Authors :
Presenting Author: Aya Khalaf, PhD – Yale
Matthew Gruen, BS – Yale University
Alberto Prat, MD – Yale
Alexandre Castro, MD – Yale
Aziza Ysmanova, MD – Yale
Emmanuel Makinde, Undergraduate – New York University
Basil Abdalla, MD – Yale
Nasrin Shahana, MD – Yale
Harry Sutherland, MD – Yale
Jay Liu, MD – Yale
Gregory Worrel, MD,PhD – Mayo Clinics
Hal Blumenfeld, MD, PhD – Yale University
Rationale:
Temporal lobe seizures frequently impair consciousness, significantly affecting patient safety and quality of life. Clinical assessment of consciousness during seizures can be challenging, particularly when behavioral testing is not available or when seizures occur without obvious behavioral manifestations. We developed a machine learning approach to predict consciousness impairment in temporal onset seizures using scalp EEG recording, providing an objective tool to assist clinical decision-making.
Methods:
We developed a machine learning framework to analyze scalp EEG recordings from temporal onset seizures (N=76 seizures, 35 impaired). Our approach incorporated common spatial pattern (CSP) analysis to identify spatial features that maximize separability between spared and impaired consciousness states. Additional extracted features included time-domain and frequency-domain seizure characteristics from both pre-ictal and ictal periods. We employed support vector machines and linear discriminant analysis for classification, with model validation performed using 10-fold cross-validation. Information fusion from pre-ictal and ictal periods was performed using feature vector concatenation and probabilistic fusion. The system was designed to support clinical decision-making for medication adjustments, therefore, the objective was to maximize overall accuracy and balanced performance across impaired and spared consciousness states.
Results:
Our machine learning model achieved 82.8% overall accuracy in classifying consciousness states during temporal onset seizures, with balanced sensitivity for spared (82.9%) and impaired (82.8%) consciousness. CSP-derived spatial features proved most informative for distinguishing consciousness states, effectively capturing the complex spatiotemporal dynamics of temporal lobe seizure propagation. These results demonstrate promising feasibility for clinical applications in temporal lobe epilepsy management.
Conclusions:
Our findings demonstrate that scalp EEG-based machine learning can effectively predict consciousness impairment in temporal onset seizures. The balanced accuracy achieved supports the potential clinical utility of this approach for guiding therapeutic decisions, particularly medication optimization. Our findings demonstrate that spatial filtering techniques can effectively capture the neurophysiological signatures of consciousness impairment specific to temporal lobe seizures. Implementation of this automated system could provide clinicians with an objective tool for consciousness assessment in temporal lobe epilepsy, potentially improving treatment precision and patient outcomes.
Funding: