Focal onset seizure prediction using convolutional networks (1805.11576v1)

Abstract: Objective: This work investigates the hypothesis that focal seizures can be predicted using scalp electroencephalogram (EEG) data. Our first aim is to learn features that distinguish between the interictal and preictal regions. The second aim is to define a prediction horizon in which the prediction is as accurate and as early as possible, clearly two competing objectives. Methods: Convolutional filters on the wavelet transformation of the EEG signal are used to define and learn quantitative signatures for each period: interictal, preictal, and ictal. The optimal seizure prediction horizon is also learned from the data as opposed to making an a priori assumption. Results: Computational solutions to the optimization problem indicate a ten-minute seizure prediction horizon. This result is verified by measuring Kullback-Leibler divergence on the distributions of the automatically extracted features. Conclusion: The results on the EEG database of 204 recordings demonstrate that (i) the preictal phase transition occurs approximately ten minutes before seizure onset, and (ii) the prediction results on the test set are promising, with a sensitivity of 87.8% and a low false prediction rate of 0.142 FP/h. Our results significantly outperform a random predictor and other seizure prediction algorithms. Significance: We demonstrate that a robust set of features can be learned from scalp EEG that characterize the preictal state of focal seizures.

• The paper introduces a CNN-based method that automates feature extraction from wavelet-transformed EEG data for focal seizure prediction.
• It optimizes the preictal prediction horizon to approximately 10 minutes, achieving an 87.8% sensitivity and a false prediction rate of 0.142 FP/h.
• Robust cross-validation on both MSSM and CHB-MIT datasets demonstrates its superior performance compared to traditional methods.

An In-depth Analysis of "Focal Onset Seizure Prediction Using Convolutional Networks"

The paper "Focal Onset Seizure Prediction Using Convolutional Networks" by Khan et al. embarks on the challenging problem of predicting focal seizures using scalp electroencephalogram (EEG) data. This paper leverages convolutional neural networks (CNNs) to extract features from EEG signals, aiming to distiniguish interictal and preictal states and specify a prediction horizon.

Key Contributions and Methodology

1. Feature Learning via Convolutional Networks: The authors employ CNN to automate the feature extraction process from wavelet-transformed EEG data, diverging from traditional hand-crafted feature approaches. Their system learns filter maps that capture relationships across time and frequency domains, essential for discerning the subtle transitions in EEG data indicative of the preictal state.
2. Seizure Prediction Horizon Optimization: A significant methodological innovation is the data-driven determination of the preictal period length, challenging the common practice of assuming these parameter values a priori. They employ grid search for selecting an optimal preictal length, ultimately identifying an approximately ten-minute prediction horizon around seizure onset.
3. Data and Validation: This paper undertakes a robust cross-validation approach using two datasets—the Mount Sinai Epilepsy Center (MSSM) and the publicly available CHB-MIT database. The paper tests its system on a multitude of EEG recordings, thus ensuring the obtained results are not confined to overfitting a specific dataset.

Results and Performance

The paper presents compelling results, showcasing a sensitivity of 87.8% with a false prediction rate of 0.142 false positives per hour (FP/h). These measures outperform several contemporary seizure prediction algorithms, including top contenders from the Kaggle seizure prediction competition. Notably, the work achieves significant statistical validation compared to an unspecific random predictor, emphasizing the practicality and soundness of the proposed prediction methodology.

Implications and Future Work

The research encapsulated in this paper provides meaningful insights into automated seizure prediction systems. On a theoretical level, the paper reaffirms the presence and detectability of a preictal phase transition around ten minutes before seizure onset. Practically, this work highlights the feasibility of developing non-invasive, accurate seizure forecasting systems, which could materially improve patient safety and care through timely alerts.

Looking ahead, potential advancements would include increasing the generalization of the current system by incorporating richer, more diverse datasets comprising intracranial EEG signals. Such efforts could refine the prediction horizon further and reduce the variability in prediction times. Furthermore, integrating end-to-end deep learning architectures focused directly on seizure prediction tasks could enhance the precision of model predictions by leveraging deeper insights into the temporal dynamics of EEG data.

Conclusion

This paper encapsulates a well-rounded approach to advancing seizure prediction technologies, marrying the complex interplay of EEG signal processing with modern machine learning paradigms such as CNNs. Their research underlines the value of leveraging automatic feature extraction to capture complex physiological phenomena and sets a precedent for future explorations in this critical area of healthcare technology.