Exploring Novel Methodological Approaches for the Analysis of Electroencephalogram Data: Machine Learning and Group Iterative Multiple Model Estimation (2024)

Undergraduate: Suryadyuti Baral

Faculty Advisor: Joseph B. Hopfinger; Kathleen M. Gates
Department: Psychology and Neuroscience

---

Machine learning, particularly Support Vector Machines (SVM), has been widely utilized in Electroencephalogram (EEG) research to identify significant information within time points related to stimuli. However, alternative models may offer comparable performance. In the present study, we re-analyzed datasets from experiments by Bae & Luck (2018, 2019) using K-Nearest Neighbors, Naïve Bayes, Random Forest and Adaptive Boosting. We address discrepancies between behavioral outcomes and decoding results in Bae and Luck’s motion perception study from 2019. We replicated the results from Bae and Luck's 2018 study using the other algorithms and explored the possibility of one model outperforming the others. While no single model emerged superior based solely on decoding accuracy, converging results from multiple models helped reconcile discrepancies in the 2019 experiment. Additionally, for Bae and Luck’s 2018 study, we explored neural mechanisms underlying significant time point clusters identified by our machine learning algorithms. Implementing Group Iterative Multiple Model Estimation (GIMME) in EEG research for the first time, we probed differences in connectivity between spatial attention and working memory using the paradigm. Our findings suggest that GIMME complements traditional machine learning models, uncovering connectivity differences not apparent in decoding results. However, challenges remain in establishing a proper measurement model and addressing estimation singularities. Successful resolution of these issues in future research will lead to a more robust exploration of effective neural connectivity using EEG.