Time-Frequency Decomposition of Scalp Electroencephalograms Improves Deep Learning-Based Epilepsy Diagnosis

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Diagnostics

Background:

• Epilepsy diagnosis from electroencephalograms (EEGs) relies on identifying Interictal Epileptiform Discharges (IEDs), a process that is time-consuming and subjective.
• Automated detection and classification of IEDs are crucial for efficient and objective epilepsy diagnosis.

Purpose of the Study:

• To develop and evaluate an effective Interictal Epileptiform Discharge (IED) detector and an automated system for classifying EEG recordings as IED-free or IED-positive.
• To investigate the performance of Convolutional Neural Networks (ConvNets) using various input features and architectures for IED detection and EEG classification.

Main Methods:

• Explored Convolutional Neural Network (ConvNet) models, including 1D and 2D architectures, with temporal, spectral, and wavelet features.
• Investigated noise injection techniques and different input signal preprocessing methods.
• Evaluated the system's performance on five independent EEG datasets using metrics such as false detection rate, sensitivity, balanced accuracy (BAC), and area under the curve (AUC).

Main Results:

• The optimized 1D ConvNet achieved a false detection rate of 0.23/min at 90% sensitivity for IED detection.
• The EEG classification system demonstrated a mean Leave-One-Institution-Out (LOIO) cross-validation BAC of 78.1% (AUC 0.839) and Leave-One-Subject-Out (LOSO) CV BAC of 79.5% (AUC 0.856).
• The classification system analyzes a 30-min EEG in seconds, significantly reducing analysis time.

Conclusions:

• The proposed ConvNet-based system provides accurate and efficient detection of IEDs and classification of EEGs for epilepsy diagnosis.
• This automated approach has the potential to significantly reduce the workload for clinicians and improve the objectivity of epilepsy diagnosis.