PaperlayerView exact revision on arXiv

Source-linked AI summary

Focal onset seizure prediction using convolutional networks

Haidar Khan, Lara Marcuse, Madeline Fields, Kalina Swann, Bülent Yener

arXiv:1805.11576v1cs.LGstat.ML

TL;DR

The paper asks whether focal seizures can be predicted from scalp EEG despite uncertainty about the interictal-to-preictal transition and optimal prediction horizon. It learns wavelet-based convolutional features and the horizon from data, identifying an approximately ten-minute transition and promising test-set performance. The authors conclude that automatically learned scalp-EEG features can characterize the preictal state of focal seizures.

  • Problem

    The work addresses the lack of a clear, robust, patient-independent definition of the preictal prediction horizon for focal seizures using scalp EEG.

  • Method

    Convolutional filters learn quantitative signatures from wavelet-transformed scalp EEG for interictal, preictal, and ictal periods while optimizing the seizure prediction horizon from data.

  • Results

    The learned horizon is approximately ten minutes before seizure onset, with test-set sensitivity of 87.8% and false prediction rate of 0.142 FP/h.

  • Takeaways & Limitations

    Automatically extracted scalp-EEG features can characterize the preictal state of focal seizures and support a feasible reliable seizure-prediction system.

  • Takeaways & Limitations

    The limited scalp-EEG database must be expanded to fully test the system's generalization ability.

Abstract

from arXiv · show

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.

I. INTRODUCTION

The paper addresses the difficult problem of predicting focal seizures from scalp EEG, where artifacts and an unclear interictal–preictal transition complicate reliable prediction. It applies automatic feature extraction toward a non-invasive alert system.

  • Motivation: Epileptic seizures are unpredictable, while scalp EEG offers a non-invasive but artifact-prone basis for prediction.Scalp EEG is more accessible than intracranial EEG but is affected by electrode shifts, muscle interference, and volume conduction.
  • Problem: The central challenge is defining a general, robust transition between interictal and preictal EEG periods.Existing estimates of the preictal period have varied from hours to minutes and have not been shown to generalize.
  • Contribution: This work applies automatic feature extraction to scalp EEG to support alerts about oncoming seizures.The broader goal is a wearable, non-invasive EEG device that could alert patients, family members, and clinicians.
  • Prior work: The study builds on evidence that some recent seizure-prediction algorithms classify interictal versus preictal data above random levels.Prior work has nevertheless included many methods that failed to predict seizures reliably.
  • Prior work: Earlier approaches combined manually extracted linear or nonlinear features with statistical and algorithmic prediction methods.Reported approaches include multivariate time-series analysis, explicit thresholds, support vector machines, random forests, and autoencoders.

II. METHODS

The study assumes a preictal phase and seeks a patient-independent prediction horizon directly from EEG data. It uses seizure-containing and interictal recordings for feature learning, horizon selection, and evaluation.

  • Problem Definition: The study assumes focal seizures have a preictal phase with an inflection point separating interictal and preictal states.The prediction horizon is defined as the time window before seizure onset in which a seizure is anticipated.
  • Problem Definition: The prediction horizon is learned from data because its optimal length is not well understood.The goal is a justifiable patient-independent horizon that supports early EEG prediction of the phase transition.
  • Experimental Design: Seizure-containing recordings support feature learning, preictal-period determination, and sensitivity evaluation, while interictal recordings test specificity.The study evaluates the model on independent MSSM and CHB-MIT datasets.

1) MSSM EEG Dataset:

The datasets contain focal-seizure and seizure-free scalp EEG recordings from MSSM and CHB-MIT. Each recording is transformed into a time–scale–channel wavelet tensor for analysis and CNN input.

  • MSSM EEG Dataset: MSSM provides 86 scalp EEG recordings from 28 patients, acquired with 22 inputs at 256 Hz.Continuous studies lasted 2–8 days and used de-identified recordings.
  • MSSM EEG Dataset: MSSM recordings retain sleep–wake, eye-blink, movement, and chewing artifacts, while multi-electrode artifacts were excluded.Expert reviewers could not distinguish a preictal signal directly in the raw EEG recordings.
  • MSSM EEG Dataset: The study includes focal clinical and subclinical seizures but excludes generalized-onset seizures.Each patient had more than one seizure during recording.
  • CHB-MIT EEG Dataset: CHB-MIT contributes scalp EEG from 22 patients, including recordings with labeled seizures and recordings without seizure activity.A suitable subset was selected for training and testing.
  • Wavelet Representation: The continuous wavelet transform represents each EEG channel across time and scales, producing a three-mode tensor.The displayed tensor contains 22 channels and 10 scales.
  • Wavelet Representation: Wavelet transformation provides multiscale frequency information and can isolate noise before CNN processing.The transformed signal is used both for analysis and as network input.

C. Methodology: Overview and Background

The method uses CNNs on wavelet-transformed EEG to learn features distinguishing interictal, preictal, and ictal states. It treats preictal duration as a learnable parameter and optimizes it across candidate lengths.

  • Aim 1: Convolutional Neural Networks for Feature Extraction: CNN features are learned to distinguish preictal, ictal, and interictal EEG examples.The first aim is automatic feature extraction; the second is estimating the preictal period and prediction horizon.
  • Aim 1: Convolutional Neural Networks for Feature Extraction: Convolutional filters capture short-term temporal dependencies and relationships between nearby frequency bands in wavelet-transformed EEG.Weight sharing and local dependencies reduce the number of trained parameters.
  • Aim 1: Convolutional Neural Networks for Feature Extraction: Max-pooling supports temporal invariance, while dropout randomly removes unit outputs during training.These components help the network identify wavelet-tensor patterns independently of their temporal location and reduce overfitting.
  • Aim 1: Convolutional Neural Networks for Feature Extraction: The network uses six convolutional layers, two dense layers, and a three-unit soft-max output representing class probabilities.The architecture was selected after comparing shallower, wider, and deeper alternatives.
  • Aim 2: Computing the Length of the Preictal Period: Incorrect preictal labels can contaminate either the preictal or interictal class and impair supervised feature learning.Both overlong and truncated preictal labels introduce mislabeled examples into the training classes.
  • Aim 2: Computing the Length of the Preictal Period: The preictal length is treated as a learning parameter and selected by grid search, with extracted features analyzed for validation.This approach addresses literature estimates that have not been shown to generalize.

D. Methodology: A Prediction System

The prediction system transforms scalp EEG into wavelet-based inputs, learns interictal, preictal, and ictal patterns with a CNN, and converts class probabilities into smoothed seizure-imminence predictions. Cross-validation selects the preictal length and other operating parameters.

  • Wavelet transforms convert each EEG recording into tensors that serve as CNN inputs.The transform uses scales from 1 to 512 and the Mexican-hat mother wavelet.
  • The system labels windows as interictal, preictal, or ictal using seizure onset and an assumed preictal length.Windows within l minutes before onset are preictal, windows after onset are ictal, and the remainder are interictal.
  • 10-fold cross-validation selects hyperparameters by minimizing average validation loss, including a 10-minute preictal length.The final settings include 1-second epochs, 0% overlap, α = 0.7, and l = 10 minutes.
  • The trained network produces three-class probabilities, with oncoming-seizure probability defined as p = p1 + p2 = 1 − p0.Here p0, p1, and p2 correspond to interictal, preictal, and ictal probabilities.
  • Exponential smoothing couples adjacent predictions, and the system declares a seizure imminent when the smoothed signal crosses a threshold.Results use an empirically determined threshold of 0.6, while performance is evaluated with ROC-AUC.
  • The CNN’s 100-unit final hidden layer yields a T x 100 feature matrix for analyzing interictal-to-preictal transitions.These compressed features are observed for each epoch.

III. RESULTS

The study evaluates whether CNN-learned features capture the transition from interictal to preictal activity using Kullback-Leibler divergence.

  • KL divergence is used to verify that CNN-learned features capture the interictal-to-preictal phase transition.

A. Extracted Feature Analysis for Phase Transition

Feature-distribution analysis identifies a consistent transition approximately 9–10 minutes before seizure onset, supporting a 10-minute prediction horizon. Test-set results include wavelet-based performance gains and MSSM prediction outcomes.

  • A. Extracted Feature Analysis for Phase Transition: The earliest significant divergence between interictal and current feature distributions occurs approximately 9–10 minutes before seizure onset.The divergence remains small initially and then increases rapidly.
  • A. Extracted Feature Analysis for Phase Transition: The divergence shift occurs at the same time across preictal-length settings and appears consistently in validation recordings.This supports the learned features’ capture of the underlying interictal-to-preictal transition.
  • B. Test Set Results: Wavelet-transformed EEG improves Matthews correlation coefficient across all test-set recordings compared with raw EEG input.The authors attribute this improvement partly to filtering across multiple frequency bands.
  • 1) MSSM Test Set:: 8 minutes was the average seizure prediction time before onset in the MSSM test set.The set contained 18 recordings, including 15 recordings with seizures.
  • 1) MSSM Test Set:: 1 of 15 tested seizures was not predicted, and the false prediction rate was .128/hr.False positives occurred in seizure recordings long before the 10-minute prediction horizon, but not in interictal-only recordings.
  • 1) MSSM Test Set:: Figure 4 overlays interictal, preictal, and ictal regions with the CNN’s oncoming-seizure probability and threshold-crossing prediction.

2) CHB-MIT Test Set:

The CHB-MIT test set evaluated prediction timing, false predictions, and error patterns across seizure and interictal recordings. Performance varied across recordings, with some errors associated with poorly conditioned wavelet matrices and difficult EEG characteristics.

  • 18 seizure recordings and 50 interictal recordings were evaluated, totaling approximately 70.5 hours of EEG.
  • 6 minutes was the average prediction time before seizure onset, with 3 of 18 seizures not predicted.
  • 0.147/h was the false prediction rate, and false predictions occurred within interictal-only recordings.
  • Error Analysis: Poorly conditioned wavelet matrices were associated with MSSM false positives, indicated by low spectral gaps and high condition numbers.
  • Error Analysis: False positives in CHB-MIT and false negatives in both datasets were not separated from error-free recordings by the analyzed linear properties.
  • Error Analysis: Four missed seizures involved low-amplitude EEG, muscle, electrode, movement, or sleep/wake and epileptiform-burst characteristics.

D. Sensitivity and Comparison to Unspecific Random Predictor

The comparison evaluates sensitivity against an unspecific random predictor and contrasts this work with prior algorithms using a ten-minute prediction horizon and observed false prediction rates.

  • The random-predictor sensitivity range was calculated with SOP = 10 minutes, observed FPr values, and α = 0.05.The analysis set the number of independent features to the upper bound d = 100.
  • Table IV compares specificity and false prediction rates for this work and other seizure-prediction algorithms.Specificity is represented by the false prediction rate, calculated as false predictions divided by interictal-only recording duration.
  • Sensitivity above the σhigh value indicates performance better than chance at the given false prediction rate.

IV. COMPUTATIONAL COMPARISON TO PRIOR WORK

The learned features were tested out of sample against published and Kaggle seizure-prediction methods. The proposed system achieved the highest reported sensitivity while also attaining a lower false prediction rate.

  • The learned features produced out-of-sample sensitivity and specificity superior to other published seizure-prediction methods.The comparison included two top-performing algorithms from the Kaggle seizure-prediction competition.
  • 87.8% sensitivity and 0.142 FP/h were achieved by this system, exceeding the other compared methods on both reported measures.

V. DISCUSSION

The automatically extracted scalp-EEG features reveal a preictal transition approximately 10 minutes before seizure onset, supporting feasibility for reliable seizure prediction. However, variability and limited data remain barriers to a complete system.

  • 10 minutes before seizure onset, the preictal phase transition appears in scalp EEG using automatically extracted features.The prediction-time spread is large, with σ = 98.7.
  • Typical EEG artifacts and sleep-wake transitions did not cause false positives.
  • The large prediction-time spread is a drawback of training across patients because patient-specific data were limited.
  • A reliable seizure prediction system could support wearable, non-invasive alerts for patients, family members, and doctors.The proposed benefits include enhanced patient safety and potentially reduced SUDEP rates.
  • A complete system requires a larger scalp-EEG database to test generalization and a clearer understanding of ictal-state termination for predictor resetting.The authors also identify end-to-end optimization as a future step.
Loading 1805.11576v1…