Interpretable machine learning-based prediction of 90-day mortality after endovascular treatment in acute large vessel occlusion stroke.

The 90-day all-cause mortality rate among AIS-LVO patients with successful endovascular treatment was 19.6%.

• 88 out of 449 patients died within 90 days.
• All patients had successful recanalization defined as modified Thrombolysis in Cerebral Infarction (mTICI) grade 2b-3.
• The cohort was enrolled at a nationally certified Advanced Stroke Center in Guangxi, southwestern China, between January 2022 and May 2025.
• Patients were randomly split 8:2 into training (n = 359) and independent test (n = 90) sets, stratified by 90-day mortality.

Random Forest achieved the best overall performance among the fifteen machine learning algorithms evaluated for predicting 90-day mortality.

• Random Forest test-set performance: AUC = 0.806, Brier score = 0.128, Hosmer-Lemeshow p = 0.315.
• Fifteen ML algorithms were optimized via Bayesian hyperparameter search with 5-fold cross-validation.
• The top five models underwent bootstrap optimism correction.
• Performance was evaluated across discrimination, calibration, clinical utility (decision curve analysis), and reclassification (NRI/IDI).
• The Hosmer-Lemeshow p = 0.315 indicates good calibration (non-significant, meaning predicted probabilities matched observed outcomes).

Random Forest demonstrated positive net clinical benefit across threshold probabilities of 0.10 to 0.60 on decision curve analysis.

• Decision curve analysis was used to assess clinical utility of the model.
• Positive net benefit across a wide range of threshold probabilities (0.10–0.60) indicates clinical usefulness for risk stratification.
• This range encompasses clinically relevant decision thresholds for mortality risk.

SHAP analysis identified eight top predictors of 90-day mortality after EVT, led by stroke-associated pneumonia, D-dimer, and age.

• The top contributors identified by SHAP analysis were: stroke-associated pneumonia, D-dimer, age, acute renal insufficiency, fasting blood glucose, NIHSS score, brain herniation, and symptomatic intracranial hemorrhage.
• SHAP (SHapley Additive exPlanations) and permutation importance were applied for model interpretability.
• The SHAP results were concordant with permutation importance rankings.
• Complication-based predictors (e.g., stroke-associated pneumonia, brain herniation, symptomatic intracranial hemorrhage, acute renal insufficiency) are incorporated into the model only after they clinically occur, supporting dynamic in-hospital prediction.

Random Forest showed comparable discrimination to LR-LASSO without significant reclassification improvement.

• Net Reclassification Improvement (NRI) = -0.472 when comparing Random Forest to LR-LASSO.
• Integrated Discrimination Improvement (IDI) = -0.067 when comparing Random Forest to LR-LASSO.
• These values indicate that Random Forest did not significantly outperform LR-LASSO in reclassification metrics.
• Despite the lack of significant reclassification improvement, Random Forest was selected for its overall performance profile including calibration and clinical utility.

The study developed an online calculator based on the Random Forest model to support dynamic post-EVT in-hospital mortality risk stratification.

• The online calculator incorporates SHAP-based interpretability to provide individualized explanations of predictions.
• The model supports dynamic prediction by incorporating complication predictors only after they clinically occur during hospitalization.
• The model was internally validated using a retrospective cohort of 449 patients.
• Prospective external validation in geographically diverse multicenter cohorts was identified as a necessary next step.

Candidate predictors spanned four domains: baseline characteristics, admission laboratory indicators, procedural variables, and post-procedural complications.

• The study retrospectively enrolled 449 AIS-LVO patients with successful EVT.
• Predictor categories included baseline characteristics, admission laboratory indicators, procedural variables, and post-procedural complications available during hospitalization.
• The study design was retrospective and conducted at a single center in Guangxi, southwestern China.
• The authors noted that existing prognostic models predominantly target functional disability rather than mortality and rarely incorporate calibration, clinical utility, or interpretability.