Discrete Wavelet Convolution for Learnable Time-Frequency Representation with Application to Seizure Prediction

Area of Science:

• Signal Processing
• Machine Learning
• Computational Neuroscience

Background:

• Accurate time-frequency representation is crucial for nonstationary signal analysis, particularly in epileptic seizure prediction using electroencephalogram (EEG) data.
• Existing deep learning methods often use fixed feature extraction and lack interpretability, hindering reliable seizure prediction.
• There is a need for adaptive and interpretable deep learning models for analyzing complex biological signals like EEG.

Purpose of the Study:

• To propose and validate Adaptive Discrete Wavelet Convolution (ADWConv) as a trainable wavelet decomposition front-end for seizure prediction.
• To enhance the adaptiveness and interpretability of deep learning models for analyzing nonstationary signals.
• To improve the accuracy and robustness of epileptic seizure prediction systems.

Main Methods:

• Developed ADWConv, a general-purpose trainable multi-level wavelet decomposition front-end that parameterizes filters and decouples kernel shapes.
• Implemented an end-to-end seizure prediction framework utilizing ADWConv.
• Employed a joint optimization strategy with regularization loss and wavelet priors for discriminative and interpretable filter learning.
• Validated the model on the CHB-MIT and SH-SDU EEG databases.

Main Results:

• The ADWConv-based model achieved 98.65% event-based sensitivity with a 0.014/h false prediction rate (FPR) on the CHB-MIT database.
• Achieved 96.53% event-based sensitivity with a 0.028/h FPR on the SH-SDU database.
• Ablation and visualization confirmed that learning wavelet kernels significantly improved performance and enabled task-specific adaptation of frequency bands.

Conclusions:

• ADWConv offers a novel approach to adaptive time-frequency representation for nonstationary signals.
• The proposed method significantly enhances the interpretability and accuracy of deep learning-based seizure prediction.
• ADWConv demonstrates competitive performance and robustness, paving the way for improved clinical applications in epilepsy monitoring.