Causality-Inspired Neural Network for the Identification of Schizophrenia

Functional connectivity (FC) analysis has emerged as a pivotal tool for identifying neural biomarkers in schizophrenia. However, existing methods often lack interpretability and fail to capture temporally dynamic causal connectivity. To address this limitation, we propose a novel Granger causality (GC)-inspired Convolutional Long Short-Term Memory (cLSTM) model for diagnosing schizophrenia. Our framework integrates a dynamically learned sparsity-inducing mask within the cLSTM architecture to prioritize causal connectivity patterns while filtering out non-informative connections, thereby enhancing computational efficiency and model interpretability. We evaluated the model on the COBRE dataset across seven parcellation atlases, achieving superior performance with a mean accuracy exceeding 90% and F1-scores of up to 92%, thereby outperforming state-of-the-art methods. The GC-inspired mask reduces redundant parameters by 40%–60%, facilitating the identification of clinically relevant biomarkers, including dysregulated prefrontal-hippocampal and default mode network (DMN) interactions. By integrating temporal dependency modeling with causal inference, our approach not only enhances diagnostic accuracy but also provides neurobiologically interpretable insights into functional disruptions associated with schizophrenia. This study bridges the gap between complex deep learning (DL) models and clinically actionable tools, demonstrating significant potential for psychological healthcare applications.