BISeizuRe: BERT-Inspired Seizure Data Representation to Improve Epilepsy Monitoring

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epilepsy diagnosis relies heavily on electroencephalogram (EEG) analysis.
• Accurate and efficient EEG-based seizure detection remains a clinical challenge.
• Transformer-based models show promise in analyzing complex biological signals.

Purpose of the Study:

• To develop and evaluate a novel BERT-based model, BENDR, for enhanced EEG-based seizure detection.
• To optimize the model's performance by examining fine-tuning strategies, architecture, and pre/post-processing techniques.
• To improve seizure detection sensitivity and reduce false positive rates in clinical EEG data.

Main Methods:

• A two-phase training approach was employed: initial pre-training on the Temple University Hospital EEG Corpus (TUEG) and subsequent fine-tuning on the CHB-MIT Scalp EEG Database.
• Extensive experimentation was conducted on the CHB-MIT dataset to optimize model architecture, pre-processing, and post-processing.
• A novel second pre-training phase was introduced before subject-specific fine-tuning to boost generalization.

Main Results:

• The optimized BENDR model achieved a significant reduction in false positives per hour (FP/h), reaching as low as 0.23, which is 2.5 times lower than the baseline.
• The model demonstrated a lower but clinically acceptable sensitivity rate.
• The BERT-based approach proved effective for EEG-based seizure detection, outperforming baseline methods.

Conclusions:

• The BENDR model offers a promising advancement in automated EEG-based seizure detection.
• The optimized model enhances clinical utility by improving accuracy and reducing false alarms.
• This approach has the potential to improve patient safety, enable personalized treatments, and generalize to new patient populations.