Interpretable EEG biomarkers for neurological disease models in mice using bag-of-waves classifiers

Area of Science:

• Neuroscience
• Genetics
• Computational Biology

Background:

• Electroencephalograms (EEGs) record brain electrical activity, revealing waveform patterns that can serve as phenotypical biomarkers.
• Identifying neurological disease genotypes, particularly epilepsy, is crucial, especially when seizures are not directly observed in clinical or animal models.
• Subtle neurological phenotypes in animal models often require advanced methods for genotype-specific activity detection.

Purpose of the Study:

• Investigate the prediction of genotypes from long-term EEG signals in freely behaving mice.
• Determine if EEG waveform patterns can reliably identify specific genetic backgrounds and disease-related genotypes (TSC1 gene knockout).
• Explore the potential of EEG as an interpretable phenotype for genotype prediction.

Main Methods:

• A machine learning approach was developed to predict genotypes based on the occurrence counts of EEG waveforms.
• Waveform dictionaries were optimized to approximate EEG windows for each genotype.
• Vectors of waveform occurrence counts were used as features for logistic regression-based genotype prediction.

Main Results:

• Waveform counts from multi-hour EEG segments enabled reliable mouse strain prediction with 70% accuracy.
• Strain-specific classifiers accurately identified the epilepsy-genotype (TSC1 knockout) in DBA2 (86%) and C57B6 (67%) mice, without overt seizures.
• A state-of-the-art method achieved higher strain classification (98%) but lacked interpretability.

Conclusions:

• EEG waveforms can serve as interpretable phenotypes for identifying epilepsy genotypes.
• The 'bag-of-waves' feature representation is effective for genotype prediction from EEG data.
• This approach offers a valuable tool for understanding genotype-specific neural activity and diagnosing neurological conditions.