Generalizable Machine Learning Methods for Electrophysiology Open Access

Nasiri Ghosheh Bolagh, Samaneh (Summer 2020)

Permanent URL:


Brain pathology is increasingly recognized as a crucial factor in many illnesses. As the availability of low-cost brain monitoring devices important, the volume of data continues to expand. The need for automated brain monitoring diagnostics is, therefore, more acute, particularly in low resource regions of the world. The ground truth for brain monitoring remains the multi-lead electroencephalogram (EEG) and standard practice is still focused on visual inspection of EEGs. However, this is a costly and time-consuming procedure. Moreover, the lack of significant public databases (of 10,000-100,000 patients) of heterogeneous populations has limited the development of verifiable algorithms that generalize well across populations. Due to the characteristics and complexities of EEG signals, accurate interpretation of EEG signals by human experts requires several years of training. 


Developing accurate classifiers with high generalizability on other datasets is a challenging task in this area. Due to the non-stationary nature of the EEG signal, the statistical characteristics of the signal vary with time; therefore, a classifier that is trained on a temporally-limited amount of data from an individual may poorly generalize on EEG data recorded at a different time on the same subject. Another issue with the low generalizability issue in EEG data is related to high inherent inter-subject variability in the way an EEG manifests, which limits the usefulness of EEG applications. This phenomenon arises due to physiological differences (e.g. skull shape) between individuals, and neural activity does not propagate in a similar manner in different subjects. In particular, cortical folding, tissue conductivity, and tissue shapes of brains are different across people. Moreover, electrode sensor montages (the points at which the electrodes are attached and the references points) may differ and different manufacturers' acquisition hardware may filter the EEG differently. Finally, when electrodes are applied, small differences in the locations on the skull may exist, reflecting the EEG technicians' variety of training or even attentiveness on a given day. All these factors lead to significant variabilities in EEG signals, which lead to different joint distributions between the feature and label space of different recordings. Therefore, the transferability of the trained model on unseen subjects is degraded. The reason behind this problem is the assumption in machine learning techniques that training and test data should be drawn from the same distribution, an assumption that does not necessarily hold in large biomedical datasets. 


  This thesis has addressed this challenge via two approaches: 1) measuring the boosting effect of machine learning and deep learning methods in a non-Euclidean space to mitigate the effects of intra and inter-subject variability in seizure detection and sleep staging; 2) developing adversarial networks with attention mechanisms and importance weighting to learn both transferable and discriminative representations, and enhance the generalizability of the model for classifying sleep stages, and seizure detection.

Table of Contents

Introduction, Boosting Automated Sleep Stage and Seizure detection Performance in Big Datasets using Population Sub-grouping, Attentive Adversarial Network for Large-Scale Sleep Staging, Importance Weighting with Adversarial Network for Large-Scale Sleep Staging, Cross-Subject Seizure Detection using Generating Transferable Adversarial Features, Conclusion

About this Dissertation

Rights statement
  • Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.
  • English
Research Field
Committee Chair / Thesis Advisor
Committee Members
Last modified

Primary PDF

Supplemental Files