Study Finds Guideline-Based Categorization Can Match Continuous Predictors in Stroke Outcome Prediction

A new study, available on arXiv (arXiv:2608.05203), explores the potential of using categorical encodings that align with clinical guidelines as substitutes for continuous predictors in machine learning models aimed at predicting outcomes 90 days post-acute ischaemic stroke. This research stemmed from a clinician user study advocating for cut-offs that reflect clinical standards. By utilizing a multi-centre European registry, the authors evaluated standard gradient-boosted models against fully categorized versions that applied treatment-specific, guideline-aligned thresholds. Patients were divided into three treatment groups. The findings revealed that the categorized models were statistically similar to continuous models in two cohorts, although one cohort experienced a notable decline in predictive accuracy. Notably, global feature importance rankings were consistent, indicating that discretizing continuous predictors does not necessarily compromise model interpretability. These results are significant for the clinical integration of machine learning models, as aligning model explanations with clinician reasoning may enhance trust and usability. The research team announced this paper on arXiv.