Simulation and temporal validation of ICU trajectories after stroke: patient condition, transfer readiness and actual ICU exit

Wait 5 sec.

Objective: To describe how intensive care unit (ICU) time is composed after stroke, and to test whether simulating patient condition, transfer readiness and ICU exit separately predicts as accurately as direct models. Materials and Methods: In 6,366 ICU stays for stroke in MIMIC IV, hourly readiness criteria divided ICU hours into organ support needs, stroke specific needs, and time not explained by these needs. An hourly simulator was built on stays up to 2016, frozen, and validated on 2017 to 2019 (n=1,714) and 2020 onwards (n=1,521). The primary measure was the 24 hour alive exit Brier score versus the strongest direct model (gradient boosting). Results: Unexplained time was 34.1% of hours in standard ICUs and 87.6% in neuro intermediate units. 37.8% of stays lost readiness for [≥]6 consecutive hours after first confirmation. The simulator and gradient boosting were indistinguishable (Brier 0.1238 vs 0.1245; difference -0.0007, 95% CI -0.0030 to 0.0016). The simulator was better calibrated at 48 and 72 hours but discriminated less well (C statistic 0.71 vs 0.74 to 0.75). All models overpredicted exit after 2016; the error lay mainly after readiness (median 40 hours simulated vs 68 observed). Discussion: Readiness was not a single event: patients often lost and regained it, and many stayed long after meeting it, especially in neuro intermediate units, which exist for monitoring. Separate modelling did not improve accuracy but placed the post 2016 error after readiness, as the unit mix changed. Conclusion: The simulator matched the strongest direct model while also generating complete trajectories. It needs external validation before any clinical use.