Argus, a multi-scale convolutional and bidirectional LSTM network with an attention mechanism for automated seizure detection. Trained and evaluated on the Temple University Seizure Corpus (TUSZ v1.5.2) using a 22-channel TCP bipolar montage at 200 Hz, achieving 98.06% ROC-AUC. The architecture is designed with clinical deployment in mind, prioritizing low false positive rates in continuous long-term monitoring settings. No manuscript in preparation at this stage.

Investigates the electrophysiologic and network-level features distinguishing conscious experience (CE) from no conscious experience (NCE) during NREM stage 2 and 3 sleep using HD-EEG combined with machine learning classification. In a serial-awakening paradigm across 140 participants and 699 awakenings, spectral power, functional connectivity via lagged coherence, and graph-theoretic metrics including modularity, global efficiency, and rich-club coefficient were extracted and screened using false discovery rate correction. A probabilistic ensemble classifier achieves ROC-AUC 0.80 and average precision 0.80 with stable 5-fold participant-level cross-validation, with the largest discriminating effects observed in posterior alpha, delta, and gamma-to-alpha power ratios.