Clinical status after acute stroke frequently changes within hours of admission to the intensive care unit (ICU). Traditional prognostic approaches typically predict fixed endpoints from admission data or assign a single trajectory label per patient. The authors sought to use high-resolution longitudinal ICU data to identify interpretable, time-varying clinical states, to characterize transitions among those states during the early ICU course, and to relate the current state to near-term clinical events.
This was a retrospective cohort study in 6,368 adult patients with acute stroke drawn from MIMIC-IV v3.1. The analysis focused on the first 72 hours after ICU admission, divided into twelve consecutive 6-hour windows. A hidden Markov model was fitted to 21 variables covering neurological status, physiological measurements and organ-support modalities.
State number was chosen against criteria prespecified before model fitting: statistical fit, restart stability, state occupancy and clinical interpretability. The authors used generalized estimating equations to link the current state to two near-term outcomes: initiation of new mechanical ventilation and new vasopressor use within 12 hours, and ICU death within 72 hours. The study also included eleven sensitivity analyses to test robustness and evaluated reproducibility across training and test splits and admission eras.
The model produced four clinically interpretable states. Window-level occupancy of these states was: neurologically preserved–low support (63.3%), neurological impairment–low support (7.8%), impairment–renal dysfunction (11.8%) and impairment–respiratory support (17.1%). These labels reflect combinations of neurological findings, physiological derangements and organ-support needs captured by the 21 input variables.
Although the model was trained without outcome data, the last-observed state before outcome assessment showed marked differences in ICU mortality: mortality by last state ranged from 2.9% to 43.9%.
Transitions among the four states were common: 40.3% of patients changed state at least once within the first 72 hours. The observed transitions occurred in both directions rather than progressing strictly along a single unidirectional severity axis, indicating dynamic improvement and deterioration among patients.
After adjustment for age, sex, stroke subtype and Charlson comorbidity index, the current state remained associated with subsequent escalation of organ support (new mechanical ventilation and vasopressor initiation within 12 hours) and with ICU death within 72 hours. These adjusted associations indicate that the time-updated state contains prognostically relevant information beyond baseline covariates.
The model solution showed stability across multiple checks. State prevalence differed by at most 1.1 percentage points between training and test datasets. Ten of the eleven sensitivity analyses produced the same four-state solution; clustering agreement metrics (adjusted Rand index) were approximately in the range 0.754–0.955, indicating substantial to very high concordance across perturbations.
The early ICU course after acute stroke can be represented as movement among a small number of clinically interpretable dynamic states. These states are reproducible within the source dataset and across admission eras and are associated with near-term needs for organ support and with ICU death, even after adjustment for baseline risk factors. Because states were discovered without using outcome information yet nonetheless predicted outcomes, time-updated state assessment may offer a compact clinical summary useful for monitoring and prognostic signal extraction.
However, the authors explicitly note that this representation requires validation in an independent database before any clinical deployment or decision support application.
The underlying MIMIC-IV data resource was reviewed by the Institutional Review Board of Beth Israel Deaconess Medical Center, which granted a waiver of informed consent for the research database. The dataset analyzed is publicly available through PhysioNet to credentialed users who complete required training and sign the data use agreement; the data cannot be redistributed by the authors. Analysis code will be made available upon reasonable request. The authors declared no competing interests.