Author :

Labonno Akter Mim

Co-Author :

Imranul Hoque Rizvee, Htet Thet, Jannatul Naima, Natasha Parker-Rauscher

Mentor:

Prof. Dr. Maryam Vatankhah

Abstract

Chronic neuropathic pain remains difficult to measure objectively, as clinical assessments rely heavily on subjective patient reports. Electroencephalography (EEG) offers a promising non-invasive approach for detecting neural signatures associated with pain. However, classification accuracy alone does not ensure that extracted EEG features are stable or generalizable across datasets. In this study, EEG data were processed using a structured pipeline in Python with the MNE framework. Power spectral density (PSD) features across standard EEG frequency bands were extracted and used to train Support Vector Machine (SVM) and Random Forest classifiers. Models were evaluated on two separate datasets and a combined dataset to assess both performance and feature stability. For Dataset 1, SVM achieved 62.5% accuracy and Random Forest achieved 68.8%. For Dataset 2, SVM achieved approximately 69% accuracy (ROC-AUC 0.753), while Random Forest achieved approximately 67% (ROC-AUC 0.733). On the combined dataset (188 samples), both models achieved an accuracy of 0.76, with SVM achieving a ROC-AUC of approximately 0.78 and Random Forest achieving approximately 0.87. Although accuracy was identical on the combined dataset, the higher ROC-AUC of Random Forest indicates better class separability. The consistency of results across datasets suggests that the extracted EEG features are stable and generalizable across subjects.

Introduction

Chronic neuropathic pain significantly impacts patient quality of life and remains challenging to diagnose objectively. Current clinical assessments rely heavily on subjective pain reports, which can vary widely across individuals. Electroencephalography (EEG) has emerged as a promising non-invasive technique for detecting neural patterns associated with pain perception. Recent studies have demonstrated that machine learning models can classify EEG signals related to pain and non-pain conditions. However, classification accuracy alone does not ensure that detected features are reliable or reproducible. Identifying stable EEG biomarkers that remain consistent across machine learning models and preprocessing methods is essential for developing robust EEG-based pain detection systems.

Aim

The goal of this study is to evaluate whether EEG features associated with chronic neuropathic pain remain stable and reproducible across datasets and machine learning models, while also assessing classification performance using multiple evaluation metrics.

Methodology

Data Processing

EEG recordings were processed using the MNE Python framework, including filtering, artifact removal, and quality control.

Feature Extraction

Spectral features were extracted using Power Spectral Density (PSD) across standard EEG bands: Delta, Theta, Alpha, Beta, and Gamma.

Machine Learning Models

Two classifiers were evaluated:

• Support Vector Machine (SVM, RBF kernel)

• Random Forest

Evaluation

Dataset: 188 samples (94 pain, 94 non-pain), 80/20 train-test split

Metrics: Accuracy, Precision, Recall, F1-score, ROC-AUC

Models were tested on Dataset 1, Dataset 2, and a combined dataset to assess generalizability.

Results

Dataset 1:
• SVM Accuracy: 62.5%
• Random Forest Accuracy: 68.8%

Dataset 2:
• SVM Accuracy: ~69% (ROC-AUC: 0.753)
• Random Forest Accuracy: ~67% (ROC-AUC: 0.733)

Combined Dataset:
• Accuracy (both models): 0.76
• SVM ROC-AUC: ~0.78
• Random Forest ROC-AUC: ~0.87

Although both models achieved identical accuracy, Random Forest demonstrated stronger class separation based on ROC-AUC.

Discussion & Conclusion

Both machine learning models achieved comparable classification accuracy across datasets. However, the Random Forest model demonstrated improved discriminative capability based on ROC-AUC. These findings indicate that relying solely on classification accuracy may not fully capture model performance. Evaluating additional metrics such as ROC-AUC provides a more comprehensive understanding of model reliability. Importantly, the consistency of results across multiple datasets suggests that the extracted EEG features are stable and generalizable across subjects.

References

Mari T. et al., Scientific Reports, 2023
Zolezzi D. et al., Mendeley Data, Chronic Neuropathic Pain EEG Dataset
Tyler Mari et al., 2023 — External validation of machine learning models for pain classification

Acknowledgment

This work is supported by the CSTEP (Collegiate Science and Technology Entry Program). Special thanks to Prof. Dr. Maryam Vatankhah for mentorship and guidance.