Development and validation of a multimodal automatic interictal epileptiform discharge detection model: a prospective...

Area of Science:

• Artificial Intelligence in Medicine
• Neurology
• Medical Image Analysis

Background:

• Automated detection of interictal epileptiform discharges (IEDs) is crucial for epilepsy diagnosis, yet visual identification is time-consuming and expert-dependent.
• Developing accurate automated models can significantly improve diagnostic efficiency and accessibility.
• This study introduces a multimodal approach to address these challenges.

Purpose of the Study:

• To develop and validate a multimodal deep learning model (vEpiNetV2) for automated IED detection.
• To assess the model's performance across multiple clinical centers.
• To evaluate the contribution of video features to IED detection accuracy.

Main Methods:

• A deep learning model (vEpiNetV2) was trained on a large dataset of video-electroencephalogram (EEG) epochs from 530 patients.
• The model integrates both video and EEG features, incorporating bad channel removal and patient detection algorithms.
• Performance was validated prospectively across three centers using metrics like AUPRC and AUC.

Main Results:

• vEpiNetV2 achieved high accuracy in IED detection across three centers, with AUPRC/AUC values ranging from 0.76-0.78/0.96-0.98.
• Integrating video features improved precision by 5-9% and reduced false positives by 24% at 95% sensitivity.
• The model demonstrated robustness despite amplitude discrepancies and bad channels, highlighting the value of video data.

Conclusions:

• The multimodal vEpiNetV2 model, combining video and EEG, offers high precision and robustness for IED detection.
• Multi-center validation confirms its potential for real-world clinical application in epilepsy diagnosis.
• Video features significantly enhance IED analysis and model performance.