Abstract: A methodology for understanding and detecting nonconvulsive seizures in individuals with subarachnoid hemorrhage is introduced. Specifically, beginning with an EEG signal, the power spectrum is estimated yielding a multivariate time series which is then analyzed using empirical orthogonal functional analysis. This methodology allows for easy identification and observation of seizures that are otherwise only identifiable though expert analysis of the raw EEG.
Introduction
Patients suffering from aneurysmal subarachnoid hemorrhage (SAH) are at significant risk for non-convulsive seizures (NCSz). Unlike typical seizures, NCSz lack outward physical symptoms, often manifesting only as subtle cognitive declines. Despite their "silent" nature, these events can cause secondary brain injury and severely impact patient recovery. While continuous EEG monitoring is the gold standard for detection, interpreting these signals is labor-intensive and prone to interference from artifacts in surface recordings. This paper presents a novel analytical pipeline using Empirical Orthogonal Functional (EOF) analysis to make these elusive seizures identifiable in surface EEG data.
Methodology: From Power Spectra to EOFs
The proposed methodology transforms raw EEG signals into a more interpretable format through a multi-step process. First, the power spectrum (PS) is estimated over fixed time intervals, converting the voltage-over-time signal into a multivariate time series of power across frequency bands (typically 1–20 Hz).
The core of the innovation lies in applying Empirical Orthogonal Functional (EOF) analysis—a technique closely related to Principal Component Analysis—to this multivariate power spectrum. This decomposes the signal into orthogonal components, allowing researchers to isolate the most significant patterns of brain activity.
The analysis focuses primarily on the "First EOF," which represents the direction of maximum variance in the frequency data. By tracking this component and the fraction of total variance it explains, the energetic shifts associated with seizure activity become visible.
Key Insights: Brain Activity Constraints
A central finding of this research is that seizures significantly constrain the brain's electrical complexity. During a seizure event, the fraction of variance captured by the first EOF increases dramatically. This suggests that the brain's activity becomes "locked" into a specific, highly energetic set of frequencies.
Furthermore, EOF analysis of surface EEG can reveal "depth seizures" that are otherwise invisible in standard surface power spectra. While standard visualizations might appear as unstructured noise, the EOF plots can clearly depict abrupt frequency changes and oscillations corresponding to seizure onset. For instance, pre-seizure frequency oscillations were observed to shift in period (e.g., from 25 to 12 minutes) at the start of a seizure.
Results and Validation
The EOF methodology was validated against a cohort of 26 SAH patients, with seizure events manually labeled by a neurointensivist as the ground truth. The results demonstrate a significant improvement over traditional methods:
- Detection Accuracy: Human observers using EOF visualizations achieved 88% accuracy in identifying seizures, nearly doubling the 48% accuracy achieved with standard Power Spectrum plots ($p = 0.008$).
- Correlation: There was a strong linear correlation ($\rho = 0.76$) between EOF-identified seizures and the expert gold standard.
Conclusion and Future Directions
While the current study is limited by a small sample size and the inherent noise of clinical data, it establishes a powerful framework for non-invasive seizure detection. Future work will focus on automating the detection process and exploring whether specific EOF "phenotypes" can predict long-term patient outcomes or guide personalized anti-seizure medication strategies.
Summary updated