[Paper Review] Improved EEG Classification by factoring in sensor topography.

Lubna Shibly Mokatren, Rashid Ansari+5 more|arXiv (Cornell University)|May 22, 2019

EEG and Brain-Computer Interfaces3 citations

TL;DR

This paper proposes a novel EEG classification method that incorporates sensor topography by modeling spatial electrode configurations via interpolation, significantly improving classification accuracy. By representing EEG data as 2D spatial maps and using CNNs, the approach achieves 5–8% higher accuracy than conventional methods that ignore spatial layout, especially in social anxiety disorder and emotion recognition tasks.

ABSTRACT

Electroencephalography (EEG) serves as an effective diagnostic tool for mental disorders and neurological abnormalities. Enhanced analysis and classification of EEG signals can help improve detection performance. This work presents a new approach that seeks to exploit the knowledge of EEG sensor spatial configuration to achieve higher detection accuracy. Two classification models, one which ignores the configuration (model 1) and one that exploits it with different interpolation methods (model 2), are studied. The analysis is based on the information content of these signals represented in two different ways: concatenation of the channels of the frequency bands and an image-like 2D representation of the EEG channel locations. Performance of these models is examined on two tasks, social anxiety disorder (SAD) detection, and emotion recognition using DEAP dataset. Validity of our hypothesis that model 2 will significantly outperform model 1 is borne out in the results, with accuracy $5$--$8\%$ higher for model 2 for each machine learning algorithm we investigated. Convolutional Neural Networks (CNN) were found to provide much better performance than SVM and kNNs.

Motivation & Objective

• To improve EEG classification accuracy by leveraging the spatial arrangement of EEG sensors.
• To investigate whether incorporating sensor topography leads to better performance in mental state classification.
• To compare models that ignore versus exploit spatial sensor configuration in EEG signal analysis.
• To evaluate the effectiveness of different interpolation techniques for spatially representing EEG data.
• To determine the relative performance of machine learning models (SVM, kNN, CNN) when spatial information is included.

Proposed method

• The method models EEG signals using a 2D image-like representation based on the anatomical locations of EEG electrodes.
• Spatial interpolation techniques are applied to estimate signal values at non-measured locations, enhancing spatial resolution.
• Two models are compared: one ignoring sensor configuration (model 1) and another incorporating it via interpolated spatial maps (model 2).
• Frequency band signals are either concatenated across channels or transformed into 2D spatial feature maps for input to classifiers.
• Convolutional Neural Networks (CNNs) are used to exploit spatial patterns in the 2D representations, outperforming SVM and kNN.
• Performance is evaluated on two datasets: social anxiety disorder (SAD) and DEAP for emotion recognition.

Experimental results

Research questions

• RQ1 Does incorporating EEG sensor topography significantly improve classification accuracy in mental state detection?
• RQ2 How do different interpolation methods affect the representation of spatial EEG patterns?
• RQ3 Can CNNs effectively learn from 2D spatial representations of EEG signals compared to traditional classifiers?
• RQ4 Is the performance gain from spatial modeling consistent across different machine learning algorithms?
• RQ5 To what extent does spatial information enhance detection of social anxiety disorder and emotional states?

Key findings

• Model 2, which incorporates sensor topography via spatial interpolation, achieved 5–8% higher classification accuracy than model 1 across all tested machine learning algorithms.
• Convolutional Neural Networks (CNNs) outperformed both SVM and kNN, demonstrating superior ability to exploit spatial patterns in EEG data.
• The 2D spatial representation of EEG channels significantly enhanced information content compared to simple concatenation of frequency bands.
• The performance gain from including spatial configuration was consistent across both the SAD and DEAP datasets.
• The use of interpolation methods improved the spatial fidelity of EEG signal representation, leading to more discriminative features for classification.
• The results validate the hypothesis that spatial sensor configuration is a critical factor in EEG classification and should not be ignored.