Abstracts

Rationale: Spike ripples, high-frequency oscillations in brain voltage recordings that co-occur with interictal discharges, have emerged as potential biomarkers for epileptogenic cortex. However, detecting spike ripples remains challenging due to the labor-intensive nature of manual interpretation and limitations in current automated methods, which are hindered by dataset heterogeneity and scarcity of training data. Here we explore the use of data augmentation to enhance the performance of long-short term memory (LSTM) neural networks in detecting spike ripples.

Methods: We developed a novel data augmentation strategy to generate synthetic spike ripple events. To do so, we analyzed 677 manually identified spike ripple events from 6 patients with epilepsy. From these events, we estimated features of the spike-wave (e.g., amplitude and timing of peaks) and ripple (e.g., center frequency and bandwidth) components. We then combined these estimated features with information from the literature to simulate synthetic spike ripple and spike events. We used these synthetic events to train and test two long-short term memory (LSTM) neural network architectures for spike ripple detection: one network operating in the frequency domain, and another in the time domain. We trained the neural networks using three different data sets: in vivo data alone, synthetic data alone, or a combination of in vivo and synthetic data. Classification performance of the neural networks was evaluated using leave-one-out cross-validation across various performance metrics, including sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve (AUC).


Results: Data augmentation produced synthetic spike ripples consistent with in vivo observations. Training with synthetic data alone or combined with in vivo data improved the classification performance of both LSTM models, although the improvement was marginal. The frequency-domain model showed a marginal improvement in AUC from 0.80 to 0.85, while the time-domain model improved from 0.80 to 0.83. Additionally, augmented training data led to improved sensitivity and accuracy for the frequency-domain method compared to training on in vivo data alone, demonstrating the effectiveness of the data augmentation strategy.

Conclusions: This study demonstrates that data augmentation can enhance the performance of LSTM neural networks in detecting spike ripples in brain voltage recordings. By generating a more comprehensive and diverse training dataset, this approach addresses the challenges posed by dataset heterogeneity and scarcity, leading to new strategies for improved diagnostic accuracy for epilepsy treatment.

Funding: NIH R01NS110669, NIH R01NS119483, NSF Award 1451384