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

Accurate and adaptive time-frequency representation is essential for analyzing non-stationary signals in critical applications, such as epileptic seizure prediction utilizing electroencephalogram (EEG) data. However, existing deep learning approaches often suffer from two major limitations: the reliance on fixed, non-adaptive feature extraction methods and the lack of model interpretability. To bridge these gaps, we propose Adaptive Discrete Wavelet Convolution (ADWConv), a general-purpose trainable multi-level wavelet decomposition front-end, with validation in an end-to-end seizure prediction framework. ADWConv parameterizes the low-frequency and highfrequency filters within each decomposition level and decouples their kernel shapes from level-specific scaling factors, enabling task-related adaptive weighting of frequency bands. This end-to-end learning is guided by a joint optimization strategy, which incorporates a regularization loss with wavelet priors to ensure the learned filters are both discriminative and physically interpretable. Extensive experiments were conducted on the public CHB-MIT database and the private SH-SDU database. On the CHB-MIT database, the ADWConv-based model attained an event-based sensitivity of 98.65%, while maintaining a false prediction rate (FPR) of only 0.014/h. Comparable results were also demonstrated on the SH-SDU database, attaining 96.53% event-based sensitivity with 0.028/h FPR. Ablation and visualization analyses confirm that learning the wavelet kernels is the primary driver of performance gains and supports task-specific adaptation of discriminative frequency bands. Overall, ADWConv improves the feature-level interpretability of a deep learning framework for seizure prediction while maintaining competitive accuracy and robustness.