CAREER: III: Trust-EEG: A Trustworthy Machine Learning Framework for Augmenting Clinical Review of Electroencephalograms

University Of Minnesota-Twin Cities, Minneapolis MN

Investigators

Abstract

Expert visual review of patient data is widespread in healthcare, which not only contributes to physician burnout but also introduces errors in clinical decisions. This is particularly emphasized in neurology where experts spend a substantial amount of time visually reviewing lengthy multi-channel time series of brain activity, called electroencephalography (EEG). Machine learning (ML) has emerged as a potential solution to ease this burden and create reliable and scalable solutions. However, most EEG ML models, which are based on supervised learning, do not yield meaningful EEG features because of labeling inconsistencies. In addition, these models have not been rigorously tested in out-of-sample settings and therefore can exhibit performance deficits during deployment leading to incorrect diagnoses or decisions. As such, there is a compelling need to develop more reliable, reproducible, and robust ML approaches for EEG review. The goal of this proposal is to develop a trustworthy ML framework to augment clinical EEG review and demonstrate its utility in real-world clinical applications. Our research will significantly improve the diagnostic capabilities of EEG while reducing physician workload. We will demonstrate the framework’s ability to augment EEG review by working closely with domain experts at the Mayo Clinic and Cleveland Clinic. We will also enable research opportunities for undergraduate and K-12 students, especially underrepresented minorities, and engage students with epilepsy in focused research projects. Finally, we will leverage the outcomes of this research to develop courses in engineering and medicine. This research will develop a suite of novel ML methods to realize a trustworthy ML framework to augment EEG review. We will undertake the following strategies to ensure trust in EEG ML: a) developing domain-guided backbone architectures for EEG representation learning, b) leveraging self and weak supervision, instead of label-hungry and error-prone supervised learning, to scale up available training data, and c) performing model diagnostics to identify and rectify failure scenarios. The core of the proposed framework will be a domain-guided foundation model for EEG data that addresses the current limitations of EEG ML. Our proposed work includes a) development of an attention-based domain-guided architecture to capture EEG spatiotemporal dynamics; b) designing domain-guided self- and weak-supervision tasks to address labeled-data scarcity; c) development of model diagnostics and adversarially robust training to handle distribution shifts; and d) real-world validation of the framework in epilepsy subtype classification and treatment outcome prediction, and further evaluation in out-of-sample settings. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.