An explainable gradient boosting machine model based on 12 routinely collected clinical variables achieved good discrimination (AUC 0.872 in test set) for predicting 3-month functional outcome after acute ischemic stroke, with NIHSS, age, and C-reactive protein identified as dominant predictors.
Key Findings
Methods
LASSO variable selection identified 12 predictors from baseline clinical and early laboratory variables for predicting 3-month functional outcome after acute ischemic stroke.
Data came from 770 adults with acute ischemic stroke admitted within 24 hours of symptom onset at a single center.
The study used a retrospective prediction-model design with an 80/20 train-test split.
Cross-validated least absolute shrinkage and selection operator (LASSO) was applied in the training set to select predictors.
Variables were drawn from routinely collected baseline clinical and early laboratory data.
Results
Multiple supervised learning models showed similar test-set discrimination, with AUC values ranging from 0.854 to 0.867 across non-final models.
Several supervised learning models were developed and compared on the same train-test split.
Test-set AUC values ranged from 0.854 to 0.867 across the compared models.
Model discrimination, calibration, and clinical utility were all evaluated as performance metrics.
The gradient boosting machine (GBM) was selected as the final model.
Results
The final gradient boosting machine (GBM) model achieved AUCs of 0.910 in the training set and 0.872 in the test set.
Training set AUC was 0.910 and test set AUC was 0.872.
Areas under the precision-recall curve were 0.814 in the training set and 0.741 in the test set.
Calibration was described as 'satisfactory in both sets.'
The model was interpreted using TreeSHAP to provide explainability.
An online Shiny-based calculator was developed to facilitate clinical use.
Results
SHAP analyses identified NIHSS, age, and C-reactive protein as the dominant contributors to GBM model predictions.
TreeSHAP (SHAP analyses) was used to interpret the gradient boosting machine model.
NIHSS, age, and C-reactive protein showed 'predominantly non-linear effects with limited interaction strength.'
These three variables were identified as the top contributors among the 12 LASSO-selected predictors.
The non-linear effects highlight that a standard logistic regression approach may not fully capture these relationships.
Conclusions
The authors conclude that independent external validation is required before the model can be considered for clinical application.
The study is described as a single-center retrospective prediction-model study.
Validation performed was internal only, using an 80/20 train-test split.
The authors explicitly state 'Independent external validation is required before the model can be considered for clinical application.'
Despite good internal discrimination and calibration, generalizability to other centers or populations was not established.
What This Means
This research developed a computer-based tool to predict how well stroke patients will recover three months after their stroke, using only information that hospitals routinely collect when a patient is admitted. The study analyzed 770 stroke patients who arrived at the hospital within 24 hours of their symptoms starting. Using a type of machine learning called a gradient boosting machine, the model correctly distinguished between patients who would have good versus poor functional outcomes about 87% of the time on new test cases it had not seen before — a level of accuracy considered good for clinical prediction tools.
The researchers also used a technique called SHAP to explain which factors mattered most to the model's predictions, finding that stroke severity at admission (measured by the NIHSS score), the patient's age, and a blood marker of inflammation called C-reactive protein were the three most important predictors. Notably, these factors influenced outcomes in complex, non-linear ways, meaning that small changes in these values could have varying impacts depending on the overall clinical picture. The team built a free online calculator so clinicians could easily apply the model to individual patients.
This research suggests that routinely available clinical data — without the need for advanced imaging or specialized tests — can be used to make reasonably accurate early predictions about stroke recovery. This could help doctors communicate prognosis to patients and families, plan rehabilitation early, and allocate hospital resources more effectively. However, the study was conducted at only one hospital, and the authors caution that the model must be tested at other institutions before it should be used in real clinical settings.
Rao Z, Xu C, Yu Y, Liu Y, Zhu C, Zhang Y. (2026). Early prediction of 3-month functional outcome after acute ischemic stroke: an explainable model based on routine clinical data.. Frontiers in neurology. https://doi.org/10.3389/fneur.2026.1852875