LightIED: Explainable AI with Light CNN for Interictal Epileptiform Discharge Detection

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Imaging

Background:

• Interictal epileptic discharge (IED) detection from electroencephalography (EEG) is crucial for epilepsy diagnosis but remains challenging.
• Current deep learning models for IED detection often lack explainability and possess complex structures, limiting clinical adoption.
• There is a need for lightweight, explainable models to assist clinicians in IED detection and reduce diagnostic workload.

Purpose of the Study:

• To introduce LightIED, a novel, lightweight, and explainable machine learning model for automated IED detection in EEG.
• To evaluate the performance of LightIED against state-of-the-art models in terms of accuracy and efficiency.
• To utilize Grad-CAM for visualizing the model's inference basis, enhancing interpretability.

Main Methods:

• EEG data was transformed into image format for input into the LightIED model.
• The LightIED model was developed using machine learning principles, focusing on lightweight architecture.
• Gradient-weighted Class Activation Mapping (Grad-CAM) was employed for visualizing model predictions and identifying key features for IED detection.

Main Results:

• LightIED achieved IED detection accuracy comparable to the state-of-the-art Satelight model and surpassed other Vision Transformer-based models.
• The proposed LightIED model utilizes less than one-third of the parameters compared to Satelight, indicating superior efficiency.
• Grad-CAM visualizations effectively highlighted the specific EEG regions indicative of IEDs, confirming the model's interpretability.

Conclusions:

• LightIED presents a highly effective and efficient solution for IED detection in EEG, offering a valuable tool for epilepsy diagnosis.
• The model's lightweight and explainable nature makes it a practical alternative to complex deep learning approaches.
• The integration of Grad-CAM provides crucial insights into the model's decision-making process, fostering trust and clinical utility.