Strategies for Automatic IPPM Generation

Anirudh Lakra¹Cai Wingfield^2,3Chao Zhang^1,4,5*Andrew Thwaites^1,3*

• ¹University College London, London, United Kingdom
• ²University of Birmingham, Birmingham, England, United Kingdom
• ³MRC Cognition and Brain Sciences Unit, School of Clinical Medicine, University of Cambridge, Cambridge, England, United Kingdom
• ⁴Tsinghua University, Beijing, Beijing, China
• ⁵Shanghai Artificial Intelligence Laboratory, Shanghai, China

Information Processing Pathway Maps (IPPMs) are a concise way to represent the evidence for the transformation of information as it travels around the brain. However, their construction currently relies on hand-drawn maps from electrophysical recordings such as magnetoencephalography (MEG) and electroencephalography (EEG). This is both inefficient and contains an element of subjectivity. A better approach would be to automatically generate IPPMs from the data and objectively evaluate their accuracy. In this work, we propose a range of possible strategies and compare them to select the best. To this end, we a) provide a test dataset against which automatic IPPM creation procedures can be evaluated; b) suggest two novel evaluation metrics—causality violation and transform recall—from which these proposed procedures can be evaluated; c) conduct a simulation study to evaluate how well ground-truth IPPMs can be recovered using the automatic procedure; and d) propose and evaluate a selection of different IPPM creation procedures. Our results suggest that the max pooling approach gives the best results on these metrics. We conclude with a discussion of the limitations of this framework, and possible future directions.