نهمین کنگره نقشه برداری مغز

نمایش مقاله

Title:

Explainable Machine Learning with Hjorth Parameters for Epileptic Seizure Detection in Bonn EEG Dataset

Subject:

Brain Mapping Applications in Epilepsy – Across Pediatric and Adult Populations

Presenter Author:
Name:	Abolfazl Hajihashemi

Author:
Name 1:	Abolfazl Hajihashemi - Corresponding - Presenter
Affiliation 1:	Department of Biomedical Engineering, Faculty of Engineering, University of Isfahan, Isfahan, Iran

Name 2:	Mahdieh Hajihashemi
Affiliation 2:	Department of Computer Engineering, Hashemi Technical and Vocational College for Women, Khansar, Iran

Name 3:
Affiliation 3:

Name 4:
Affiliation 4:

Name 5:
Affiliation 5:

Name 6:
Affiliation 6:

Name 7:
Affiliation 7:

Name 8:
Affiliation 8:

Name 9:
Affiliation 9:

Name 10:
Affiliation 10:

Abstract

Original Research

introduction :

Epilepsy is one of the most common neurological disorders worldwide, affecting more than 50 million individuals and characterized by recurrent, unprovoked seizures. These events frequently go undetected or are misdiagnosed due to the inherent limitations of manual electroencephalography (EEG) review, which is time-consuming, heavily dependent on expert experience, and susceptible to human error. Recent advances in machine learning have enabled automated seizure detection; however, the majority of existing models suffer from a lack of transparency, often described as "black-box" systems, which significantly impedes their clinical acceptance. The present study was designed to address this critical gap by developing a model that achieves high diagnostic accuracy while providing clear explanations for its decisions, thereby promoting greater trust among clinicians. For this purpose, the Bonn University EEG dataset a well-established benchmark in epilepsy research was employed. Hjorth parameters were incorporated due to their computational simplicity and direct relevance to key EEG signal characteristics, aiming to improve detection performance while maintaining low resource requirements suitable for deployment in constrained clinical environments.

Methods:

The Bonn dataset was sourced from the University of Bonn's public repository, featuring 500 EEG recordings sampled at 173.61 Hz and segmented into 23.6-second windows reshaped into 11,500 one-second samples. Classes included healthy (eyes open/closed), epileptic interictal (healthy/tumor areas), and ictal segments. Hjorth parameters were computed for each segment: activity as signal variance, mobility as the square root of the variance ratio of the first derivative to the signal, and complexity as the mobility ratio of the second to first derivative. These were appended to the raw 178 time-domain features. Data were normalized using StandardScaler and stratified-split: 70% training (8,049 samples), 15% validation (1,726), and 15% test (1,725). A Random Forest classifier was trained with 300 estimators, optimized via default hyperparameters for baseline performance. Evaluation metrics included accuracy, precision, recall, specificity, F1-score, and ROC-AUC. Explainability was achieved through SHAP, with summary plots visualizing feature importances on a 200-sample test subset.

Results:

The model demonstrated strong performance across splits. On the test set, accuracy reached 97.51%, with precision 93.90%, recall 93.62%, specificity 98.48%, and F1-score 93.76%. Validation results were similar (accuracy 97.45%, precision 92.63%, recall 94.78%, specificity 98.12%, F1 93.70%), suggesting minimal overfitting despite perfect training accuracy a phenomenon attributable to the dataset's inherent separability. The confusion matrix showed few false positives/negatives, with most errors in borderline interictal-ictal transitions. ROC-AUC was near-perfect at 0.99, and precision-recall curves confirmed balanced performance. SHAP analysis indicated that Hjorth mobility and complexity were among the top contributors, positively influencing ictal predictions by capturing signal irregularity, while activity had moderate impact.

Conclusion:

This explainable framework, incorporating Hjorth parameters and SHAP, delivered high diagnostic accuracy with transparent decision-making. It can alleviate clinician workload and support personalized epilepsy treatment. Limitations include dataset size and single-center origin. Future studies should validate on larger, multicenter data and explore real-time applications.

Keywords:

Epilepsy; Seizure Detection; Electroencephalography; Machine Learning; Explainable Artificial Intelligence; Hjorth Parameters