A Closed-Loop Stroke Digital Twin: Integrating BCI-Derived Electrophysiology and Spatiotemporal Multi-Omics via Transformer-GNN Fusion Networks

Background Even with optimal secondary prevention, the residual risk of recurrent ischemic stroke remains clinically significant: epidemiological data indicate that approximately one in six patients experiences a recurrent ischemic event within five years [1, 3, 49]. Conventional risk stratification tools rely on cross-sectional data, which precludes detection of the continuous molecular and electrophysiological changes that precede recurrent events. Ferroptosis—an iron-dependent, lipid peroxidation-driven form of regulated cell death—and the downstream failure of SIRT1/PGC-1α-mediated mitochondrial biogenesis have emerged as mechanistically important but clinically unmonitored contributors to this residual risk. Methods We developed M4-RIS (Multimodal Multi-timescale Recurrence Ischemic Stroke), a closed-loop Stroke Digital Twin framework integrating two sensing layers: a low-frequency biochemical layer employing spatial transcriptomics and longitudinal UPLC-MS/MS multi-omics to quantify ferroptosis-related metabolic flux [5, 6], and a high-frequency physical layer deploying non-invasive brain-computer interfaces (BCI; high-density EEG) to extract electrophysiological microstates. A multimodal fusion deep neural network (MF-DNN) combining Transformer and Graph Neural Network (GNN) architectures decoded the cross-modal electro-metabolic interactions, with SHapley Additive exPlanations (SHAP) providing patient-level interpretability. Results In a prospective cohort of 428 patients with acute ischemic stroke, the MF-DNN achieved an AUC of 0.91 (95% CI: 0.88–0.94) for one-year recurrence prediction, substantially outperforming established clinical scores. SHAP analysis identified dysregulation of the AMPK/SIRT1/PGC-1α axis and accompanying lipid peroxidation as the dominant prognostic determinants [16, 17, 18]. In the closed-loop neuromodulation subgroup, electroacupuncture at ST36 activated PROKR2Cre-marked somatosensory neurons [26], engaging the vagal-adrenal axis to suppress systemic inflammation; patients in whom this intervention normalised EEG microstate C duration showed a concurrent reduction in circulating 4-HNE concentrations. Conclusions Fusing multi-timescale biosensing with interpretable deep learning establishes a mechanistic foundation for proactive, data-driven secondary stroke prevention. The M4-RIS framework advances clinical management from episodic, static observation toward continuous, closed-loop intervention guided by individual biological trajectories. Pending multi-centre validation, this approach offers a concrete and practically achievable pathway toward reducing the residual recurrence risk that current standard-of-care therapies have been unable to eliminate.