Abstract / Summary
Abstract Background Recurrent ischemic stroke remains a major source of preventable disability after an index event. Conventional risk assessment relies mainly on clinical variables measured at isolated time points and may not capture evolving systemic metabolism or brain-network organisation. We evaluated whether longitudinal metabolomic information and high-density electroencephalographic (HD-EEG) features could be integrated to distinguish patients with and without one-year recurrence and to determine how much of the observed signal was retained after model simplification. Methods This single-centre observational cohort included 428 adults with acute ischemic stroke enrolled between January 2020 and December 2022 and followed for one year; 51 participants had recurrent ischemic stroke and 377 remained recurrence-free. Longitudinal plasma UPLC-MS/MS metabolomic features and 128-channel resting-state HD-EEG microstate and functional-connectivity features were represented using a Transformer-based temporal encoder and a graph-based connectivity encoder. The source analysis used five-fold cross-validation with Bayesian hyperparameter optimisation. Because the archived fold-level and event-by-visit files were insufficient to verify fully nested preprocessing or the temporal eligibility of every longitudinal observation, the present report interprets performance as retrospective one-year recurrence-status classification rather than validated dynamic prediction. Results Participants with recurrence were older than recurrence-free participants (67.4 vs 63.2 years, P = 0.008) and had higher admission NIHSS scores (median 8 vs 5, P < 0.001). The full multimodal model yielded a reported ROC AUC of 0.91 and Brier score of 0.076; the source analysis also reported a 95% CI of 0.88–0.94 for the AUC, although the CI construction could not be independently reconstructed. A model without EEG yielded an AUC of 0.87 and a metabolomics-only model an AUC of 0.85, corresponding to nominal differences of 0.04 and 0.06 from the full model. At the observed recurrence prevalence (11.9%), a constant-risk reference has a Brier score of 0.105; thus the reported model Brier score represents a descriptive 27.6% reduction in mean squared probability error relative to that reference. This comparison does not substitute for calibration slope or intercept. Formal paired uncertainty estimates for between-model differences were unavailable. Conclusions Longitudinal metabolomic and HD-EEG features contained a strong internally evaluated signal for one-year recurrence-status classification, and the reduced-modality results suggest that metabolomics accounted for a substantial share of that signal. The nominal additional discrimination from HD-EEG warrants formal paired testing, but it should not yet be interpreted as confirmed incremental value. A participant-level, leakage-resistant nested reanalysis with auditable pre-event eligibility, full calibration and independent multicentre evaluation is required before this framework is described as prospective patient-level risk prediction. Clinical trial registration Chinese Clinical Trial Registry (ChiCTR2400079697) registered on 10 January 2024.