Convolutional neural network with autoencoder-assisted multiclass labelling for seizure detection based on scalp...

Area of Science:

• Medical technology
• Artificial intelligence in medicine
• Neurology

Background:

• Automatic seizure detection using convolutional neural networks (CNNs) shows promise for reducing epileptologist workload.
• Existing CNN models achieve high seizure detection performance but struggle with false-positive alarm rates.

Purpose of the Study:

• To reduce false alarms in automatic seizure detection by combining CNNs with patient-specific autoencoders (AE).
• To develop a more reliable seizure detection system by incorporating an AE for logging abnormalities.

Main Methods:

• Combined a CNN processing EEG plot images with patient-specific AEs analyzing EEG signals.
• Developed a 3-class CNN (seizure, non-seizure-but-abnormal, non-seizure) based on expert logs and AE errors.
• Utilized an accumulative measure of consecutive seizure labels to trigger alarms.

Main Results:

• The AE-CNN achieved comparable second-by-second classification performance to the original CNN.
• The AE-CNN significantly reduced the median false alarm rate to 0.034 h⁻¹, a fivefold decrease from the original CNN's 0.17 h⁻¹.
• "Non-seizure-but-abnormal" labels effectively interrupted false-positive seizure alarms.

Conclusions:

• The "non-seizure-but-abnormal" label provides practical benefits for seizure detection accuracy.
• Modifying CNNs with AEs is a valuable approach for unsupervised abnormality labeling, reducing demands on epileptologists.
• This AE-CNN method offers a more efficient and accurate tool for long-term video-monitoring seizure detection.