Stroke outcome AI: guideline categories match continuous predictors in most cohorts

Machine learning models show strong accuracy for predicting 90-day outcomes in acute ischaemic stroke, but clinical adoption stalls because model explanations don't align with how doctors reason. A new paper on arXiv (2608.05203) from Esra Zihni and colleagues directly tests a solution: replacing continuous predictors with categorical encodings based on clinical guideline thresholds. Using a multi-centre European registry split into three treatment cohorts, they trained standard and fully categorised gradient-boosted models, with categorisation thresholds tailored to stroke treatment guidelines.

The results are promising but mixed. In two of three treatment cohorts, the categorised models performed statistically on par with their continuous counterparts, while one cohort saw a significant accuracy drop. Importantly, global feature importance rankings remained consistent across all groups, meaning the core hierarchy of prognostic factors was preserved even after discretisation. The authors conclude that guideline-based categorisation is a viable design choice for stroke-outcome prediction, offering a path toward models that are both accurate and interpretable to clinicians.