Accuracy-Constrained CNN Pruning for Efficient and Reliable EEG-Based Seizure Detection

CNN's capture local temporal and spectral patterns, which enables improved predictive performance, but they are often large and computationally expensive. The multitude of filters added by a CNN will impose some redundancy, where some filters are irrelevant to predictive performance. This redundancy elicited prohibitive costs incurred in resource-constrained real-time clinical settings. Existing compression and pruning methods are variably acceptable in that they all seek to limit the size of the model at the expense of accuracy. Specifically, when accuracy applies, price-performance tradeoffs are unacceptable if the application is medical, where an even greater expectation of reliability exists alongside cost-to-performance considerations. Methods derived from the vision literature that were developed to mitigate the impact of redundancy on the real-time operationalization of convolutional networks are unlikely to be effective with data obtained from EEG signals, due to the unique nature of those signals [4]. In contrast to vision data, these data sources are highly susceptible to noise and temporal variations. This paper presents a pruning framework that is sensitive to seizures for the detection of seizures. We first train a light-weight 1D CNN using early stopping to stabilize the generalization from the training data to the test data.