[Paper Review] A Generalised Seizure Prediction with Convolutional Neural Networks for Intracranial and Scalp Electroencephalogram Data Analysis

Nhan Duy Truong, Anh Duy Nguyen+4 more|arXiv (Cornell University)|Jul 6, 2017

EEG and Brain-Computer Interfaces21 references59 citations

TL;DR

The paper presents a generalized, minimal-feature-engineering seizure prediction method using CNNs on Short-Time Fourier Transform representations of 30-second EEG windows, applicable to both intracranial and scalp data.

ABSTRACT

Seizure prediction has attracted a growing attention as one of the most challenging predictive data analysis efforts in order to improve the life of patients living with drug-resistant epilepsy and tonic seizures. Many outstanding works have been reporting great results in providing a sensible indirect (warning systems) or direct (interactive neural-stimulation) control over refractory seizures, some of which achieved high performance. However, many works put heavily handcraft feature extraction and/or carefully tailored feature engineering to each patient to achieve very high sensitivity and low false prediction rate for a particular dataset. This limits the benefit of their approaches if a different dataset is used. In this paper we apply Convolutional Neural Networks (CNNs) on different intracranial and scalp electroencephalogram (EEG) datasets and proposed a generalized retrospective and patient-specific seizure prediction method. We use Short-Time Fourier Transform (STFT) on 30-second EEG windows with 50% overlapping to extract information in both frequency and time domains. A standardization step is then applied on STFT components across the whole frequency range to prevent high frequencies features being influenced by those at lower frequencies. A convolutional neural network model is used for both feature extraction and classification to separate preictal segments from interictal ones. The proposed approach achieves sensitivity of 81.4%, 81.2%, 82.3% and false prediction rate (FPR) of 0.06/h, 0.16/h, 0.22/h on Freiburg Hospital intracranial EEG (iEEG) dataset, Children's Hospital of Boston-MIT scalp EEG (sEEG) dataset, and Kaggle American Epilepsy Society Seizure Prediction Challenge's dataset, respectively. Our prediction method is also statistically better than an unspecific random predictor for most of patients in all three datasets.

Motivation & Objective

• Motivate a generalizable seizure prediction approach that reduces per-patient feature engineering.
• Demonstrate CNNs can learn from STFT-based image representations of EEG across datasets.
• Evaluate cross-dataset performance on Freiburg iEEG, CHB-MIT sEEG, and Kaggle seizure datasets.

Proposed method

• Convert 30-second EEG windows to STFT spectrogram-like matrices (frequency x time).
• Apply a simple CNN with three convolution blocks to extract features and classify preictal vs interictal segments.
• Balance training data by sliding a 30-second window over preictal signals to generate extra samples.
• Use batch normalization, ReLU, and max pooling across blocks, followed by two fully connected layers with dropout.
• Post-process predictions with a k-of-n rule to reduce isolated false positives.
• Evaluate using SPH=5 minutes and SOP=30 minutes under leave-one-out cross-validation per subject.

Experimental results

Research questions

• RQ1 Can a CNN learn discriminative features from STFT EEG representations across both iEEG and sEEG without patient-specific feature engineering?
• RQ2 What are the predictive performance and false alarm rates when applying the same CNN architecture and preprocessing across diverse EEG datasets?
• RQ3 Does the proposed approach achieve statistically significant prediction above a random predictor across datasets?

Key findings

• Sensitivity around 81% across Freiburg iEEG and CHB-MIT sEEG datasets with FPRs between 0.06 and 0.22 per hour.
• Kaggle dataset achieved 75% sensitivity with FPR of 0.21 per hour.
• CNN-based method generalizes to both intracranial and scalp EEG using STFT inputs with minimal feature engineering.
• Validation strategy and SPH/SOP definitions are clearly stated to enable robust benchmarking.
• The approach shows statistical significance over random predictors for most patients/datasets.