PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine Learning for Predicting Epileptic Seizures Using EEG Signals: A Review

Khansa Rasheed, Adnan Qayyum, Junaid Qadir, Shobi Sivathamboo, Patrick Kwan, Levin Kuhlmann, Terence O'Brien, Adeel Razi

arXiv:2002.01925v1cs.LGeess.SPq-bio.NC

TL;DR

Epileptic seizures are recurrent and unpredictable, making early prediction clinically important, but progress is constrained by limited data and unresolved methodological challenges. This paper reviews ML and DL techniques using EEG signals, surveying prediction workflows, reported results, pitfalls, and open research directions. The reviewed literature includes strong reported performance in some studies, while generalization, reproducibility, data quality, and long-term open EEG availability remain concerns.

  • Problem

    Seizure prediction remains difficult because labeled and long-term EEG data are limited, despite the need to anticipate adverse seizure consequences.

  • Method

    The paper comprehensively reviews ML and DL approaches for EEG-based seizure prediction, including preprocessing, feature extraction, classification, evaluation, and research challenges.

  • Results

    94% sensitivity and 0.11 FPR on average were reported for an SVM using permutation-entropy features on Freiburg hospital data.

  • Takeaways & Limitations

    The review highlights future work requiring further investigation across feature selection, prediction techniques, evaluation methodologies, and open research problems.

Abstract

from arXiv · show

With the advancement in artificial intelligence (AI) and machine learning (ML) techniques, researchers are striving towards employing these techniques for advancing clinical practice. One of the key objectives in healthcare is the early detection and prediction of disease to timely provide preventive interventions. This is especially the case for epilepsy, which is characterized by recurrent and unpredictable seizures. Patients can be relieved from the adverse consequences of epileptic seizures if it could somehow be predicted in advance. Despite decades of research, seizure prediction remains an unsolved problem. This is likely to remain at least partly because of the inadequate amount of data to resolve the problem. There have been exciting new developments in ML-based algorithms that have the potential to deliver a paradigm shift in the early and accurate prediction of epileptic seizures. Here we provide a comprehensive review of state-of-the-art ML techniques in early prediction of seizures using EEG signals. We will identify the gaps, challenges, and pitfalls in the current research and recommend future directions.

I. INTRODUCTION

Epilepsy involves recurrent, difficult-to-predict seizures whose consequences motivate early prediction from EEG. The paper reviews ML methods, identifies research gaps and pitfalls, and outlines future directions.

  • Epilepsy is a neurological disorder involving recurrent seizures, while some seizures arise from acute insults without indicating epilepsy.
  • Seizures are difficult to predict, and their severity and duration are uncertain, creating injury and safety concerns for patients and families.
  • EEG-based seizure prediction distinguishes pre-ictal from interictal brain states and requires analysis of recordings that may span days, weeks, or months.
  • Machine learning and deep learning address the substantial human effort required to examine long-duration EEG recordings and learn patterns from large datasets.
  • The review emphasizes ML advances for epilepsy prediction, introduces relevant neuroscience and brain-study tools, and considers their use for prediction.
  • Unlike earlier surveys, this review focuses in depth on ML methods and compares its coverage with existing surveys.

II. BACKGROUND ON EEG, NEUROSCIENCE, AND EPILEPSY PREDICTION

The paper introduces neuroscience, neuroimaging, EEG, and epilepsy prediction, emphasizing EEG’s temporal advantages, lower cost, and role in ML-based analysis.

  • Neuroscience integrates disciplines studying the brain’s structure, chemistry, electrical properties, and cognitive functions.
  • Neuroimaging includes structural and functional approaches that examine brain anatomy, activity, tissue diffusion, metabolism, and pathology.
  • EEG records cerebral-cortex electrical activity, offering higher temporal insight but lower spatial resolution than functional MRI.
  • EEG prediction workflows extract features or spectral information using Fourier or wavelet transforms before training ML classifiers, while DL can automate feature extraction.
  • EEG is attractive for seizure-prediction research because it tracks brain changes, costs less hardware, and supports recordings over longer durations.

2) Analysis Techniques:

EEG analysis methods are organized into time, frequency, time-frequency, and nonlinear domains, each targeting different signal characteristics and limitations.

  • EEG analysis methods are classified into time-domain, frequency-domain, time-frequency, and linear or nonlinear approaches.
  • Time domain methods: Time-domain methods use approaches such as PCA, LDA, and ICA to summarize, reduce, or decompose high-dimensional EEG data and remove artifacts.
  • Frequency domain methods: Frequency-domain methods measure seizure-related frequency changes using Fourier-based power-spectrum estimation, including Welch, MA, AR, and ARMA methods.
  • Time-frequency domain methods: Wavelet transform addresses time-domain and frequency-domain limitations by producing multiresolution sub-band signals through filter banks.
  • Non-linear methods: Nonlinear methods characterize harmonic coupling and provide epilepsy-classification features through higher-order spectra, entropy measures, and Lyapunov-related parameters.

C. EEG Signal Analysis for Epilepsy

EEG analysis supports epilepsy seizure identification by distinguishing seizure-related states, but prediction methods have evolved amid concerns about reproducibility and algorithm comparison.

  • EEG states: EEG recordings help distinguish epileptic seizures from non-seizure activity and identify pre-, ictal, interictal, and post-ictal periods.The pre-ictal state may provide predictive changes before onset, while ictal and interictal states describe seizure activity and intervals between seizures.
  • EEG analysis: Visual EEG examination is time-consuming and vulnerable to faulty detection because EEG data are dynamically nonlinear with high temporal and spatial complexity.Computerized parameter extraction and analysis are presented as potentially beneficial for diagnosis.
  • Validation challenges: Extensive and unseen EEG datasets challenged earlier findings based on selected, inadequate data because those results could not be reproduced.This motivated competitions using common datasets to standardize algorithm comparisons.
  • Validation challenges: Seizure-prediction competitions used shared datasets and AUC, but complex submitted algorithms made it difficult to determine which features or ML algorithms were best.One described challenge included 942 seizures recorded over more than 500 days, while another involved 1139 seizures.

III. ML APPROACHES FOR ES PREDICTION

This section reviews ML applications in healthcare and neuroscience, emphasizing how data-driven methods support analysis of complex biomedical and neuroimaging data.

  • Scope: The review examines ML-based methods for epileptic seizure prediction while highlighting ML’s potential in healthcare and neuroscience.It frames the discussion around applications of data-driven techniques to biomedical and brain data.
  • Healthcare applications: Large-scale biomedical data and advances in ML are enabling data-analysis approaches for developing medical tools.The passage notes that complex medical data make manual detection of representations impractical.
  • Neuroscience applications: ML and DL are presented as enabling technologies for processing and analyzing neuroimaging data and exploring theories of brain functioning.The section introduces supervised, unsupervised, and reinforcement learning in neuroscience.
  • Supervised learning: Supervised learning trains on expert-labeled data to learn relations between inputs and labels for classifying unseen data.The passage illustrates this process using labeled images and their categories.

2) Unsupervised Learning:

Unsupervised learning uses unlabeled EEG data, while seizure-prediction pipelines process noisy recordings, extract features, and classify pre-ictal and interictal patterns.

  • Unsupervised Learning: Unsupervised learning trains algorithms using unclassified or unlabeled data rather than expert-provided labels.The section motivates this approach through the brain’s formation of working models from repeated information.
  • Unsupervised Learning: Reinforcement learning develops policies that maximize rewards from interactions between an agent and its environment.Its central factors include a policy, reward signal, value function, and environment model.
  • EEG preprocessing: EEG recordings require noise reduction because eye movements, blinks, cardiac signals, and muscle activity contaminate brain-activity measurements.Filtering and noise-reduction methods are used before identifying pre-ictal and interictal stages.
  • Prediction pipeline: Classical ML uses processed EEG features for pattern identification, whereas DL can learn patterns directly from raw or minimally processed data.The distinction concerns whether feature extraction is performed before model training.
  • EEG preprocessing: Filtering artifacts and normalizing EEG data reduce their influence during feature extraction and improve comparability across patients.Examples include band-pass, wavelet, finite impulse response, and adaptive filters; corrupted data may also require handling.
  • Feature Extraction and Selection: Bivariate measures captured pre-ictal changes at least 240 min before seizure onset, compared with 5–30 min for univariate measures.Linear measures performed better or similarly to nonlinear measures in the cited comparison.

3) Classification:

Classification methods identify pre-ictal and interictal EEG patterns using engineered features and ML algorithms, with reported successes accompanied by concerns about evaluation and generalization.

  • Classification: ML classifiers such as ANNs, k-means, decision trees, SVMs, and fuzzy logic identify pre-ictal and interictal patterns from EEG data.Many studies use feature-value thresholds, while ML research emphasizes extracting optimized predictive features.
  • Evaluation: One study evaluated bispectral entropy features using F1 score and p-values but omitted sensitivity and specificity for a defined seizure prediction horizon and occurrence period.These measures are commonly preferred for seizure-prediction performance analysis.
  • Feature-based classification: Permutation entropy features combined with an RBF-kernel SVM achieved 94% sensitivity and 0.11 FPR on average with a mean SPH of 61 min.The study analyzed 83 seizures from 19 patients using 5 sec feature segments from Freiburg iEEG data.
  • Generalization: Training on randomly selected small EEG subsets can overfit because pre-ictal and interictal patterns vary across patients and within patients.Such models may not show significant performance in real-time scenarios.

D. Introduction to DL For Healthcare

Deep learning models process data through multiple nonlinear layers that learn useful representations, offering automated feature extraction for EEG-based seizure prediction. The reviewed approaches include CNNs, wearable implementations, and methods addressing patient variability and limited samples.

  • D. Introduction to DL For Healthcare: Deep learning models use layered nonlinear modules to extract relevant features and produce outputs from raw data.Higher layers amplify input features significant for tasks such as classification.
  • D. Introduction to DL For Healthcare: CNNs are widely applied to EEG because they can reduce noise and detect intricate patterns in high-dimensional data.EEG is a high-dimensional signal domain where automatic feature learning is useful.
  • D. Introduction to DL For Healthcare: Handcrafted-feature ML methods can achieve good accuracy, but noise, artifacts, and limited samples hinder generalized automatic seizure prediction.Deep learning instead learns more distinguishing and robust features automatically.
  • D. Introduction to DL For Healthcare: CNN-based methods transform raw EEG into two-dimensional STFT representations for learning and classifying pre-ictal versus interictal states.The approach was trained and tested on Freiburg hospital iEEG and CHBMIT scalp EEG databases.

2) Unsupervised DL Method for ES Prediction:

Unsupervised and recurrent deep-learning approaches address limited labeled EEG data and investigate architectures for seizure prediction. Reported results include strong AUC and sensitivity values, while raw-EEG prediction without extensive feature engineering remains open.

  • 2) Unsupervised DL Method for ES Prediction:: A GAN-based unsupervised method addresses limited labeled data by learning features from STFT spectrograms without manual feature extraction.The trained discriminator supplies features for seizure prediction.
  • 2) Unsupervised DL Method for ES Prediction:: With a 5-minute SPH and 30-minute SOP, the GAN approach achieved 77.68% AUC on CHBMIT, 75.47% on Freiburg, and 65.05% on EPILEPSIAE.These results were compared with supervised model-training methods.
  • 3) Use of RNN for ES Prediction:: LSTM-3, a two-layer architecture with 128 memory units per layer, performed best among the three evaluated LSTM architectures.The models used EEG feature vectors from time, frequency, and graph-theoretic measures.
  • 3) Use of RNN for ES Prediction:: LSTM-3 achieved average sensitivities of 99.28%, 99.35%, 99.63%, and 99.84% for 15-, 30-, 60-, and 120-minute pre-ictal periods, respectively.The study also compared raw EEG input with engineered feature vectors.
  • 2) Unsupervised DL Method for ES Prediction:: Public EEG databases from hospitals and research centers support development and evaluation of ML and DL seizure-prediction models.Examples include databases from Children’s Hospital Boston and Temple University Hospital.

H. Evaluation Metrics for ES Prediction

ES prediction evaluation combines confusion-matrix metrics with timing measures that describe when an alarm occurs relative to the seizure occurrence period. Sensitivity captures correctly predicted seizures, while SPH and SOP define the clinically relevant prediction window.

  • H. Evaluation Metrics for ES Prediction: Sensitivity is the ratio of correctly predicted seizures to all seizures, while false prediction rate captures incorrect alarms.These metrics were proposed as performance parameters for ES predictors.
  • H. Evaluation Metrics for ES Prediction: Accuracy is computed from true positives and true negatives divided by all positive and negative prediction outcomes.The displayed formula uses TP, TN, FP, and FN.
  • H. Evaluation Metrics for ES Prediction: True positives are correctly predicted seizures, true negatives are correctly predicted no-seizure periods, and false negatives are missed seizures.The supplied definitions distinguish correct predictions from missed events.
  • H. Evaluation Metrics for ES Prediction: False positives are non-seizure periods incorrectly predicted as seizures.False positives contribute to the false prediction burden.
  • H. Evaluation Metrics for ES Prediction: SOP is the interval in which a seizure may occur, whereas SPH is the time between the alarm and the SOP’s beginning.For precise prediction, seizure onset should follow SPH and occur within SOP.

IV. PITFALLS OF EPILEPTIC SEIZURE PREDICTION METHODS

ML has reduced manual EEG-analysis burdens, but seizure prediction remains constrained by data availability, missing observations, computational demands, hardware costs, and unresolved implementation needs.

  • IV. PITFALLS OF EPILEPTIC SEIZURE PREDICTION METHODS: Deep learning can learn features from raw or minimally processed data, reducing reliance on manual and time-consuming analysis.The review notes that automatically learned features may be more effective than analytically extracted features.
  • IV. PITFALLS OF EPILEPTIC SEIZURE PREDICTION METHODS: Long-term EEG data and reproducible code remain urgently needed because a prior 91-minute prediction result has not been reproduced on public datasets.The review identifies limited access to long-duration recordings as a core research problem.
  • IV. PITFALLS OF EPILEPTIC SEIZURE PREDICTION METHODS: Missing observations from communication failures or limited storage can lower prediction performance and motivate missingness indicators in models.The review specifically discusses zero or near-zero recorded values as a form of data corruption or dropout.
  • IV. PITFALLS OF EPILEPTIC SEIZURE PREDICTION METHODS: Excessive feature extraction increases time and computational cost, while deep learning requires abundant data for effective prediction.Real-time systems therefore require fast prediction on comparatively low-power and inexpensive hardware.
  • IV. PITFALLS OF EPILEPTIC SEIZURE PREDICTION METHODS: Cost-effective and efficient hardware implementation remains a bottleneck despite initial implementation efforts.The review states that substantially more work is needed to manufacture devices with optimized results.

A. Curse of Data Dimensionality

The review highlights dimensionality, annotation, data-sharing, and interpretability challenges in EEG-based seizure prediction, and surveys future directions addressing these issues.

  • A. Curse of Data Dimensionality: Multi-electrode EEG recordings increase signal dimensionality and make multi-channel analysis difficult.Suggested strategies include converting multi-channel data into a single-channel representation or using single-channel EEG.
  • B. Data Annotation: Supervised seizure prediction requires expert-labeled EEG data, but annotation quality and large-scale labeling are costly, time-consuming, and difficult to sustain.The review also emphasizes the importance of a true validation set for evaluating trained models.
  • B. Data Annotation: Automated labeling approaches such as active learning, generative models, and unsupervised clustering are proposed to address infeasible large-scale EEG annotation.The review cites GAN-based seizure prediction as an example of generative-model-assisted labeling.
  • B. Data Annotation: Efficient ML/DL development requires sharing distributed patient data across departments and hospitals and integrating data from different domains.The review also points to recurrent models and NLP for extracting knowledge from clinical notes.
  • B. Data Annotation: Deep-learning models can lack interpretability and make uncertainty quantification difficult, which is especially consequential in life-critical healthcare applications.The review therefore identifies explainability and interpretability as open requirements for seizure-prediction systems.
  • VI. CONCLUSIONS: The review synthesizes feature selection, prediction techniques, evaluation methodologies, and future research directions across the seizure-prediction literature.Its scope extends beyond surveys focused mainly on EEG analysis or limited prediction developments.
Loading 2002.01925v1…