INTRODUCTION: Neuroscience circuitry at any scale has always involved time series analysis or even causality, where we try to discover the influence that a brain region or a group of neurons exerts over time. This is particularly relevant if we want to understand diseases that involve the transfer of information.

AIM(S): We present an end-to-end AI framework for directed graphs, incorporating explainable AI techniques, aimed at modeling brain connectivity in stroke patients. Additionally, we explore the integration of time series analysis using foundation models inspired by large language models to enhance temporal dynamics understanding.

METHOD(S): Those machine learning pipelines combine different types of causal estimators, such as Granger causality, or reservoir computing, combined with directed graph analysis to highlight even more differences between healthy subjects and subjects with trauma or diseases. Directed graphs are constructed from these connectivity measures and classified using a directed graph convolutional network. Explainable AI tools are employed to interpret the disrupted brain networks and identify relevant biomarkers. The inclusion of foundation model-based time series analysis enhances the temporal resolution and robustness of the connectivity features. Ultimately, a general framework is proposed to discover causality from MRI, EEG time series or other types of data.

RESULTS: The proposed framework achieved a classification level where the reservoir computing was relatively superior to the traditional Granger Causality. Time series foundation models have also been shown to be an alternative tool for time series prediction.

CONCLUSIONS: This approach demonstrates the potential of combining reservoir computing, directed graph analysis, foundation model-driven time series analysis, and explainable AI to improve patient stratification in brain diseases. Our technical innovations advance the understanding of effective brain connectivity and pave the way for more interpretable AI-driven clinical tools.

FINANCIAL SUPPORT: This work was supported in part by the Polish High-Performance Computing Infrastructure PLGrid (HPC Centers: ACK Cyfronet AGH) under Grant PLG/2024/017108; in part by the project of the Minister of Science and Higher Education ‘‘Support for the Activity of Centers of Excellence established in Poland under Horizon 2020’’ under Contract MEiN/2023/DIR/3796; in part by the European Union’s Horizon 2020 Research and Innovation Program under Grant 857533; and in part by the International Research Agendas program of the Foundation for Polish Science, co-funded by the European Union under the European Regional Development Fund.