Interictal epileptiform discharge characteristics underlying expert interrater agreement

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Interictal epileptiform discharges (IED) on electroencephalograms (EEG) are crucial for epilepsy diagnosis.
• Inter-rater agreement (IRA) among experts on IED presence is often imperfect, potentially leading to misdiagnosis.
• Identifying attributes that influence expert agreement can enhance automated IED detection systems.

Purpose of the Study:

• To determine which electroencephalogram (EEG) signal attributes best predict expert agreement on the presence of interictal epileptiform discharges (IED).
• To develop computational models that emulate expert diagnostic capabilities for IED detection.
• To investigate the impact of the number of expert opinions on model performance.

Main Methods:

• 18 clinical neurophysiologists annotated IEDs in 200 EEG segments.
• 5538 waveform features, including wavelet coefficients and morphological characteristics, were extracted.
• Support vector regression (SVR) models were trained to predict expert opinions using selected features and local EEG normalization.

Main Results:

• Wavelet features, particularly specific basis functions, were most effective in predicting expert annotations.
• Local EEG normalization significantly improved model performance.
• Model performance plateaued with approximately 11 expert annotators, indicating a
• wisdom of the crowd
• effect.

Conclusions:

• Wavelet features combined with local EEG standardization are key predictors of expert agreement on IEDs.
• Computational models trained on a larger consensus of expert opinions (over 10) achieve optimal performance.
• These findings support the development of more accurate automated IED detection systems by leveraging quantifiable EEG features and collective expert judgment.