Message in a Bottleneck: Interpretable Deep Learning for Dynamic Effective Brain Connectivity.

Deep learning (DL) approaches find increasingly more applications in medical imaging analysis, where they excel at predictive tasks such as diagnosis classification, but often offer little insight on the underlying mechanisms governing complex systems, a key goal of scientific inquiry. One way to bridge this gap is to design DL architectures for modeling the underlying system reflected in imaging data, which can give insights into the system through the analysis of fitted models. In this spirit, we present a novel approach for deep extraction of causally informed features via restricted architecture (DECIFRA), designed for multivariate time series modeling. DECIFRA introduces an interpretable bottleneck in its computation flow, where dynamically generated transition matrices exclusively model the directed influences between channels over time. These matrices function as both the model's internal computational mechanism and its primary, interpretable output that gives insights into the cross-channel interactions. To validate our method, we apply it to the challenging task of learning dynamic brain connectivity from functional magnetic resonance imaging (fMRI) data. We demonstrate that DECIFRA learns stable, reproducible connectivity patterns that are discriminative for group-level analysis. Moreover, these learned dynamics can be fine-tuned for specific classification tasks, improving predictive performance while retaining interpretability. Our findings show that DECIFRA captures functionally grounded brain connectivity patterns, with certain dynamics associated with age or schizophrenia classification tasks and others consistently preserved across tasks.