Interictal Epileptiform Discharge Detection Through Probabilistic Diffusion Models with Maximization of Precision...

Area of Science:

• Neuroscience
• Medical Informatics
• Machine Learning

Background:

• Automated Interictal Epileptiform Discharge (IED) detection aids early epilepsy diagnosis by analyzing Electroencephalographic (EEG) data.
• Existing methods struggle with imbalanced datasets (IEDs vs. background data) and poor cross-institutional precision.
• This leads to inefficient neurologist review due to false positives.

Purpose of the Study:

• To develop a novel approach for robust and generalizable automated IED detection.
• To address data scarcity and imbalance issues in EEG datasets.
• To improve the precision and F1-score of IED detection models, especially in cross-institutional settings.

Main Methods:

• Utilized probabilistic diffusion models for data augmentation to synthesize IED data.
• Employed Area Under the Precision-Recall Curve (AUPRC) maximization for model training.
• Balanced training datasets by combining real and synthesized IEDs.

Main Results:

• Achieved a 4.5% increase in precision and a 0.7% improvement in F1-score in within-data evaluation.
• Demonstrated significant improvements in cross-data evaluation: 40.04% precision and 18.74% F1-score enhancement.
• The proposed method shows robustness and generalizability across different datasets.

Conclusions:

• The novel approach effectively tackles data imbalance and scarcity in IED detection.
• Probabilistic diffusion models and AUPRC maximization enhance model performance and reliability.
• This method offers a promising solution for improving automated IED detection in clinical practice.