Articles

IACM Colloquium

 

Speaker:

Dr. Antonios G. Dougalis

Independent Researcher

Formerly Head of the Laboratory for Human Cellular and Network Neurophysiology, University of Eastern Finland

 

Title: Interpretable Machine Learning and EEG Biomarkers in Neurodegenerative Disease

 

Abstract:

Machine learning approaches applied to electrophysiological recordings often prioritize classification accuracy while providing limited physiological insight into the neural mechanisms underlying disease states. This presents an important challenge for the development of clinically useful biomarkers in neurodegenerative disorders, where interpretability, robustness, and generalizability are essential for diagnosis, disease monitoring, and clinical trial applications.

In this presentation, I will discuss a framework for the analysis of resting-state EEG recordings based on physiologically interpretable representations of brain dynamics. Rather than relying on abstract raw signal descriptors with unclear biological significance, the approach integrates established electrophysiological measures linked to cortical and network function, including oscillatory spectral activity, phase synchronization, cross-frequency coupling, aperiodic dynamics, neuronal avalanche statistics, scale-free activity, and functional connectivity measures. These physiologically motivated features are combined with modern machine learning approaches, including attention-based transformer architectures, while preserving interpretability at multiple levels of the analysis. Using Parkinson's disease as the primary experimental model, I will show how complementary electrophysiological descriptors capture distinct aspects of disease pathology and medication-related neural state modulation. Beyond classification performance, I will discuss how feature importance analyses, channel-level representations, and attention mechanisms may contribute to physiologically meaningful biomarker discovery. More broadly, the work raises questions concerning how alterations in neuronal and network physiology become expressed in large-scale brain dynamics and how physiologically informed representations may improve the interpretability and clinical utility of machine learning approaches. 

Although the examples presented focus on Parkinson's disease, the underlying framework may have broader applications for biomarker discovery in neurodegenerative disorders and for understanding brain dynamics across multiple spatial and temporal scales.

 

Short Bio:

Antonios G. Dougalis is a neuroscientist and neurophysiologist whose research combines cellular neurophysiology, electrophysiology, computational neuroscience, signal analysis, and machine learning to investigate brain dynamics in health and disease. He received his Ph.D. 

in Cellular Neurophysiology and Neuropharmacology from the University of Sunderland, United Kingdom, in 2005 and has held research positions at several international institutions, including Imperial College London, the University of Manchester, the University of Ulm, Dublin City University, and the University of Eastern Finland, where he formerly led the Laboratory for Human Cellular and Network Neurophysiology. His research focuses on understanding how alterations in neuronal and microcircuit physiology become expressed at the level of large-scale brain dynamics. Current interests include electrophysiological biomarkers, neural connectivity, critical dynamics, multiscale brain organization, and physiologically interpretable machine learning approaches for neurodegenerative disorders. More broadly, his work aims to bridge cellular neurophysiology, systems neuroscience, and applied mathematics to understand how brain states emerge, evolve, and transition across multiple spatial and temporal scales.