CINP: A Causality-Informed Framework for Generalizable and Replicable Neuroimaging Prediction of Antipsychotic...

Treatment response to antipsychotic drugs in schizophrenia (SCZ) is highly variable, necessitating predictive biomarkers for personalized treatment. While neuroimaging-based predictive models (NPMs) offer promise, their reliance on correlation-based methods renders them vulnerable to confounders. A general framework is required to integrate NPMs with biologically causal information for generalizable and replicable predictions.

We propose causality-informed neuroimaging prediction (CINP), a framework that incorporates Mendelian randomization-derived causal effects as biological priors into model inference. This framework aims to transcend purely correlational approaches by establishing biologically constrained neuroimaging predictions. To test feasibility, we collected multimodal magnetic resonance imaging data from patients with SCZ across 2 independent longitudinal datasets (Peking University Sixth Hospital: n = 37; Zhumadian Psychiatric Hospital: n = 58). Antipsychotic responses were quantified using the percentage reduction in Positive and Negative Syndrome Scale (PANSS) scores from baseline to follow-up. CINP models were trained to predict individualized PANSS score reduction using neuroimaging data. Model accuracy was evaluated via internal cross-validation and generalizability by intersite cross-validation.

CINP achieved an average threefold improvement over conventional NPMs in predicting antipsychotic treatment response. Among various neuroimaging modality configurations, the white matter tract-based CINP yielded the highest prediction accuracy (mean r = 0.652) and cross-site generalizability (mean r = 0.551). Furthermore, predictive features' weights were highly consistent across datasets (similarity = 0.397 to 0.510), indicating replicable patterns.

Beyond feasibility and validity in predicting antipsychotic response, CINP provides a flexible computational framework for incorporating diverse causal constraints, thereby enabling robust neuroimaging-based predictions in multiple clinical contexts.