Does time-incorporated admission severity improve stroke prognostication?

IntroductionThe dynamic relationship between time since symptom onset and severity on admission is poorly incorporated in stroke prognostic models. We sought to develop a novel approach and compare it with alternatives. MethodsWe developed admission stroke severity centiles to contextualise a patients stroke severity relative to others admitted at a similar time post ictus. We conducted a secondary data analysis to compare four prognostic models, including the novel severity centiles and other time-incorporated approaches. Models were evaluated using Pseudo-R squared, Bayesian Information Criterion, and Area Under the Curve. ResultsTime since symptom onset not only predicts outcomes but also modifies the prognostic value of severity, with delayed admission amplifying its impact on functional dependency and mortality in acute ischaemic stroke (AIS) patients [Adjusted Odds Ratio (OR) (95% CI) = 1.02 (1.01 to 1.04); Adjusted OR (95% CI) = 1.03 (1.01 to 1.05), respectively]. Centile-based method offered a dynamic interpretation of stroke severity for both stroke types, demonstrating a prognostic signal comparable to traditional time-incorporated models. For AIS patients, each unit increase in centile corresponded to a 5% increase in the odds of functional dependency Adjusted OR (95% CI) = 1.05 (1.04 to 1.05) and a 5% increase in the odds of mortality Adjusted OR (95% CI) = 1.05 (1.04 to 1.07). Similar results were obtained for intracerebral haemorrhagic (ICH) patients. While covariate adjustment improved overall model performance, it did not alter the ranking of time-incorporated methods. Time-incorporated approaches did not enhance model performance for ICH patients, possibly due to a smaller sample. ConclusionsWe are the first to construct centiles for admission stroke severity. Severity centiles based on a reference population offer a graphical tool to support risk assessment and outcome prediction. While this approach provides clinically interpretable measures, our results suggest that explicitly incorporating time into prognostic models offers slightly better predictive performance. We recommend using statistical models that allow the effect of stroke severity to vary according to the time since symptom onset to assessment.