Multivariate Machine Learning Techniques for Data Preprocessing, Decomposition, and Integration: Radiomics, Sparse Partial Least Squares, and Transformers

In this talk, different machine learning techniques will be introduced that may help neuroscientific researchers in multiple ways ranging from data preprocessing to decomposition or integration. The focus will be the Sparse Partial Least Squares (SPLS) algorithm as well as its extension, the multi-block SPLS (mbSPLS). It will be shown how SPLS and mbSPLS can extract various layers of multivariate effects between any combination of two or more data matrices. Investigations into neuroinflammation-brain patterns as well as deeply layered genetic-brain-behavior signatures in the early-age PRONIA cohort will be given as specific examples. Furthermore, forays into hostility and neurodestructive processes in early and late-stage psychosis will be presented.
A smaller part of the talk will revolve around novel approaches that go beyond data analysis. Specifically, a brief introduction to radiomics will highlight the potential for machine learning techniques in extracting new and innovative features from MRI scans, such as texture, entropy, and contrast. Moreover, transformer-based pipelines will be shown that allow the researcher to integrate multiple data domains from different cohorts to answer specific research questions without any requirement for data or sample homogenization.
Overall, the talk will cover the fundamentals of these techniques, however, the focus will be on the application and translational value of these approaches

This talk is hosted online only. To join, please use the link below:
zoom.us/j/92620728590?pwd=s1JefrGff6bN0nZZcHSTBkCw8Z1RlT.1
Meeting ID: 926 2072 8590
Passcode: 196542