Abstract Background and aims Stroke severity evolves in the acute phase yet is typically measured only on admission and treated as a static single-timepoint measure. This study uses routinely collected repeated NIHSS assessments to characterise the dynamic nature of stroke severity and assess their value for prediction. Methods We analysed real-world data from 1,622 ischaemic stroke patients in the Bergen NORSTROKE study with at least two NIHSS assessments within 48 hours of symptom onset. Stroke severity trajectories were examined using descriptive statistics, linear mixed-effects models (LMM), and group-based trajectory modelling (GBTM). Seven logistic regression models incorporating different representations of severity dynamics were compared for predicting favourable short-term functional outcomes (modified Rankin Scale), with performance assessed using AIC, BIC, and AUC. Results NIHSS scores varied considerably over time. Most patients had mild strokes (median NIHSS = 2) and showed improvement within 48 hours. GBTM identified three latent groups: Very Low-Stable (40.3%), Moderate Low-Stable (41%), and High-Mildly Improving (18.7%). Models incorporating stroke severity dynamics outperformed those using admission NIHSS alone (AUC 0.835 vs. 0.778). The best predictive performance was achieved using random intercepts and slopes from the LMM. While GBTM improved model fit (AIC = 1624), its added discriminatory power was limited (AUC = 0.792). Conclusions Stroke severity can change substantially in acute phase, and in this dataset, models capturing these dynamics outperformed those based on a single admission score. As serial NIHSS is not routine, these findings are exploratory. Future work should assess when and how often reassessment is feasible, and whether updating predictions improves decision-making across centres. Conflict of interest Nothing to disclose for all authors
Lu et al. (2026) studied this question.