Automated Seizure Detection Using a Random Forest Classifier in a Mouse Model of Focal Cortical Seizures
Abstract number :
703
Submission category :
3. Neurophysiology / 3F. Animal Studies
Year :
2020
Submission ID :
2423043
Source :
www.aesnet.org
Presentation date :
12/7/2020 9:07:12 AM
Published date :
Nov 21, 2020, 02:24 AM
Authors :
Elizabeth Tilden, Washington University School of Medicine; Nicholas Rensing - Washington University School of Medicine; Michael Wong - Washington University School of Medicine; Thomas Foutz - Washington University School of Medicine;
Rationale:
Manual EEG scoring of seizure activity is the gold standard in research and clinical settings. This process can significantly limit the throughput of experiments as well as introduce bias when scoring groups with different experimental interventions. Random Forest classification is a form of machine learning that uses a set of defined features to sort a dataset into predetermined categories. The utility of a Random Forest classifier evaluating seizures is far-reaching and can be incorporated in various applications, including post hoc scoring of seizures, or real-time detection. Here, we present a Random Forest classifier trained to detect seizures from electrocorticographic recordings in a mouse model of focal cortical seizures.
Method:
4-aminopyridine (50 nL of 50 mM solution) was focally injected into the motor cortex (AP: +1, ML: +1, DV: -1 from Bregma) of adult mice with mixed C57Bl/6 and SV129 genetic background. Electrocorticographic 4-channel recordings were captured using implanted screw electrodes, with a sampling rate of 250-600 Hz. Signals were preprocessed with manual artifact removal and 1-100 Hz filtering. The training was performed using a pooled dataset (21 seizures) from two separate animals. EEGs were manually scored into three categories: early seizure, late seizure, and interictal behaviors. Features were selected through brute force optimization. Identifying features that delivered consistent results across experiments were included: power spectral density (1-48 Hz), total band power (1-10 Hz and 1-48 Hz), correlation coefficients, and eigenvalues in the frequency domain.
Results:
A cross-validation performance analysis of the algorithm was performed on the validation data set, which consisted of 84 seizures across the same two animals. The total performance was 76.7% and was consistent across animals (75.5%, 78.5%).
Conclusion:
This Random Forest classifier provides an unbiased alternative to manual EEG scoring of seizure in a mouse model of focal cortical seizures. This proof of concept demonstrates that the classifier has the potential to provide real-time unsupervised detection of seizures.
Funding:
:R01 NS056872/NS/NINDS NIH HHS/United States, R21 NS104522/NS/NINDS NIH HHS/United States
Neurophysiology