Clinical Guideline Categorization for Stroke Prediction Maintains Accuracy

Machine learning models have shown high accuracy in predicting 90-day outcomes for acute ischaemic stroke, but their adoption in clinical settings is hindered by a mismatch between model explanations and clinicians' reasoning. This research explored whether converting continuous predictors into clinically relevant, guideline-based categorical thresholds could bridge this gap without compromising predictive power. The study compared standard gradient-boosted models with fully categorized versions, where continuous variables were discretized according to stroke guideline-aligned, treatment-specific thresholds. Across a multi-center European registry, the fully categorized models performed statistically indistinguishably from their continuous counterparts in two out of three treatment cohorts. Only one cohort showed a significant, albeit small, drop in accuracy. Crucially, global feature importance rankings remained consistent between the continuous and categorized models, indicating that the core prognostic factors were preserved. This suggests that guideline-based categorization is a viable design choice for stroke-outcome models, potentially enhancing their interpretability and facilitating clinical adoption by aligning with established medical practices.