Interpretable machine learning models for EEG analysis
I have long-standing interests in understanding which brain regions are related to the emergence of psychiatric disorders and specific symptoms, such as auditory verbal hallucinations (AVH). A very promising way to identify potential neural correlates or "fingerprints" is offered by machine learning models. Although a plethora of algorithms were proposed in the last decade, most of them struggled when applied to real-world data. Besides, interpretability is often omitted from the consideration which does not allow to gain insights from data and hence improve knowledge about the system of interest. Combined with poor generalization, it leads to the rare application of machine learning to clinical decision support. We address these problems by using a framework of generative modelling to develop an algorithm that is able to a) recover neurological mechanisms underlying psychiatric disorders from neuroimaging data in a data-driven manner; b) differentiate between multiple psychiatric disorders and/or symptoms; and c) generalize to multiple brain imaging modalities. In this project, we have made a step towards an in-depth examination of functional brain connectivity during rest and dichotic listening in healthy controls and patients with schizophrenia who do or do not have AVH. Our method demonstrates high classification performance and learns AVH-related mechanisms that are consistent with current state-of-the-art knowledge and theory. This project is in collaboration with Nico Hoffmann (HZDR) and Maksim Zhdanov (Technical University Dresden).
Maksim Zhdanov, Steinmann S*, Nico Hoffmann* (2022). Learning Generative Factors of Neuroimaging Data with Variational auto-encoders. arXiv preprint arXiv:2206.01939
Maksim Zhdanov, Steinmann S, Nico Hoffmann (2022). Investigating Brain Connectivity with Graph Neural Networks and GNNExplainer. arXiv preprint arXiv:2206.01930