Clinical thresholds did not worsen the prognosis for stroke—but not everywhere
Sh0ny
7 августа 20262 min read
In short
The study shows that a model for predicting stroke outcomes can be made easier for doctors to understand by replacing continuous variables with clinical categories. However, this optimization comes at a cost: in one of the three treatment groups, accuracy decreased significantly.
Medical models face an unfortunate trade-off: the closer their explanations align with a clinician’s familiar logic, the greater the risk of losing accuracy. New research shows that for predicting 90-day outcomes following an ischemic stroke, this trade-off may be less significant than it seems.
The researchers compared standard gradient boosting models with versions in which continuous variables were replaced with categories and thresholds from clinical guidelines. For example, the model does not use just any numerical value for a feature, but rather whether it falls within a specific clinical range—making it easier for a doctor to compare the prediction with standard clinical practice.
The results were mixed. In two of the three treatment groups, the categorized models did not differ statistically in accuracy from the models with continuous features. In the third group, accuracy decreased significantly. However, the ranking of the most important factors remained consistent across all groups: simplifying the input data did not alter the underlying hierarchy of prognostic features.
The practical conclusion is not that it is time to convert all medical models to threshold-based ones as recommended. Rather, this is a practical way to bring model explanations closer to clinical reasoning, provided that the cost of such a change is assessed in advance for each patient group. For one treatment regimen, the trade-off in accuracy may be negligible; for another, it may not be.
There are significant limitations here: the conclusion is based on a multicenter European registry and three treatment cohorts, rather than on the day-to-day implementation of the model in clinical practice. The study also highlights the preservation of accuracy and the overall importance of clinical features, but does not prove that physicians will actually make better decisions using this model.
If you were to implement such a system, what would be more important to you: a few points of accuracy or an explanation based on familiar clinical thresholds? Source: cs.AI updates on arXiv.org