External validation of dynamic clinical states in acute stroke: transportability and independent rediscovery across 176 hospitals
Year: 2026
Type: preprint
Abstract: Objective Four dynamic clinical states were identified in the first 72 h of stroke intensive care in one centre. We tested whether this representation transports to an independent multicentre cohort, predicts the next state, and is independently recovered. Materials and Methods 8279 adults in 176 hospitals (eICU-CRD v2.0) contributed 70 630 six-hour windows. Phenotype, dictionaries and eligibility rules were frozen before any state was assigned; the model was applied unchanged and judged against five prespecified criteria. Generalised estimating equations related state to subsequent organ support and ICU death; prediction was assessed against a persistence null at 6, 12 and 24 h; a model was fitted de novo in the strictest scope. Results All five transportability criteria were met in all four scopes. All nine state–outcome comparisons reproduced the direction of association, with exact rank order for invasive ventilation and ICU death; between-hospital intraclass correlations were 0.012–0.027. States persisted across 90.7% of pairs, so prediction was scored on change: AUROC (95% CI) 0.730 (0.723–0.736) externally, 0.735 (0.721–0.749) internally. Fitted de novo, eICU favoured four by BIC but three by restart reproducibility; neither recovered neurological impairment–low support, the other three matched closely in both (r = 0.962–0.991). Discussion The representation transports and carries forward-looking information but is not fully rediscoverable. The unrecovered state is defined by a combination, not a feature: eICU places its windows consistently but never pairs impairment with absent organ support. Conclusion Three of four states are strongly supported. Transportability and independent rediscovery are distinct and should be reported separately.
Source: medRxiv
Authors: Ping Lei, Yanyi Xu, Yuxia Zhang
Institutions: University of Geneva, Fudan University, Zhongshan Hospital
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212009727