11/04/2025
By Danielle Fretwell

The Francis College of Engineering, Department of Electrical and Computer Engineering, invites you to attend a Doctoral Dissertation Proposal defense by Masoumeh Farhadi Nia on: "Machine Learning for Electroencephalography and Electrooculography Classification in Assistive Brain–Computer Interfaces."

Candidate Name: Masoumeh Farhadi Nia
Degree: Doctoral
Defense Date: Friday, November 7, 2025
Time: 2:30 – 4:30 p.m.
Location: Ball 313

Committee:

  • Advisor: Kelilah Wolkowicz, Assistant Professor; Mechanical and Industrial Engineering, Robotics; UMass Lowell
  • Co-Advisor: Kavitha Chandra, Assoc Dean - Undergrad Affairs, Electrical and Computer Engineering, UMass Lowell
  • Charles Thompson, Professor, Electrical and Computer Engineering, UMass Lowell
  • Farshid Alizadeh-Shabdiz, Professor of the Practice, Computer Science, Boston University

Abstract:
Brain–Computer Interfaces (BCIs) enable direct communication between the brain and external devices, offering new pathways for independent mobility among individuals with severe motor impairments. In the non-invasive modalities, Electroencephalography (EEG)–based Motor Imagery (MI) is particularly promising for wheelchair navigation, as it allows users to issue control commands through imagined movements rather than muscle activity. However, the practical reliability of EEG-based BCIs remains limited due to the low signal-to-noise ratio, inter- and intra-subject variability, and inconsistent classification accuracy across sessions. To address these challenges, this research investigates how repetition and real-time feedback, two promising yet understudied factors, affect EEG signal quality, user adaptation, and classification performance. Electrooculography (EOG) is also examined as a complementary modality for simple eye-based control, though the primary focus remains on EEG-based MI.

This dissertation will develop a robust EEG processing and classification framework that integrates the effects of repetition and feedback using the Graz Datasets. Graz Dataset B includes recordings from nine subjects, each participating in five sessions over two weeks. The first two sessions were recorded under open-loop conditions (without user feedback), and the remaining three under closed-loop conditions with real-time user feedback. In the open-loop sessions, each subject performed 60 trials per Motor Imagery (MI) class (i.e., imagined left-and right-hand motion), each lasting four seconds with short and randomized breaks. During the closed-loop sessions, participants completed four runs of twenty trials per class, receiving real-time smiley feedback that turned green for correct and red for incorrect MI classification. The proposed framework will compare open- and closed-loop paradigms and examine how repetition and feedback affect EEG signal quality, neural stability, and classification performance in MI-based BCIs.

The proposed methodology includes time-series data analysis, segmentation of motor imagery (MI) time slots, band-pass filtering within the Alpha (8–13 Hz) and Beta (13–30 Hz) bands, Independent Component Analysis (ICA) for artifact removal, and spatial filtering to isolate motor-related cortical activity. Feature extraction includes the use of Common Spatial Patterns (CSP) and its multi-band variant Filter Bank CSP (FBCSP), band power estimation within μ and β ranges, wavelet-based time–frequency decomposition, entropy-based complexity measures, and Riemannian covariance analysis. Feature extraction is followed by classification via Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), Random Forest (RF), Ensemble, and AdaBoost models. In parallel, deep learning architectures, including EEGNet, Temporal Convolutional Networks (TCN), and CNN–LSTM hybrids will be explored to further enhance accuracy and reliability.

Preliminary experiments on the Graz Dataset B show that traditional machine learning models achieve about 10% higher classification accuracy under feedback conditions compared to non-feedback sessions, indicating enhanced neural consistency and user engagement. These findings support the hypothesis that repetition and feedback improve both the reliability and safety of EEG-driven assistive navigation. Future work will extend this framework to the Graz Dataset A, enabling multi-directional (left, right, forward, and backward) control for advanced wheelchair navigation. It will also explore integrating Electrooculography (EOG) as a complementary modality for simple eye-based commands, such as stopping or selecting path directions to further enhance user safety and adaptability.