Enhanced Detection of Epileptic Seizure Using Supervised and Unsupervised Algorithms

Ananthakrishnan Gopalakrishnan, Sharon Priya Surendran, Aisha Banu Wahab, Aarthi Gopalakrishnan, Yogesh Kumar BalajiAci@icaids 2023

Introduction

Epilepsy is a chronic neurological disorder characterized by recurrent, unprovoked seizures caused by abnormal electrical activity in the brain. Currently, the clinical "gold standard" for diagnosis involves medical experts manually reviewing Electroencephalogram (EEG) waveforms. However, this manual process is inherently time-consuming, highly subjective, and prone to human error due to the non-stationary and noisy nature of EEG signals. To address these challenges, this study investigates the development of an automated machine learning pipeline capable of accurately categorizing EEG data to assist in clinical diagnosis.

Methodology

The researchers implemented a multi-stage pipeline designed to transform raw brainwave data into actionable classifications. The process begins with data collection and proceeds through rigorous feature engineering and model building phases.

The multi-stage machine learning pipeline for EEG signal classification.

Feature Extraction and Preprocessing

Raw EEG signals are complex and difficult to interpret directly. The team utilized Discrete Wavelet Transform (DWT) to decompose these signals into specific frequency components, such as alpha, beta, and gamma waves. From these components, statistical features like standard deviation were calculated to create a functional feature space. Preprocessing steps included:

  • Missing Value Imputation: Ensuring data integrity.
  • Handling Categorical Data: Converting non-numeric identifiers for model compatibility.
  • Feature Scaling: Normalizing the magnitudes of various brainwave readings to prevent algorithmic bias.

Classification Strategy

The study compared four distinct machine learning approaches to determine the most effective method for seizure detection:

  1. Supervised Learning: Support Vector Machine (SVM), Random Forest, and K-Nearest Neighbor (KNN).
  2. Unsupervised Learning: K-means Clustering.

The dataset, sourced from Kaggle, consisted of 2,216 rows and 668 columns of numerical EEG readings from specific electrodes (FP1, FP2, F3, F4, T5, and T6). The data was split into an 80:20 ratio for training and validation.

Key Results and Discussion

The performance metrics revealed a surprising and significant advantage for unsupervised learning methods in this specific medical context.

Algorithm Accuracy (%) Precision (%) Recall (%) F1-Score (%)
K-means 98.1 - - -
Random Forest 83.5 83.6 83.5 83.5
SVM 83.5 70.8 54.2 40.3
KNN 80.0 ~80.0 ~80.0 ~80.0

The Unsupervised Advantage

The most striking finding was that K-means clustering achieved a superior accuracy of 98.1%, far outperforming the best supervised model (Random Forest at 83.5%). This suggests that the inherent structure of seizure-related EEG data is so distinct that clustering algorithms can identify the "seizure state" more effectively than classifiers attempting to map features to pre-defined labels.

The SVM Performance Gap

While the SVM model achieved a respectable accuracy of 83.5%, its recall was notably low at 54.2%. In a medical diagnostic setting, low recall is particularly concerning as it indicates a high rate of false negatives—missing nearly half of the actual seizure events.

Comparison of recall values across supervised learning models, highlighting the significant performance gap in the SVM algorithm.

Conclusion and Future Work

This research demonstrates that automated systems, particularly unsupervised clustering, can provide highly accurate detection of epileptic seizures from EEG data. By utilizing DWT for feature extraction, the researchers created a robust foundation for machine learning application.

Future work will focus on transitioning toward deep learning architectures, such as Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs), which may eliminate the need for manual feature engineering. Additionally, validating these models against larger, real-world clinical datasets containing more varied seizure types and environmental artifacts is essential for moving toward clinical implementation.

Summary updated

BibTeX
@inproceedings{gopalakrishnan2023enhanced,
  author = {Ananthakrishnan Gopalakrishnan and Sharon Priya Surendran and Aisha Banu Wahab and Aarthi Gopalakrishnan and Yogesh Kumar Balaji},
  title = {{Enhanced Detection of Epileptic Seizure Using Supervised and Unsupervised Algorithms}},
  booktitle = {Aci@icaids},
  year = {2023},
}
Figure